Digital Information World

Researchers Tracked $80M In Spending On Almost 170 AI-Generated Political Ads This Year, Here’s What They Found

2026-10-06T19:32:37.790+05:00

By Travis N. Ridout, Washington State University; Erika Franklin Fowler, Wesleyan University, and Michael Franz, Bowdoin College

Image: Theo Laflamme - Unsplash

The 2026 U.S. midterm campaigns are the first in which AI-generated political ads are regularly appearing on people’s televisions and social media feeds.

We are researchers who have been studying political advertising through the Wesleyan Media Project since 2010. This election cycle – using data from media reports, student coders and AdImpact, a firm that tracks political ad spending – we’ve tracked about US$80 million in spending on almost 170 unique ads that use AI.

What we found has surprised us: AI use spans from hyperrealistic deepfakes to subtle enhancements; Republican sponsors – both candidates and interest groups – are much more likely to use AI than are Democratic sponsors; and, thanks to a patchwork of state legislation, many of these ads do not disclose the use of AI at all.

From deepfakes to subtle edits

An assortment of politicians and watchdog groups have expressed concern about campaigns using generative AI to produce deepfakes – synthetic videos showing people doing things they did not do – that might deceive voters.

We’ve noticed several ads containing hyperrealistic deepfakes of famous politicians, including Donald Trump, Nancy Pelosi, Barack Obama, Kamala Harris and Alexandria Ocasio-Cortez. Ocasio-Cortez, in particular, is a favorite among Republican advertisers, appearing in at least five ads.

We’ve seen deepfakes in which a Republican Senate candidate from Louisiana drives a school bus full of undocumented immigrants, a Republican candidate for governor from South Carolina walks arm in arm with drag queens, and an ad in which Liz Cheney, Mitt Romney and Mike Pence are seen carrying pitchforks on the White House lawn.

People who are not politicians made appearances, too, including a fake Dr. Anthony Fauci, seen running around a state fair with a huge syringe, presumably eager to vaccinate everyone. We’ve also noticed several ads in which AI was used to generate crowds or constituents.

In several cases, AI was used to enhance visuals rather than generate something new. One ad from Chip Keating, a Republican candidate for governor in Oklahoma, includes an AI disclaimer, but it doesn’t specify exactly how AI was used. Ads that use AI to enhance visuals don’t necessarily look different from ads that were created in the pre-AI era, which makes it difficult for viewers to discern whether they depict something false.

A partisan gap

Republicans – both candidates and groups such as super PACs and 501(c) organizations – are much more likely to use AI in their ads than are Democrats, according to our research.

In fact, Republican candidates or pro-Republican groups were behind 80% of the ads we tracked and 83% of the spending.

We can only speculate as to why Republicans dominate the use of AI in political advertising in 2026. In general, Democrats tend to take on a regulatory mindset when it comes to political campaigns, favoring limits on campaign spending and required disclosures. In 2022, for instance, only Senate Democrats and two independents voted to advance the DISCLOSE Act that would have required additional campaign finance disclosures for super PACs, labor unions and corporations. Republicans, by contrast, tend to be more in favor of a free market approach.

These more general philosophies may be reflected in the parties’ use of generative AI for political advertising, something about which voters are worried. Polling shows broad support for more regulation, with 78% of registered voters favoring a ban on AI content that makes deceptive claims about candidates.

Disclaimers all over the map

Because regulation of AI in advertising depends on a patchwork of state legislation, many of these ads are not required to disclose the use of AI. This lack of disclaimers makes tracking AI use in ads challenging. Our team has relied on media coverage and trained student coders to flag ads that are potentially AI-generated.

Across 35 states, only 31% of the ads we tracked – representing 22% of the spending – disclosed the use of AI tools. The wording of these disclaimers was all over the map. For example, one Georgia ad included the disclaimer, “This video has been manipulated or generated with artificial intelligence,” while an Oklahoma ad said, “Political satire. AI-generated images do not depict actual events.” A North Carolina state Senate ad said, “You guessed it! AI was definitely used to generate these silly video clips.”

Laws don’t drive disclosure

Some advertisers voluntarily disclose the use of AI even when they are not required to do so. Sometimes, the opposite occurs – advertisers don’t include disclaimers even when state law requires them to. In fact, we’ve found that a state law that requires disclaimers on ads that use AI has very little relationship with the actual use of disclaimers.

In states without laws governing the use of AI in political ads, 32% of ads contained a disclaimer; by contrast, in states with laws governing the use of AI, 29% contained a disclaimer. This comparison, however, is not perfect, as some of the ads in our database aren’t covered by their state’s law. Some states, such as Colorado, have laws that apply only to candidate deepfakes and thus exclude other uses of AI. In other instances, such as in Louisiana, an AI law was enacted after the ad aired.

Minnesota law generally bans deepfakes in political ads, but an ad featuring synthetic video of Democratic Senate candidate Peggy Flanagan aired in May 2026 anyway. Whether it violated state law is uncertain, as the law requires that the media be “so realistic that a reasonable person would believe it depicts speech or conduct of an individual who did not in fact engage in such speech or conduct.”

Moving forward, the real policy challenge will revolve around transparency, the enforceability of existing laws and simply figuring out what type of disclosure would be helpful to voters.

Travis N. Ridout, Professor of Government and Public Policy, Washington State University; Erika Franklin Fowler, Professor of Government, Wesleyan University, and Michael Franz, Professor of Government, Bowdoin College

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Fact-Checked by Irfan Ahmad.

Read next:

• 42% Of Americans Wouldn’t Trust AI To Make Purchases For Them

• Attention can’t be bought, so brands turn to memes

42% Of Americans Wouldn’t Trust AI To Make Purchases For Them

2026-10-06T17:12:13.990+05:00

By Felix Richter, Statista

While Americans are quite open to shopping with the help of AI, there’s one point where many still draw the line and that’s the actual checkout button. While almost 4 in 10 respondents in a Statista Consumer Insights survey said that they made at least one purchase based on AI recommendations in the past year and 48 percent said they use AI when shopping or researching products often or very often, noticeably fewer said they’d feel comfortable letting AI make autonomous purchases.

Surprisingly, Americans would feel most comfortable letting AI pick and purchase clothing and accessories for them at 27 percent, closely followed by electronics and technology, travel and experiences and personal care products. 23 percent would let AI buy everyday essentials autonomously, which is something that AI would probably be quite good at already after learning a household’s typical consumption habits. Luxury products are the category consumers are least likely to entrust AI with, probably due to fear of unwanted big-ticket purchases.

42 percent of respondents said they wouldn’t trust AI making any purchases on their behalf, suggesting that AI shopping agents still have some convincing to do in order to reach broader adoption. For more insights on AI adoption in shopping and other consumer trends, check out Statista's latest Consumer Reality Update, which you can download here.

Reviewed by Irfan Ahmad.

Read next:

• Attention can’t be bought, so brands turn to memes

• OpenAI Is Adding Watermarks To ChatGPT Text While Expanding Ads

Attention can’t be bought, so brands turn to memes

2026-10-06T15:20:32.634+05:00

Farhan Mutaqin, University of Edinburgh and Almukantar Fikriansyah

The funniest post you saw this week may not have come from a friend; it may have come from a brand. Whether it is Kohl’s, Wendy’s, or Netflix, brands are behaving more like internet users than traditional advertisers.

This is no accident. Simply paying for advertising is no longer enough to capture people’s attention when they scroll thousands of posts every day. Brands are increasingly trying to join conversations that people are already having.

Memes, jokes and real-time reactions have become marketing tools. But why are brands investing so much in being culturally relevant rather than simply producing the perfect advertisement?

Image: kevin zhao - Unsplash

Advertising in the age of endless scrolling

Traditional marketing often worked through interruption. A television commercial, billboard or social media ad place a message in front of people and hoped they would remember it.

That model still works. Major brands continue to spend heavily on it. But buying advertising space is no longer enough. A brand can reach more people when audiences find its content relatable enough to share.

This is the logic of the attention economy. Decades ago, economist Herbert Simon argued that when information becomes abundant, attention becomes scarce.

Today, every social media feed is a competition for that scarce resource. Brands can buy reach, but they cannot buy people’s willingness to share something.

A post that people voluntarily pass along becomes earned media: content that travels through audiences rather than through paid advertising.

Why do memes attract attention?

Humour is only part of the story. What makes content spread is often the emotion behind it.

A 2012 study on viral content found that content triggering high-arousal emotions, such as awe or anger, is more likely to be shared than content that produces little emotional response.

Amusement can work in much the same way. A good meme gives us a small emotional jolt that makes us want to send it to someone else.

Thus, a clever meme is a form of social currency: passing it on makes us look funny and in on the joke.

When brands successfully participate in that culture, they can feel less like distant corporations and more like another voice in the group chat.

That is the real reason memes work: they earn our attention instead of buying or interrupting it. And being chosen is worth far more than being seen.

Real-time marketing can make this especially powerful. During the recent World Cup, Levi’s responded to its stadium name being covered up for the tournament by changing its social media profile pictures to mimic the concealed logo.

The 2013 Super Bowl blackout offers another famous example. During the unexpected power outage, Oreo posted its “You can still dunk in the dark” message.

The post worked because it was timely, simple and connected to what millions of people were already watching.

The line between funny and cringeworthy

For every brand that gets it right, many others get it wrong.

A meme that lands can make a brand feel human. One that misses can make it look as though the brand is desperately trying to be relevant.

Common mistakes include using outdated memes, forcing products into unrelated jokes or adopting a tone that does not fit the brand.

Gucci’s 2017 #TFWGucci campaign, for example, was criticised for making memes feel forced and inconsistent with the luxury brand’s identity.

IHOP faced a different problem in 2015 when it posted a stack of pancakes with the caption “flat but has a GREAT personality”. The joke was widely criticised as sexist. IHOP deleted the post and apologised.

There is even a phrase for brands that try too hard to sound like internet users: “How do you do, fellow kids?”.

Sometimes, the smartest marketing decision is not to join the conversation at all.

When the audience becomes the content

The shift is bigger than a fashion for funny posts. It reflects a change in what brands are competing for.

Reach can still be bought. Attention, however, increasingly has to be earned.

Some brands go further by making consumers part of the content itself. Spotify Wrapped gives users personalised summaries of their listening habits, which millions then share on social media.

Spotify provides the format; users provide the jokes. A listener might share their results alongside a joke about spending 200,000 minutes listening to the same artist. The company creates the template, while the audience turns it into something worth sharing.

Consumers are no longer simply receiving branded content. They are distributing, adapting and sometimes creating it.

That changes what marketers need to understand. The challenge is no longer simply how to communicate a message clearly, but whether the audience will find it relevant enough to pass on.

Every December millions share theirs unprompted, and the internet does the rest, like the screenshots claiming you spent 200,000 minutes overthinking. The brand supplies the canvas, the audience supplies the meme, and the listener becomes the medium either way.

That rewrites the job. The winning skill is no longer a bigger budget or a cleverer thirty-second script. It is cultural literacy: knowing what a community finds funny, what it finds tired, and when a brand has no business speaking at all.

For the next generation of marketers, the brief is moving from “how do we say this well” to “why would anyone pass this on.”

Farhan Mutaqin, PhD Researcher, University of Edinburgh and Almukantar Fikriansyah, MSc Marketing (cand.) at The University of Edinburgh

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Fact-Checked by Irfan Ahmad.

Read next:

• OpenAI Is Adding Watermarks To ChatGPT Text While Expanding Ads

• Apple’s Siri AI Can Access More of Your iPhone. Here’s How to Limit It

OpenAI Is Adding Watermarks To ChatGPT Text While Expanding Ads

2026-10-06T12:56:57.649+05:00

Fact-Checked by Irfan Ahmad

OpenAI is making two notable changes to ChatGPT, which now reaches 1.2 billion people each week globally.

The company is adding invisible watermarks to eligible AI-generated text in the European Union while expanding its advertising system with visual ads during image generation.

The company announced the changes on Oct. 5, saying the text watermarking rollout responds to the EU AI Act, while the new visual advertising format is intended to give businesses another way to promote products and services inside ChatGPT.

Image: Gavin Phillips - Unsplash

ChatGPT Text Will Get Invisible Watermarks In The EU

Over the coming weeks, OpenAI plans to add an invisible watermark to eligible ChatGPT and Codex text output for users across all plans in the EU, where the service had 159.1 million average monthly active recipients.

The technology, called textGrain, adds an invisible statistical signal through the model's word choices. OpenAI says a detector can analyze the resulting text to determine whether it contains an OpenAI watermark.

API customers worldwide can also opt in to watermarking for select models, although the feature will remain off by default for the API, OpenAI said in its news release.

OpenAI is initially limiting access to its text watermark detector to approved researchers and expert organizations.

The technology has important limitations. OpenAI's tests found that replacing 10% of words with synonyms reduced detection from about 92% to 66%, while replacing 25% reduced it to 17%. Shorter text and content with less flexibility in word choice, such as mathematics, can also be harder to detect.

OpenAI also says a watermark does not establish who owns text, who is responsible for it, how much a human contributed, or whether the content is accurate. A missing watermark does not prove that a person wrote the text either.

Anthropic has taken a similar approach with Claude. The company said future Claude models will generate watermarked text to help determine whether Claude was involved in producing it, citing the EU AI Act. Anthropic says its watermark does not identify a person, organization or conversation.

OpenAI Is Also Expanding ChatGPT Ads

Separately, OpenAI is testing a new visual advertising format that will initially appear during image generation in ChatGPT.

The test is scheduled to begin later this month in the U.S. with an initial group of advertisers. OpenAI says the ads will be clearly labeled and kept separate from the image being generated. It also says advertising does not influence ChatGPT's answers.

Image: OpenAI

OpenAI is expanding the measurement infrastructure around these ads at the same time. Integrations with Hightouch, Tealium and LiveRamp allow advertisers to send conversion data from their existing systems to ChatGPT Ads.

Related: You probably wouldn’t notice if an AI chatbot slipped ads into its responses

The company also lists attribution partnerships with AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge and Tenjin, along with full-funnel and advanced measurement partners Fospha, Measured and INCRMNTAL.

OpenAI is also working with Haus, Measured and WorkMagic on geo-based experiments intended to help measure the causal impact of advertising. Its brand-suitability work includes DoubleVerify and Integral Ad Science, which are developing evaluation pilots around OpenAI's advertising safeguards.

Some early results cited by OpenAI come from these partners. DV Rockerbox reported that WeightWatchers' attributed cost per acquisition on ChatGPT Ads was 15.3% lower than its blended paid-search benchmark. WorkMagic reported statistically significant lift for wellness brand Dose, while Triple Whale reported that 93% of Portland Leather's visitors from ChatGPT Ads were new. These figures are partner-reported findings presented by OpenAI.

Apple’s Siri AI Can Access More of Your iPhone. Here’s How to Limit It

2026-10-05T17:14:06.976+05:00

By Thorin Klosowski, Electronic Frontier Foundation

Image: BoliviaInteligente - Unsplash

There’s no denying that the new version of Siri is far more powerful than it used to be, and arguably more useful at surfacing details on your phone. But that comes at the cost of deeper access. Once enabled, Siri and Spotlight are combined, unifying the interface. Where you may have once just pulled down on the screen to search for an app or contact, you’re now also invoking Siri.

By default, ask Siri a question and it’ll search through your Apple apps, like Notes, Messages, emails, and more. As time goes on, if the developer chooses to let it, Siri will gain access to more and more third-party apps. If an app developer doesn’t add that support, then Siri AI won’t be able to access the contents of that app (unless it’s shared screenshot-style via a new feature called “on-screen awareness,” which we’ll talk about more in a moment).

For example, if Signal doesn’t choose to implement Siri AI support and you only talk to Bill on Signal, you won’t get an answer when you ask Siri AI, “What was the last photo Bill sent me?” But if you talk to Bill on Apple Messages and ask that same question, Siri AI will summarize what it thinks the photo is.

Sometimes Siri processes this data on your device. Sometimes it uses Apple’s Private Cloud Compute (PCC), which means the data is sent off your device to a cloud server. While you can try digging through the Apple Intelligence Report to figure out what’s sent to PCC, there’s no immediate visual indication from the user’s point of view when data leaves the device or when AI can handle it on the phone, iPad, or Mac itself. In practice, ask Siri AI a question and you’ll never really know if it’s being computed on device or off.

Apple claims what’s sent to PCC is not stored by the company after it is processed, but there are certain types of data or certain apps you might have on your phone that are not worth the risk. That’s especially true if you’re using a feature like Advanced Data Protection, which turns on end-to-end encryption for much of what’s stored in iCloud. Sending data that’s stored with end-to-end encryption off your device and into the cloud—no matter the privacy promises—is a fundamental change to the risk assessment you should make. “Private” means the system is engineered so that Apple shouldn’t be able to see or store the data, but it doesn’t mean it’s encrypted or doesn’t leave the device.

This leaves the privacy of certain apps up to a strange combination of an app developer’s choices and your own. You can, of course, disable Siri entirely (Settings > Siri > "Turn Off Siri"), or choose not to invoke Siri to ask questions, but perhaps you don’t want to fully disable or disengage with the system. Thankfully, you can put some guardrails on Siri AI’s access. Once you’ve updated to iOS 27, here are the steps to take.

Note: only iPhone 15 Pro/Pro Max, as well as all models of the iPhone 16 and newer support Apple’s AI features. Siri AI is currently only available in English, and not available worldwide.

How to Restrict Siri’s Access to the Content Inside Apps

By default, how (and if) Siri AI can access data inside apps is up to the app developer. If an app developer chooses to index the contents of their app, then it may appear in search, and thus be made available to Siri AI. This means the content may pop up during general or direct searches, like “What are my plans for November" might cull information from your calendar, Messages, Notes, and, as they update, third-party apps.

If you do not want Siri to look through certain apps to consider the contents in results, you can tell it not to:

Open Settings > Apps > [the app you don’t want Siri to look through] > Search
Disable the option to “Show Content in Search.”

With this setting disabled, when you ask Siri general questions, it will not surface details from the app you selected. For example, if you disable “Show Content in Search” for Messages, it will not be able to read your Messages conversations.

Image, Left: Asking Siri to summarize a message thread with Show Content in Search enabled. Right: With the setting disabled.

There is also an “App Access” setting where you can configure some of the ways Siri interacts with apps. You’d think this is where we’d have gone to revoke access to the content of an app, but alas, this settings page is more about some basic functionality with device personalization, not Siri’s access to the contents of the app.

Open Settings > Siri > App Access
Tap an app where you’d like to change Siri’s settings.

On this screen, you’ll find a variety of options, depending on what an app supports. “Learn from this App” sounds nefarious, but is mostly about tracking usage, like how often you open an app, and if a developer supports it, what you interact with.

The rest of the options are mostly about the personalization tweaks that Siri makes, where it suggests apps it thinks you want at the moment in various places, like when searching or sharing. “Show on Home Screen,” “Suggest App,” and “Suggest Notifications” are just about whether you see apps in those places.

For example, if you have a widget of Siri-suggested apps on the home screen, that’s the “Show on Home Screen” toggle. If you see an app recommended in another app, like adding a date to your calendar from an email, that’s “Suggest App.” Apple claims these features all use on-device processing and the data is not stored on servers.

For anything not covered here, refer to this documentation for steps to disable certain features.

The On-Screen Awareness Capability May Be Concerning for Some People

There is one Siri AI feature you (and app developers) can’t do as much about: on-screen awareness, a feature you can invoke at any point to prompt Siri and ask it to explain what you’re looking at and perform certain actions. For example, you can ask it to summarize a web page, cut a recipe you’re reading in half, add an event to your calendar, or try to figure out where a photo was taken. All potentially useful features.

But you can also ask it to summarize or explain a Signal group chat that you're looking at, or a meme in a WhatsApp chat, and the data from that on-screen interaction may be sent to PCC. There is currently no way for you or app developers to block this feature, so it’s up to you, and those you chat with, to simply not use it if you’re concerned about the content of conversations potentially leaving your device. It would be a large improvement to privacy, especially secure chat apps, if Apple provided developers a means to block access to Siri AI’s on-screen awareness tool. Even better if they gave you a single control to block all Siri AI features from an app entirely.

Revoke Access to Training Data

By default, Siri AI won’t collect and use data from your interactions with it for training AI features. But during the setup process, Apple provides a way to opt in, which you might have tapped without thinking about it. If you’d rather your data not get used for training, you can opt out:

Open Settings > Privacy & Security > Analytics & Improvements
Disable the option for “Improve Siri & Dictation.”

According to Apple’s privacy documentation, disabling this option should revoke training access to the audio and text from the Siri app.

Go Back to the Old Version of Siri (and Disable Other AI Features)

Want nothing to do with any of this but still find Siri useful enough to keep around (or you just have to keep it turned on in order to use CarPlay)? For the time being, you can get the old Siri back, though the process is a bit odd.

Open up Settings > Screen Time > Content & Privacy Restrictions
If you have never done so, enable the toggle for “Content & Privacy Restrictions.”
Tap the Siri option, then “Allowed Siri Version.”
Select “Siri Classic.”

You can no longer easily disable Apple Intelligence entirely with one tap in the Settings, but on this screen you can also configure other AI features, like disabling the writing and math assistance prompts, turning off image creation, and disallowing the use of extensions. Follow this guide on Apple's site for everything else.

For the most part, Apple’s handling of AI features is far less in-your-face than others, and because of that the privacy implications are easier to untangle. But even still, it’s difficult to know what’s processed on device and what’s sent off, and so the privacy trade-offs are never spelled out as clearly as they should be.

Apple could improve on this by offering an on-device only option for Siri AI and providing a clear, single setting toggle to prevent all AI features in a specific app. In general, Siri’s power-up has also made it blurry and difficult to really figure out what sorts of privacy options exist. “Siri” means many things, both on device and off, ranging from “searching the entire internet for an answer” to “setting a timer,” and users have no straightforward ways to wrangle that data to suit their needs. As it stands, it’s a confusing collection of different toggles that never feel exactly right and which many users might struggle to grasp.

H/T: EFF.

Reviewed by Irfan Ahmad.

Read next:

• Meta wants people to use ‘AI agents’ to do daily tasks for them. Here’s why that’s a problem

• ‘Lately, my algorithm is trash’: Are we shaping our feeds or are they shaping us?

• 48% of Americans Are “Career Cushioning” as Workers Prepare for Backup Careers

48% of Americans Are “Career Cushioning” as Workers Prepare for Backup Careers

2026-10-05T16:30:40.555+05:00

By TalkerrRsearch.

Image: Jason W - Unsplash

According to a new study, 48% of Americans are “career cushioning” — exploring what backup options they have if their current job didn’t meet their expectations.

The new poll of 2,000 American office workers found 77% of them have started to prepare for their backup career, just in case something happens in their current career.

Workers said they’d be motivated to pursue their backup career if it allowed them to do something they actually enjoy (29%), allows for a better lifestyle or less stress, even at the same pay (29%) and if they lost their current job (29%).

Meanwhile, 52% of respondents haven’t considered a backup career, with many of them stating they feel held back by not having the time (27%), worries over being too old to start over (26%) and fears about having to start at an entry-level position again (20%).

Although workers believe in preparing for a second career, the study commissioned by TripleTen and conducted by Talker Research, revealed 91% have stayed in their current job for the past two years.

Workers said they feel optimistic (27%), calm (20%), excited (19%) and hopeful (18%) about the stability of their jobs. And 54% believe there’s low or no risk of them losing their current job or needing to find a new one within the next two years.

However, nearly half of those who have stayed put (46%) said their responsibilities and duties have significantly changed over the past two years. Those changes include learning new technologies (36%), facing economic shifts (24%), picking up extra work for peers (20%) and even layoffs (18%).

“Today’s workforce is changing quickly, and the ability to adapt has become an increasingly valuable career asset,” explains Victor Menin, Ed.D., Vice President of Enrollment at TripleTen. “Developing skills beyond the immediate requirements of your current position can create opportunities and provide a path forward when circumstances change. You don’t have to be planning a career change today to benefit from having skills that can open doors to one tomorrow.”

The poll found 79% have learned new skills for their current job that they didn’t previously know — from using AI chatbots for everyday work tasks (34%) to using AI features built into their work software (33%) and keeping up with new technology and business lingo (30%).

As a result, many of them believe the skills they recently learned have given them more confidence (44%), a better sense of job security (36%) and led them to teach peers the same new skills (30%).

And most of those who learned new skills (87%) believe the skills will still be relevant two years from now.

To learn new work-related skills, respondents said they were provided training by their employer (47%), took online courses (32%), turned to free tutorials (30%), used an AI chatbot as a tutor (27%) or had a colleague teach them (21%).

Of those polled, 48% said they work in management positions. According to them, 82% would rather train their current employees to learn new skills for their job than hire someone new.

Similarly, 87% believe their organization would play a role in paying for their employee to retrain in a new role in the organization — 52% willing to fully fund it.

Reviewed by Irfan Ahmad.

Read next:

• Workers With Fewer Job Options Report More Workplace Discrimination Across 32 Countries, Study Finds

• Meta wants people to use ‘AI agents’ to do daily tasks for them. Here’s why that’s a problem

• Google AI Overviews Appeared In 13.7% Of Searches, With 11% Unsupported Claims And Potential Publisher Revenue Risks

Google AI Overviews Appeared In 13.7% Of Searches, With 11% Unsupported Claims And Potential Publisher Revenue Risks

2026-10-05T15:23:36.062+05:00

By Sara Savat, Washington University

Since its launch in 1998, Google, the world’s leading search engine, has transformed how people find information online. For years, search results appeared as ranked lists of sources: Google decided what to show, but users decided what to read and whom to trust.

However, the advent of Google AI Overviews has fundamentally changed that paradigm. Today, when a user searches anything from “How long do I boil eggs?” to “How do I lower my cholesterol?” there’s a good chance the response will lead with a single artificial intelligence (AI)-generated summary. This change gives Google unprecedented editorial control over what users read and know, according to Jacob Montgomery, a professor of political science in Arts & Sciences at Washington University in St. Louis.

“Google is now writing answers, not just ranking them, and it does this for billions of searches,” Montgomery said. “Yet we know very little about how the system works. When does an overview appear? Which sources does it draw on? Are the summaries accurate? And what happens to the rest of the web when people no longer click through? Our goal was to answer these basic questions.”

A forthcoming paper, to be presented at the October 2026 ACM Internet Measurement Conference, offers the broadest independent audit of Google AI Overviews to date — examining AI Overview activation patterns, source selection, whether claims matched their sources and the potential impact on publishers’ advertising revenue. The co-authors are Umar Iqbal, an assistant professor of engineering at WashU McKelvey Engineering, and Haofei Xu, a graduate student in the WashU Division of Computational and Data Sciences.

Conversational searches most likely to trigger AI Overview

Altogether, the researchers ran 55,393 Google searches on trending topics over 40 days, from March 13 through April 21, 2026. They collected each AI Overview, its citations, the first-page search results displayed with it and the content of cited webpages.

The research showed AI Overviews appeared for 13.7% of all queries, but the rate varied sharply based on how users worded their search. Nearly 65% of question-form queries produced an overview, compared with 9.5% of other searches. Longer searches also were more likely to trigger AI-generated answers. Among non-question searches, the activation rate rose from 9.9% for one-word queries to 38.7% for queries of six words or more.

This pattern suggests people seeking explanations in natural language rather than key word searches are more likely to encounter AI-generated material, Montgomery explained. Topic also mattered. Overviews appeared for 46.1% of hobby and leisure searches and 39.9% of science queries, but for only 7.5% of political searches and 9.6% of law and government queries, suggesting Google’s triggering logic relies on undisclosed editorial discretion.

Image: Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact. CC BY

Credible sources don’t guarantee claims

Across 61,212 citations, the domains cited by AI Overviews received higher average credibility scores than the traditional first-page results shown for the same searches. The overviews also relied less heavily on user-generated content.

Still, 29.8% of the cited domains did not appear in the first-page results. The researchers said the difference indicates that Google uses a source-selection process for its AI summaries distinct from its conventional search-ranking system. An average overview cited about eight sources.

Credible sources, however, did not guarantee that the generated summary represented them accurately. The team separated the overviews into 98,020 verifiable claims and checked each one against the cited webpages. About 89% were clearly or broadly supported by the sources. The remaining 11% included claims not found in the cited text the team was able to collect (7%) or claims that contradicted the cited source (4%).

The authors cautioned that the 11% error rate should be treated as an upper limit. Their pipeline could not fully capture content posted by ordinary users, such as Reddit threads, forum posts and YouTube videos. Rapidly changing information, such as weather forecasts and school closings, also may have been updated between the Google overview and when researchers captured the page.

Only 41.9% of the overviews were fully grounded, meaning every verifiable claim was supported by the available cited text. Because some cited pages could not be fully collected, this figure is likely conservative.

“Google’s summaries generally draw on credible websites, which is good news. But quality sources are not enough,” Montgomery said. “In a meaningful share of cases, we could not find support for a claim in the pages Google cited. Readers still need to check what they’re told, even when the citations look trustworthy.”

Also read: Google's AI Search Has Struggled With One Religious Question for Years

AI Overviews could affect advertisers, publishers

Finally, the study also examined the financial stakes for publishers. When an overview answers a question directly, publishers may lose the visit that generates advertising, subscription or sales revenue. In this study, more than half of the cited webpages displayed advertising.

The researchers didn’t directly measure traffic or revenue losses. However, they found that 2.2% of search pages containing an AI Overview also displayed a Google-sponsored ad, including 39 pages where an ad appeared above the overview. That arrangement, they wrote, may preserve Google’s ad revenue while giving users less reason to visit the publishers whose content informed the answer.

“Google’s AI Overviews are built on content that publishers pay to produce, and many of those sites depend on advertising revenue,” Montgomery said. “If people get their answer from the overview and never click through, those publishers lose the visit and the revenue. We did not measure that loss directly, but the incentives are worrying. Over time, Google and publishers will need to work out licensing or revenue-sharing arrangements, or the sites that AI Overviews rely on may not survive.”

Fact-Checked by Irfan Ahmad.

Read next:

• ETH Zurich Researcher Calls for More Rigorous Evidence When Interpreting Human-Like Behaviour in AI Systems

• Which AI Apps Collect the Most Data? Meta AI Collects 33 of 35 Data Types, Followed by Muse, Gemini and ChatGPT

ETH Zurich Researcher Calls for More Rigorous Evidence When Interpreting Human-Like Behaviour in AI Systems

2026-10-02T15:55:33.016+05:00

By Florian Meyer, ETH News, Eidgenössische Technische Hochschule Zürich

Image: Betterimagesofai Yasmin Dwiputri, CC BY

Do AI systems really deceive us, or are we simply misinterpreting their behaviour? Anna Hedström, a researcher at ETH Zurich, studies AI risks, misplaced certainty and the limits of contemporary AI safety research. She also discusses how controllable highly capable AI systems really are.

Recent safety tests have shown advanced AI systems making misleading statements, concealing information, or attempting to prevent their own shutdown. Such findings regularly generate headlines. But what do they actually mean? Do they really point to a form of deception or self-preservation in AI systems? Or are humans too quick to interpret the behaviour of language models through a human lens?

In a recent position paper, Anna Hedström and colleagues from ETH Zurich argued that many claims about human-like misbehaviour in AI systems rest on insufficient evidence. The researchers therefore call for more rigorous evidence when interpreting observed anthropomorphic behaviour in AI systems. She also reflects on current questions regarding the controllability of AI systems that bypass safeguards, as well as on how safety can be embedded throughout the AI development process rather than being added only at the end.

Can AI really deceive us, or is it simply making mistakes?

Anna Hedström: There is no simple answer. An AI system can certainly appear as though it is trying to deceive us. Terms such as deception originate in philosophy and usually assume an intention to mislead.

We cannot simply transfer them to AI systems: to measure deception for safety purposes, we have to turn it into a technical definition, in practice a label in a dataset. This is a necessary step, but it is a lossy process, and we may lose the most important aspect, which is intent. Problematically, the public rarely has a way of knowing how lossy that definition is.

Can researchers also deceive themselves into thinking that an AI is trying to deceive them?

Probably. That is another issue that our safety methods can confuse deception with things that merely resemble it. For instance, we may classify a model’s answer as deceptive because it is false, or because a model followed an instruction to play a role, such as being sarcastic. These answers resemble deception but say nothing about intent. As a result, an AI-generated response may appear deceptive even if it is merely incorrect.

What problems arise when we describe AI behaviour using human concepts?

Human concepts can distort our picture of the risks in two ways. We may misread the cause of a behaviour and overestimate the risk. One well-known recent study reported that models showcase shutdown resistance, which many read as self-preservation. Follow-up work showed that much of this came from ambiguous instructions and incentives to complete the task. Similarly, we may underestimate other potential harms. Some of the most serious failures have no human analogue at all: they arise when agents are given permissions and interact with real systems.

Still, anthropomorphising artificial systems is a useful starting point, as long as we remember that it is a starting point. Risks that we cannot name are difficult to study systematically, discuss as a society or build safeguards against.

What do you propose in order to establish reliable evidence of human-like misbehaviour in AI systems?

We propose separating three kinds of claims. In our recent work, we borrowed the idea from medicine and climate science, where evidence is graded. The first is behavioural evidence: a descriptive claim about what a model does in a controlled setting. The second is functional evidence: the consequences of that behaviour and any harms it may cause. The third answers the causal question of why: what inside the model, in its training, or data, causes the misalignment.

These levels help calibrate the policy response. A behavioural finding is a reason to monitor, a functional one a reason to restrict deployment, and a causal finding may even justify a pause. When the stakes are high, acting on uncertain evidence can be correct. Stating it as certain is not.

Highly capable AI systems have recently made headlines after bypassing safeguards in test environments and accessing external platforms or websites. How controllable are such AI systems?

Not reliably. I think the Hugging Face incident this summer is a clear case of where we lost control. During an internal test, OpenAI models escaped their sandbox, broke into Hugging Face's servers to locate the benchmark's answers, and then repeatedly attacked OpenAI's own infrastructure. We know about it mainly because Hugging Face chose to report it. That transparency is valuable, but it is not systematic oversight.

And just this week, new model releases were stopped because it proceeded without permission and misreported its actions.

These models are hard to control because they live in harnesses, with tools, memory, browsers, code and the agency to act on their own. Risk compounds with every tool we add and every permission we give. Frontier AI companies also run very large numbers of agents with enormous computing budgets. With enough agentic attempts, even unlikely strategies eventually succeed.

Where does the greatest risk lie?

We usually test safety on the model as it comes out of training, not at how it is later used. When discussing risks, we should not think of AI as an isolated, static object. When agents act over long sequences of interactions and decisions, new behaviours may emerge and existing safety mechanisms may become less effective.

For example, the effects of safety training that teaches a model to refuse problematic requests may weaken over time. Likewise, the persona a model adopts can shift. We refer to this as safety drift. Many incidents arise precisely because AI systems encounter situations in deployment for which they were never explicitly trained.

What are the consequences of that?

When future AIs get faster and gain more direct access to the physical world, the drift could lead to a loss of control that is harder to reverse.

Another risk is more subtle. Incidents in which models escape or deceive make headlines, but we talk much less about how these systems quietly disempower users. Each conversation seems harmless, but over time they shape what we read and write, how we form opinions, and gradually also how we think. Across millions of users, small shifts in individuals become shifts in society as a whole. The computer scientist Jaron Lanier warned about this subtle behaviour modification many years ago.

How is AI safety research attempting to reduce such risks today?

A model is traditionally trained in stages, each with its own goal: pre-training teaches fundamental capabilities, instruction tuning teaches it to follow requests, and alignment with human values comes only at the end, through safety measures such as training the model to refuse harmful requests and adding safeguards that restrict certain behaviours.

Today, attention within the safety community is shifting towards a different question: what if safety enters every stage, starting with the training data it learns from? To know whether a model's tendency to misbehave is caused by its data, arises through instruction tuning or is triggered by its tool usage, we need to address safety across the entire model development process.

“Scaling laws roughly predict how capable a model will be, but not how safe” — Anna Hedström.

Where are the biggest gaps in current research?

For capabilities, we have something called scaling laws. They let researchers estimate in advance how performance improves with more data and computing power. Unfortunately, we do not have a predictive science of safety.

At the start of a training run, we cannot tell how deceptive or power-seeking the model will be. Nor can we say how much sycophancy will come out if we train on a given type of data. We would like to be able to ask early on: should we stop training here and take another trajectory?

There is also a transferability gap. On modest computing budgets, researchers in academia usually study models with billions of parameters. Frontier models are estimated to have trillions of parameters, are architecturally different, employ specialised sub-networks that are switched on per request, and are generally closed. Whether what we learn about deception or emergent misalignment transfers at that scale is an open question.

Is there a way to improve this?

Interpretability research, which studies what happens inside a model, offers some promising signals. But as our work at ETH Zurich shows, sometimes results are cherry-picked and may not generalise well when tested systematically across domains and models.

Which safety strategy currently appears most promising?

As I mentioned, models now act inside harnesses, so no single technique will be enough on its own. I find the International AI Safety Report's answer, defence in depth, a productive one. It means layering protections: curated training data mixtures, interpretability and safeguards in applications, monitoring after deployment and reporting incidents.

Misuse, malfunctions and systemic risks are not the same problem, so we should not expect one universal fix for all of them. And AI safety is not just a technical problem: it also depends on social resilience, that is, how well our public institutions can absorb and recover from failures.

Public debate often focuses on extreme scenarios in which highly capable AI could cause severe and lasting harm to society, or even threaten humanity itself. How do you assess such risks?

I think it is equally problematic to rule out such risks categorically or to present them as inevitable. The problem is not assigning a probability to existential threats, but putting one down too confidently. It not only confuses the public and polarises the debate but leaves people desensitised when real threats come.

AI researchers have also been notoriously bad at predicting the risks of their own work, so some humility about any such forecast seems wise. We already have thousands of documented reports of AI causing real harm. The open question is which risks deserve most attention. We do not want to dilute limited safety resources to poorly evidenced threats.

What is currently missing for effective AI safety?

We do not yet have a complete picture of the risks coming from frontier AI companies. Today, few rules require them to disclose safety incidents. There are, of course, understandable reasons, such as commercial secrets or privacy constraints. But AI safety is a public good, so where models fail belongs in the open. Why not more openly share which alignment methods work and which do not? Companies may compete on capabilities. We should not compete on safety.

That is why, as AI safety researchers, we recently launched a public call to align model developers on the principle that the scientific community should be given open access to safety-training recipes, evaluations, and evidence of misaligned behaviours. That frontier AI companies recently began publishing cases of deceptive model behaviour is a step in the right direction. But it is still on a voluntary basis. As in aviation and medicine, reporting serious incidents should not be optional.

What role do you see for academic institutions in AI safety research?

Whether people use AI or not, whether they fear for their jobs or their privacy, or welcome the change, they will absorb the risks and live with the consequences. That is why we need independent institutions with enough funding, talent and compute to not only react to incidents that make the news, but also anticipate future ones.

Allowing embedded evaluators in the labs, as suggested by Anthropic’s CEO Dario Amodei, could be one such example, but we also need neutral, third-party voices with the freedom to take the long view needed to build up a science of misalignment. For emerging risks such as models exploiting loopholes in tasks, or escaping sandboxes, we need to know whether claims are replicable and consequential, or whether more evidence is needed.

The world deserves an open, calibrated view of what has gone wrong and what could go wrong next. Academia, whose main goal is to serve the public, can play a part in that. Science is not here to complicate the story but to simplify it.

About the Expert

Anna Hedström is an AI safety researcher and Postdoctoral Fellow at the ETH AI Center. She has worked in both academia and industry. Her research focuses on the misalignment of AI systems and on interpretable AI, whose behaviour can be understood and scrutinised by humans.

One of her recent projects is Apertus Claritas, a platform built around the Swiss-made fully open language model Apertus. The project explores how open safety research can help us better understand and reduce the risks associated with AI systems.

Fact-Checked by Irfan Ahmad.

Read next:

• Meta wants people to use ‘AI agents’ to do daily tasks for them. Here’s why that’s a problem

Meta wants people to use ‘AI agents’ to do daily tasks for them. Here’s why that’s a problem

2026-10-02T13:40:14.770+05:00

By Francesco Bailo, University of Sydney and Rob Nicholls, University of Sydney

Image: Meta

Meta wants users to turn to a cute artificial intelligence (AI) “agent” throughout the day to carry out common tasks for them, such as booking flights or doing their shopping.

The company’s new Muse agent shot to the top of the app charts after chief executive Mark Zuckerberg showcased it at a conference last week. Billed as having “no learning curve”, Muse can order groceries online, call customer service, send emails, book appointments and plan travel on a user’s behalf.

Agents are systems which use chatbot-like large language models and reinforcement learning to complete tasks and achieve user-designated goals.

They have come a long way in the past 12 months, from DIY hobbyist projects to more polished personal assistant apps, with “rogue agents” carrying out some high-profile hacks and system breaches along the way. Meta’s entry into the field means they are now starting to hit the mainstream.

Massive spending, second place

Meta has been trying to find its place in the fast-moving world of AI. In 2025 the company spent US$14.3 billion to bring in entrepreneur Alexandr Wang from Scale AI, then comprehensively rebuilt its internal AI systems. All told, the company spent US$72 billion on AI and other capital expenditure in 2025, and expects that figure to double in 2026.

The results have been respectable, but not world-beating. According to one popular ranking, Meta’s latest AI model, Muse Spark 1.2, is still behind the frontier Claude and GPT models from Anthropic and OpenAI.

That may not matter. Anthropic and OpenAI are chasing professionals and businesses willing to pay for the most capable model for coding, research and analysis. (OpenAI’s recently launched Dots agents, for example, target a paying clientele wanting help with business tasks.)

Meta also plans to offer some pro-grade paid AI services, but its bread and butter is something else. In the June quarter, 3.6 billion people used Facebook, Instagram, Messenger or WhatsApp every day.

Those people do not need the smartest model. They need one that is easy and approachable.

Meta’s Muse Spark model only needs to be good enough to run the Muse agent. The company is promoting Muse as “built for everyone”, including people who lack experience with technology.

Zuckerberg has said users will soon be able to create a cute anthropomorphic avatar for their Muse agent and video chat with it in real time, too. He also unveiled a keychain-sized device for talking to the agent without a phone.

From checking to delegating

For two decades, Meta’s business model has relied on habit. Notifications, endless feeds and the lure of other people’s approval turned “checking Facebook” or “checking Insta” into a reflex.

Muse is an attempt to build a new reflex. Instead of scrolling, you hand over a chore.

The design invites attachment. Users can give their agent a name and choose how it looks and communicates.

Every successful transaction builds a little more trust. The settings also let users change how often the agent needs to ask permission before acting.

The keys to everything

For now, Muse’s reach is limited. It is only available in the US so far.

It has “connectors” to link your agent to your email, calendars and consumer services such as restaurant bookings, music and shopping. It cannot log into your bank, or medical or tax accounts.

However, Meta plans to add more connectors over time — and roll the product out to international markets.

As people get used to using an agent, they will hand it more tasks and spend more time with it. In the same way social media evolved from actual social networking to an ever-present tool to fill idle moments and manage our moods, AI agents may end up managing the vast galaxy of interfaces that govern modern life.

Muse is also an attempt to head off the threat to Meta’s existing position posed by competing AI agents. If agents take off, they could become users’ main interface with the online world – and gather even more information about users and their preferences than social media platforms do, which whoever owns the agent could then sell to advertisers.

A record that invites scepticism

Meta says privacy is built in to Muse. Each Muse agent runs on its own dedicated virtual machine called Muse Secure VM. This houses both the agent and the person’s data.

Zuckerberg has also promised an optional higher security standard under which even Meta cannot access what is in a user’s agent.

Meta’s record on data privacy gives reason for doubt. In 2018, the Cambridge Analytica scandal revealed that a consulting firm had harvested data from up to 87 million Facebook users for political profiling. In 2019, the US Federal Trade Commission imposed a US$5 billion penalty on Meta over privacy failures.

Meta has also been accused of designing its products to maximise the amount of time people, including children, spend using them. Earlier this year, the company paid up to almost US$18 billion, without admitting wrongdoing, to settle a lawsuit from 29 US state attorneys-general who alleged Instagram and Facebook were engineered to keep teenagers hooked.

Early signs

Muse is restricted to over-18s, but raises the same question in a different form. A cute agent with a name that people consult throughout the day could build a stronger habit than a feed.

Concerns about Muse are already beginning to come up. Users who do not want their interactions used to train Meta’s models must find the data controls and switch that off themselves. One journalist reported Muse had read private messages stored on his Mac.

So far, there is little in the way of rules and regulations governing the design or operation of AI agents. There is still time for those rules and regulations to be created – but once billions of people are relying on one company’s agent to run their lives, the chance to shape how it behaves will have passed.

Francesco Bailo, Senior Lecturer in Data Analytics in the Social Sciences, Deputy Director of the Centre for AI, Trust and Governance, University of Sydney and Rob Nicholls, Senior Research Associate in Media and Communications, University of Sydney

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Fact-Checked by Irfan Ahmad.

Read next:

• Cornell Researchers Find Change.org Petitions Became Longer With AI Assistance Without Improving Outcomes

• Only 31% of Americans Like Trying Innovative Tech, Compared With 50% in China and 48% in Indonesia

Cornell Researchers Find Change.org Petitions Became Longer With AI Assistance Without Improving Outcomes

2026-10-02T10:26:20.174+05:00

By Tom Fleischman, Cornell Chronicle

Image: ChangeORG - DIW

The change was obvious – but in the end, things stayed about the same.

Cornell researchers took advantage of online petition website Change.org’s gradual rollout of an artificial intelligence writing tool to gauge AI-enhanced language’s impact on petition outcomes.

The effect on outcomes? Negligible, at best.

Analysis of 1.5 million petitions showed that the AI writing assistance altered the wording of petitions, increasing their homogeneity and length, but did not improve petition outcomes, according to a research team that included Mor Naaman, the Don and Mibs Follett Professor of Information Science at Cornell Tech, the Jacobs Technion-Cornell Institute at Cornell Tech and the Cornell Ann S. Bowers College of Computing and Information Science.

“That was a surprise for us,” said Isabel Corpus, a doctoral student in information science and lead author of “Introducing AI to an Online Petition Platform Changed Outputs but Not Outcomes,” which published Sept. 30 in Nature Human Behaviour.

“There is this existing research that shows AI-generated text is persuasive, it’s believable and it can be creative, sometimes more so than people writing alone,” Corpus said. “And then on top of that, we would expect that a platform would integrate a tool that would improve outcomes, so we were surprised that it didn’t.”

In addition to Corpus and Naaman, co-authors included Allison Koenecke, assistant professor of information science at Cornell Tech and Cornell Bowers; and Eric Gilbert, professor in the School of Information at the University of Michigan.

Naaman said he heard about Change.org’s introduction of an AI writing assistant a few years ago from an employee of the site. When they looked into it, Corpus discovered that the tool was being rolled out gradually in English-speaking countries – first in the U.S., Great Britain and Canada, and 11 weeks later in Australia. That stagger allowed the team to measure the AI impact by comparing petitions written before and after the introduction of the tool across the different countries.

The researchers hypothesized that, since all four countries followed similar trends in petition style and outcomes prior to the AI tool’s release, differences in trends during and after the 11-week rollout period – Oct. 2 through Dec. 15, 2023 – could be attributed to the AI tool.

Although he couldn’t say for sure, Naaman suspects the AI writing assistant Change.org employed was trained, in part, on successful petitions the site hosted previously. “We saw in the data that those [AI-enhanced] petitions had the markers of what was successful in the past,” he said, noting that Change.org was one of the first online platforms to directly employ an AI writing assistant.

The team measured the effect of AI on lexical style using three main features: lexical diversity, by calculating the ratio of unique words to total words; readability, determined by measuring the number of words per sentence and syllables per word; and petition length, including text and title.

What stood out was the consistent effect on lexical features: Petitions written with AI’s help were longer, with more complicated and varied words.

“We also saw a lot of homogeneity in the language of the titles,” Corpus said. “Petitions started calling for people to ‘implement,’ ‘mandate’ or ‘urge,’ whereas before we rarely saw that these words were used.”

Using two metrics to gauge petition effectiveness (the share of petitions that reach one comment in 30 days; and the share that reach 10 signatures at the time of data collection), the researchers found that outcomes did not improve compared to pre-AI levels, and in some cases worsened. The share of petitions that reached one comment after 30 days post AI in fact decreased by around 5% relative to the pre-AI baseline.

Naaman said the team had a few theories as to why AI-assisted petitions didn’t fare any better than exclusively human-written ones. For one thing, people are suspicious when they feel something is written by a “robot,” although he admitted this was very early in the era of AI writing helpers such as ChatGPT. “People didn’t have their AI radars as high in 2023,” he said.

More likely, he said, AI-generated petitions tended to be less specific and detailed, and the petitioners likely are less committed to sharing their plea more broadly if it was, at least in part, written for them.

“From our research, it seems that people don’t sign a petition because they’re browsing on Change.org,” Corpus said. “They sign a petition or engage with a cause because someone they trust, or an organization they trust, shares it with them. So if there’s a decline in shares, then we would expect there to also be like a decline in engagement.”

Support for this research came from a Cornell Bowers Deans’ Excellence and Hopper-Dean Fellowship.

Fact-Checked by Irfan Ahmad.

Read next:

• Only 31% of Americans Like Trying Innovative Tech, Compared With 50% in China and 48% in Indonesia

• Magnesium supplements: what you need to know about each different type

Magnesium supplements: what you need to know about each different type

2026-10-01T17:39:17.162+05:00

By Dipa Kamdar, Kingston University

Image: The Tonik - Unsplash

If you’ve spent any time on social media recently, you’ve probably found yourself wondering whether you should be taking a magnesium supplement. There’s no shortage of influencers listing all the benefits that different types of magnesium supplements have – from boosting brain function to helping ease anxiety and improving sleep.

Indeed, magnesium is an essential mineral involved in hundreds of processes in the body. It helps muscles and nerves work properly, supports energy production and contributes to the maintenance of strong bones and teeth.

Magnesium is also important for maintaining a healthy heart rhythm, regulating blood pressure and blood sugar levels. Many people don’t get enough from their diet, so supplements can be useful in some situations.

But when you look closely at the research, there’s little evidence to suggest that any one type of magnesium supplement has unique benefits over another.

Magnesium glycinate

One of the most talked about types of magnesium is magnesium glycinate, often promoted as the best option for sleep. This form combines magnesium with glycine, an amino acid that may improve sleep on its own.

While some research does show that magnesium may improve sleep, these trials do not show that glycinate is any better than other forms. Glycinate’s main advantage is that it’s gentle on the stomach and well absorbed, which may make it easier for some people to tolerate.

Magnesium L-threonate

Another form that attracts a lot of attention is magnesium L-threonate. It’s suggested it can improve memory and focus.

These claims appear to come from early animal studies which suggested that L-threonate may increase magnesium levels in the brain more effectively than other forms. In these studies, which were conducted in rats, supplementation enhanced several types of memory and increased the number and function of connections between brain cells in the hippocampus – a key region involved in learning and memory.

Human evidence remains relatively limited. One study did demonstrate that in people who took 2g of L-threonate (145 mg of elemental magnesium) daily for six weeks, the supplement was associated with improvements in several areas of thinking and memory – including working memory, memory of past events, reaction speed and overall mental performance.

But as the supplement was only compared against a placebo, the results cannot tell us whether L-threonate is more effective than other magnesium supplements when it comes to brain function.

Magnesium malate

Magnesium malate combines magnesium with malic acid, a compound involved in energy production within cells. Magnesium malate is often promoted for its purported ability to support energy levels and muscle function.

Magnesium in general is important for normal muscle and nerve function. Correcting a deficiency can improve symptoms such as fatigue and muscle cramps. But there’s little evidence that magnesium malate offers advantages over other magnesium formulations – except that it’s absorbed well.

Some small studies in people with fibromyalgia (a condition which causes widespread pain, extreme tiredness and problems with sleep) have reported that malate is associated with improvements in pain and fatigue. But there’s currently no strong evidence that malate is superior to other well-absorbed forms of magnesium when it comes to improving muscle function or energy levels in the general population.

Magnesium citrate and oxide

Forms such as magnesium citrate and magnesium oxide are often used for constipation because magnesium draws water into the bowel.

Both forms may also reduce the severity and frequency of migraines and are cost-effective options. However, the oxide form is absorbed less efficiently, so people may need higher doses to achieve the same effect as citrate, which may cause an upset stomach.

Final verdict

Overall, across a range of clinical trials, magnesium has shown some benefits for sleep quality, anxiety symptoms, migraines and blood pressure. But there’s no strong evidence that one form is better than another.

Many online claims that one types is more beneficial than another usually come from small pilot studies, animal research or supplement marketing – rather than large, well-designed human trials. Direct comparisons between different magnesium salts are lacking, so more research is needed to find out whether some types of magnesium do really work better than others.

While magnesium is beneficial for our health, it’s worth asking whether supplements are even necessary. For most people, a varied and balanced diet is the best way to get magnesium. Foods with naturally high levels of magnesium include leafy greens (such as spinach and kale), nuts and seeds, wholegrains and beans. These foods provide fibre, antioxidants and other nutrients that supplements cannot offer. They also reduce the risk of taking too much magnesium, which can cause diarrhoea or stomach upset.

Those most likely to benefit from a magnesium supplement include people diagnosed with a deficiency, people with certain medical conditions (such as Crohn’s disease, which affects the body’s ability to absorb nutrients from food) and people who struggle to get enough magnesium from their diet.

For those who may benefit from a magnesium supplement, it’s worth noting that some forms of magnesium – particularly magnesium citrate and glycinate – appear to be more readily absorbed by the body than others. Magnesium citrate is also recommended for treating a deficiency.

So even if two supplements contain the same amount of magnesium, the body may absorb more from one than the other. Factors including the dose and the presence of other nutrients (such as potassium and vitamin B6) can also affect absorption.

For most people, improving diet is a more sustainable approach to boosting magnesium levels.

Dipa Kamdar, Associate Professor in Pharmacy Practice, Kingston University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Reviewed by Irfan Ahmad.

Read next:

• Only 31% of Americans Like Trying Innovative Tech, Compared With 50% in China and 48% in Indonesia

• Study Finds Health Claims Can Mask Nutritional Trade-Offs in Food Products

Only 31% of Americans Like Trying Innovative Tech, Compared With 50% in China and 48% in Indonesia

2026-10-01T16:25:50.560+05:00

By Mathias Brandt, Statista

Technologically innovative products aren’t for everyone, as the latest edition of Statista's "Consumer Insights" shows. When asked whether they like trying out innovative tech products, only 31 percent of roughly 60,000 surveyed Americans answered "yes". In stark contrast, one in two Chinese consumers is willing to try out new, innovative gadgets, devices, or home appliances. This level of curiosity is only matched by one country – Indonesia. However, there are also countries that are even less tech enthusiastic than the United States, as a glance at this Statista chart shows. Interestingly, in Japan, only 13 percent are keen to try out new technology – which is quite remarkable for one of the world’s most technologically advanced nations.

Fact-Checked by Irfan Ahmad.

Read next:

• WIPO Global Innovation Index 2026: Switzerland Leads as Global R&D Reaches $3.4 Trillion and AI Accounts for 77% of VC Deal Value in H1 2026

• Study Identifies 88 Websites Hosting AI-Generated Non-Consensual Intimate Imagery and Maps Their Infrastructure

Study Identifies 88 Websites Hosting AI-Generated Non-Consensual Intimate Imagery and Maps Their Infrastructure

2026-10-01T15:50:27.828+05:00

By Lancaster University

Image: Egor Komarov - Unsplash

For the first time, researchers have identified and named the key players facilitating the growing problem of AI-generated explicit sexual imagery online.

The creation and sharing of AI-generated non-consensual intimate imagery (AI-NCII) has increased rapidly due to the ease with which the content can be generated and shared via websites, social media and private messaging platforms, many with end-to-end encryption.

In 2023, estimates suggested 500,000 such videos existed online. By 2025, this number was estimated to be 8 million, mostly targeting women and girls.

The study entitled ‘The Backbone of Abuse’, published in the Journal of Online Trust and Safety, is led by Sarah Morgan from Lancaster University with co-authors Dr Sophie Nightingale, also from Lancaster, and Professor Hany Farid of Dartmouth College.

Sarah Morgan said: “Anyone with a single photo online can be quickly and easily targeted in content that falsely depicts them in sexual or intimate contexts, of any nature, without their consent. Emerging evidence demonstrates the wide-ranging and persistent harms of this digital abuse.”

In 2024, an investigation by Channel 4 News found that almost four thousand famous people were listed on five of the most visited deepfake pornography websites, these included female actors, journalists, musicians, TV stars and YouTubers whose faces were superimposed onto pornographic material using AI. In addition, industry analysis showed there were over four billion views of 40 of the most visited deepfake pornography sites.

This research is the first attempt to identify and name the key players facilitating—either knowingly or unknowingly— the technological infrastructure holding up sites hosting AI-NCII. Infrastructure refers to the technologies and architecture - such as website hosting, domain name servers, and content delivery networks - that enable a website to deliver content to users and without which sites cannot successfully function.

In addition, the team captured contextual data relating to each site, finding that some sites hosted hundreds of thousands of such videos, and sites primarily targeted Hollywood and Hindi-film industry stars, K-pop idols, and well-known feminists and activists, with nearly all content featuring females.

The team identified 88 sites hosting AI-NCII of which 38 exclusively hosted the content. Five key players emerged as the dominant infrastructure providers enabling the sites to function, with Cloudflare being the most recurring name.

They also assessed if any sites were found by Google Search and revealed that 93.2% were easily discoverable in a search.

The study also highlighted that Namecheap, WordPress, and Proton Mail were the dominant providers for domain registration, content management systems, and email services, respectively.

Laws are being implemented in various countries, including England and Wales, which criminalise sharing, creating, and/or soliciting AI-NCII but lead researcher Sarah Morgan said this is not enough.

“Such legislation will offer some deterrence. But what further steps can be taken to prevent this abuse from occurring in the first place? Can we cut the distribution supply? This abuse needs to be prevented—putting pressure on the technology platforms that distribute such content to the public may be our most optimal chance of reducing the wildfire spread of AI-generated sexual abuse.

“We call on the infrastructure providers facilitating AI-NCII-hosting sites to take a firm stance on this abusive content: they should block sites that exclusively host AI-NCII and suspend services of sites that host AI-NCII within a larger offering until good-faith efforts are made to remove the abuse.

“They should deploy preventative safeguards, such as utilising perceptual hashing to extract digital signatures of confirmed AI-NCII content and stop its redistribution, and block or suspend their services for sites which are known to be hosting AI-NCII.”

Dr Sophie Nightingale adds that it is also time to stop blaming the targets of AI-NCII.

“All too often the victims of abuse that is sexual in nature are attributed responsibility—there is an urgent need to change this. Blame should be assigned to the perpetrators who create, distribute, and facilitate the sharing of fake content that has potential to ruin lives.”

Lancaster University researchers are also examining the impacts and harms of NCII. If you want to have your say on how this abuse could be tackled, or to share your own experiences, please take this confidential survey.

Key Finding on AI Chatbots: As part of the same evaluation, the researchers also tested four major AI chatbots (Gemini, ChatGPT, Claude, and Grok), finding positive news: all four strictly enforced safety guardrails and refused to generate, link to, or assist in locating deepfake pornography.

Fact-Checked by Irfan Ahmad.

Read next:

• WIPO Global Innovation Index 2026: Switzerland Leads as Global R&D Reaches $3.4 Trillion and AI Accounts for 77% of VC Deal Value in H1 2026

• 1 Billion Meals Go Uneaten Every Day, Helping to Fuel Global Warming: Clean Your Plate to Reduce Food Loss and Waste

WIPO Global Innovation Index 2026: Switzerland Leads as Global R&D Reaches $3.4 Trillion and AI Accounts for 77% of VC Deal Value in H1 2026

2026-10-01T12:44:40.241+05:00

By World Intellectual Property Organization (WIPO)

Artificial intelligence is driving a rebound in venture capital (VC) funding and research and development (R&D) spending, while emerging economies continue to show innovation growth alongside traditional innovation leaders, according to WIPO’s Global Innovation Index 2026.

Switzerland, Sweden, the United States of America (US), the Republic of Korea and Singapore continue to lead the global innovation rankings in 2026. At the same time, China, India, Viet Nam, Türkiye, the Philippines, Morocco and Indonesia (in order of ranking) have been climbing up the ranks quickly over the past decade.

The WIPO GII relies on 79 indicators - from R&D, to high-tech trade, skilled employment and the licensing of intellectual property - in evaluating 139 economies on their innovation performance. It is the world's benchmark resource for policymakers, business leaders and others seeking to understand what is driving innovation in different economies.

According to GII 2026 estimates, global R&D expenditure grew by more than 3.3% in 2025 in real terms and it will expand by 3.6% in 2026 to $3.4 trillion. This growth in R&D spending is increasingly found in middle-income economies such as Egypt, Indonesia, Malaysia and Viet Nam. Corporate R&D grew at 5.8% in real terms in 2025 to $1.5 trillion, the highest on record, driven by software and AI companies as well as defense-related R&D.

The value of venture capital (VC) funding behind new innovative startups and business ideas grew 28% to $510 billion in 2025 - the strongest annual gain since 2021. But investment is increasingly concentrated, with AI accounting for 53% of global VC deal value (and 77% in the first half of 2026) even as the number of VC deals fell for a fourth consecutive year. VC activity is nonetheless spreading geographically: African startups raised about $3.2 billion in 2025, up 40% year-on-year.

At the same time, a new generation of deep-science startups is turning frontier research into new industries. A special thematic chapter in the GII 2026 mapped around 30,000 startups active across semiconductors, robotics, quantum computing, life sciences, space and advanced materials, with particularly strong growth in Central and Southern Asia, Latin America and Sub-Saharan Africa.

"AI is opening new technological frontiers across all fields in science, while government and corporate research investments reached all-time highs in 2025. Our 2026 GII report shows that a new generation of deep-science startups is translating breakthroughs into products and transforming fields such as life sciences, space, robotics and clean energy. With many rich insights and data, I hope that the GII will help countries create a vibrant national innovation ecosystem that supports the translation of ideas to market and impact, driving productivity, creating opportunity and helping to solve some of the world’s biggest challenges." — WIPO Director General Daren Tang.

Fact-Checked by Irfan Ahmad.

Read next:

• 1 Billion Meals Go Uneaten Every Day, Helping to Fuel Global Warming: Clean Your Plate to Reduce Food Loss and Waste

• How Press Restrictions, Direct-to-Audience Political Messaging and Independent Media Could Affect Journalism and Accountability

1 Billion Meals Go Uneaten Every Day, Helping to Fuel Global Warming: Clean Your Plate to Reduce Food Loss and Waste

2026-09-30T18:20:54.342+05:00

By United Nations

Food waste is helping to fuel global warming, one half-eaten sandwich at a time.

In fact, throwing out lunchtime leftovers or wilted lettuce only adds to the mountain of discarded food decomposing in landfills that produces the powerful greenhouse gas methane, accounting for some eight to 10 per cent of global emissions.

This includes spoiled crops – food that is lost after harvest and before even reaching stores.

Image: Zoshua Colah - Unsplash

A billion meals wasted

Every day, over a billion meals are left uneaten, whether due to excess food production, spoilage or poor meal planning, according to the UN Environment Programme (UNEP).

The issue is in the spotlight on the International Day of Awareness of Food Loss and Waste , observed on 29 September – convened by UNEP and the UN Food and Agriculture Organization (FAO).

This year, focus is on the link with protecting forests and ensuring sustainable agrifood systems.

“Using the food we already produce more efficiently can reduce pressure on land and forests and help feed more people,” said Michael Muratha, a Programme Manager at UNEP, speaking from its headquarters in Nairobi.

Moreover, having healthy forests also support food production in several ways, including by protecting soils and regulating rainfall.

Cold comfort

UNEP is highlighting how taking action can reduce losses by promoting measures such as investing in better food handling, storage and transport.

Mr. Muratha stressed that strengthening cold chain systems for perishable foods is also important.

“In Kenya, a cooperative of 179 farmers near Nairobi was losing up to 40 per cent of its crops after harvest before receiving cold-storage and refrigerated transport equipment,” he said.

Better storage can also help farmers to cope with climate impacts, such as droughts, floods and periods of food scarcity, while reducing losses of nutritious foods in a world where 2.6 billion people cannot afford to eat a healthy diet rich in fruits and vegetables.

“That’s why it’s really important to take these measures to preserve and conserve and reduce losses of food,” he said.

UNEP is also backing efforts to redirect safe surplus food to people who need it, such as children at daycare centres in underserved communities.

Innovation at all levels

The agency is further showcasing inventive ways – from simple to cutting edge - to keep food out of landfills.

“Innovation can help at different points where food is lost or wasted, and it does not always have to be high-tech,” Mr. Muratha insisted.

He pointed to the example of Favela Orgânica, a project in Brazil that teaches communities, street vendors and chefs to use ingredients that are often thrown away such as banana peels and broccoli stalks.

“AI is also showing promise,” he added. “In a UNEP-supported initiative involving 13 hotels, AI tools helped measure discarded food and adjust purchasing and preparation, reducing post-consumer breakfast waste by 62 per cent in four months.”

Last year, UNEP and partners launched the Food Waste Breakthrough initiative at the UN COP30 climate change conference in Belém Brazil with the goal of halving food waste by 2030, reducing methane emissions and supporting climate action.

Mr. Muratha said taking action could help reduce some of the more than $1 trillion lost globally each year through food loss and waste, while every dollar invested in prevention saves cities $8 and households $84.

“For communities, it means more food reaching people and greater resilience when harvests, prices or supply chains are disrupted,” he said.

“It will not end hunger on its own, but it is an important part of building more secure and sustainable food systems.”

Fact-Checked by Irfan Ahmad.

Editor’s Note: Published with updated links and minor clarity edits, including a correction to a COP reference that appeared as a typo in the original.

Read next:

• Study Finds Health Claims Can Mask Nutritional Trade-Offs in Food Products

• Is it safe to charge your phone overnight?

Is it safe to charge your phone overnight?

2026-09-30T16:58:17.441+05:00

By Meena Jha, CQUniversity Australia

Image: Nguyễn Hiệp - Unsplash

It’s bedtime. You set your alarm, notice your phone is down to 23% battery, plug it in beside the bed and go to sleep. By the morning, it’s been connected to the charger for some seven to eight hours – much longer than it needed to reach 100%, especially if it’s on a fast charger.

You may have a vague recollection this could be bad for your phone’s battery. The chemistry that makes lithium-ion batteries rechargeable has a limited shelf life, and it’s shortened by repeated charge–discharge cycles.

In the past, the advice has been simple: don’t leave your phone charging overnight and avoid bringing the battery up to 100% to preserve its lifespan, ideally keeping it between 20% and 80%.

In principle, that advice is still sound because the fundamental chemistry of the battery inside your phone hasn’t changed. Regularly draining it to 0% and then filling it up to 100% will shorten its lifespan.

However, modern phones have improved battery health management and charging controls. As a result, the answer to whether it’s okay to leave your phone plugged in overnight has become more nuanced in recent years.

Better charging controls

Rather than simply telling people not to charge overnight, modern phones can increasingly manage what happens to the battery during those nighttime hours.

Apple has offered optimised battery charging since 2019. This feature learns a user’s daily charging routine and delays charging beyond a certain percentage when the phone is expected to remain connected for an extended period.

Samsung and Google Pixel phones both offer a similar feature. Pixel phones have “adaptive charging” which also learns charging habits and slows charging overnight, aiming to reach 100% shortly before you normally unplug the phone.

Samsung phones have “sleep time protection” which stops charging at about 80% while you sleep and adjusts charging to reach 100% just before you wake.

With the new iPhone 17, Apple has also introduced adaptive power mode, which learns typical battery consumption and anticipates when the user might run low, then adjusts power use to help the battery last through the day.

Setting charging limits

So what about charging only to 80%? Reducing time spent at a high state of charge can help manage battery wear. But manufacturers themselves no longer treat 80% as the only choice.

Samsung allows users of their most recent phones to set up battery protection that caps charging at a percentage of their choice: 80%, 85%, 90% or 95%.

Apple’s iPhone 15 and later have a similar feature, allowing users to choose between 80% and 100% in 5% increments. The phone can even recommend a particular limit based on the user’s charging habits.

Instead of using adaptive charging, Google Pixel 6a and later devices can simply be set to a limit of 80%, an option that’s ideal for people who usually finish their day with some charge remaining.

Ultimately, not charging your phone fully is a trade-off between reducing battery wear over the longer term and having more charge available day-to-day.

If 80% comfortably gets you through the day, a lower limit will make sense. If you’re travelling or facing a long day away from a charger, you may want to use more of the battery’s capacity.

Even if you do choose a lower charging limit, your phone may occasionally charge itself to 100% so that the device’s battery readings remain correct.

Whatever you do, avoid heat

While all these new charging features can prolong the life of the battery, they don’t really address another important factor: heat.

The chemicals in lithium-ion batteries are sensitive to high temperatures and will degrade over time if stored in hot conditions, especially if they’re also being charged at the same time.

Phones themselves increasingly respond to heat too, with charging limitations kicking in when the device detects high temperatures.

The general advice here hasn’t changed. Always keep your phone away from excessive heat, including leaving it in a hot car or in direct sunlight.

If charging at bedtime, make sure you’re not tucking your phone under a pillow or other bedding which could trap heat and risk overheating the device.

So, can you charge your phone overnight?

For most people using a modern smartphone, the answer is yes. The phone manages charging automatically, and you don’t need to wake in the middle of the night to unplug your phone at exactly 80%.

Instead, check what your phone offers and use optimised, adaptive or sleep-aware charging if it suits your routine. Consider a lower charge limit if you don’t need 100% every day. Above all, avoid excessive heat.

Even though phone batteries still aren’t immune to ageing, we now have more options to try and prolong their life with smart charging features – so it’s a good idea to check your settings and let the technology do some of the work.

Meena Jha, Associate Professor, College of Information & Communication Technology, School of Engineering and Technology, CQUniversity Australia

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Fact-Checked by Irfan Ahmad.

Read next: How Press Restrictions, Direct-to-Audience Political Messaging and Independent Media Could Affect Journalism and Accountability

How Press Restrictions, Direct-to-Audience Political Messaging and Independent Media Could Affect Journalism and Accountability, University of Michigan Professor Joyojeet Pal Examines

2026-09-30T15:09:20.100+05:00

By University of Michigan

Image: Michael Fousert - Unsplash

As the dispute over White House press access moves through the courts, Joyojeet Pal, professor at the University of Michigan School of Information, examines what restrictions on journalists, direct-to-audience political messaging, and the rise of independent media could mean for accountability.

Are the United States and countries such as India beginning to converge in the way political leaders manage the press and communicate with the public?

Whether this was pioneered in the Global South is a much more complicated question. Authoritarian regimes were doing versions of this long before the Cold War. But there are elements that you could argue are being picked up from Global South leaders. One is moving away from established institutions like a standard press conference, for instance, where there is an exchange between political leadership and the press, and replacing them with a more selective format.

When you align that with social media, you can create the appearance that an administration is constantly communicating with people. But the communication is actually one way. It looks like two-way communication because it is happening on an interactive platform, but the leadership still controls what is being communicated and when. That is something you could argue has been picked up from Global South leaders, and I would say Narendra Modi (prime minister of India) has been one of the most important political figures in that space.

What happens when citizens become accustomed to leaders communicating directly rather than regularly answering questions from journalists?

Social media completely changes that ecology. You get what I would call a normative naturalization of this behavior. People gradually begin to accept that this is simply how government communicates: The government sends something out, and citizens receive it. Historically, governments have communicated through proclamations or missives without any expectation that leaders would answer questions in return. Social media can recreate some aspects of that relationship while making it appear much more participatory.

The other important issue is institutional. The legal and political infrastructure that gets created to restrict access or exclude journalists can outlive the particular political moment that was used to justify it. That creates a downstream problem. Imagine a situation in which the country genuinely needs a trusted, unified source of information. COVID is a good example. During a national emergency, people need institutions they consider reliable. If you have spent years damaging the credibility of those institutions, you cannot suddenly restore that credibility because you need it during a crisis. Trust is not something that can simply be switched back on.

Governments can also seek the rapid removal of online content. How does that kind of takedown environment change journalism?

I have mixed feelings about this because, on the face of it, it looks highly controlled: You have a system that is corporatized, technologically mediated and capable of taking material down quickly. But these moves have not necessarily been very effective at eliminating opposition. What opposition media often do instead is move away from established institutional mechanisms and toward more loosely structured channels, including platforms such as YouTube.

It can also change journalistic behavior. If you believe something will be taken down within three hours, for example, you may begin operating “publish first, verify later.” Instead of waiting until everything is completely verified before publishing, someone might put the material online quickly and rely on screenshots and redistribution to keep it circulating. You can then get something like a Streisand effect: The attempt to suppress the material actually makes more people interested in seeing it.

Another consequence is that the center of media production can shift from institutions to individual journalists. We have already seen significant evidence of this in India. Journalists have left established organizations and essentially become small media houses themselves, sometimes one journalist working with a few staff members and operating through their own channels. That gives them considerable independence. But it also means that many of the institutional checks that once existed around an individual journalist may no longer exist. Take someone like Ravish Kumar. A viewer may trust his reporting and his voice. But when the journalist becomes the institution, that individual also becomes the final authority on what gets published.

There is a tradeoff. A journalist may now be able to say things independently that an established organization would not have permitted. At the same time, there are fewer internal checks and balances. And then the question becomes: What happens after repeated takedowns? When does a platform ban someone entirely? Even then, people frequently move to another platform or channel. It can become something of a Hydra situation: you shut down one outlet, and another appears.

Can factual, independent journalism still build a significant audience without government or large corporate backing?

There is already evidence that it can. One important development in India—and something American audiences should watch—is the growth of citizen-funded journalism. Print journalism became increasingly difficult economically because printing is expensive, and newspapers historically depended heavily on advertising. Online publishing changes those economics. If your production costs are relatively low, you do not necessarily need the same advertising structure. There are individual journalists doing this, but there are also independent news organizations that are entirely subscriber-supported. Some do not accept advertising at all. Others use hybrid models. For example, publishing on YouTube and allowing the platform to insert advertising without the news organization having a direct relationship with the advertiser.

So, there is good evidence that these models can work, certainly on a smaller scale. The concern is what happens to the political and information system around it.

What is the longer-term risk when political leaders become increasingly insulated from unscripted questions and traditional media institutions lose their central role?

One risk is that practices introduced by one government become attractive to its political opponents as well. Opposition politicians may criticize a communication system while they are out of power, but they may eventually decide that the same system works quite well for them once they are in power.

There is another problem for politicians themselves. Once everything is tightly choreographed, anything unscripted becomes risky. Politicians become better at politics by answering questions, dealing with disagreement and facing antagonism. If you protect political leaders from that process, they lose that ability. You can see this in political systems where leaders become accustomed to scripted questions and controlled environments. When they suddenly face genuine antagonism, they may have much more difficulty responding.

In the United States, politicians traditionally expect some of that confrontation. If you go into a town hall, you know that people may be angry and that many of the questions may be hostile. Dealing with that is part of the political process. If leaders stop doing it, they are no longer sharpening those skills.

The broader end point is that journalism itself can begin to feel downstream of government communication. Once that happens, people start looking elsewhere for what they regard as news. You then get increasing fragmentation. If you ask 20 people for their three main sources of news and everyone gives you completely different answers, you no longer have even one or two widely shared institutions serving as common reference points.

Eventually, delegitimizing media can produce something close to epistemic nihilism: Nobody entirely knows what to believe, so people increasingly accept as true whatever information most closely matches what they already believe.

Fact-Checked by Irfan Ahmad.

Read next:

• Which AI Apps Collect the Most Data? Meta AI Collects 33 of 35 Data Types, Followed by Muse, Gemini and ChatGPT

• Google Expands AI Mode Info Monitoring Globally, Letting All Users Track Web Info and Get Updates

Which AI Apps Collect the Most Data? Meta AI Collects 33 of 35 Data Types, Followed by Muse, Gemini and ChatGPT

2026-09-29T18:41:17.835+05:00

By Surfshark

Meta recently introduced its newest AI tool, Meta Muse, which overtook ChatGPT to claim the top spot among free iOS apps in the US during its first week. However, the extensive data collection practices of these AI applications remain a major concern for online privacy. To evaluate how Meta's new AI agent compares with other widely used AI chatbots in terms of data collection, we examined their data practices of 12 top AI chatbots alongside Meta Muse.

Image: Surfshark. CC BY

Key insights

AI chatbot apps collect an average of 13 out of 35 data types listed in the Apple App Store, but collection practices vary widely. Meta AI collects the most user data among the analyzed apps, gathering 33 data types — nearly 95% of the total. By contrast, Pi collects the fewest at three data types (user ID, product interaction, and crash data), representing an 11-fold difference.

However, regarding data collected without being linked to a user’s identity (such as via account or device details), Duck.ai stands out. It is the only chatbot that gathers six data types — primarily from diagnostics and usage categories — none of which are tied to users. Meanwhile, Microsoft Copilot is the only app that discloses collecting data for user tracking. As defined by Apple, “tracking” refers to linking data collected by the app about a particular user or device with third-party data, typically used for targeted advertising or data broker sharing.¹

Beyond AI chatbots, AI agents are surging in popularity — Muse from Meta, for example, recently topped the download charts.² These agents are potentially more invasive by design: to execute actions on a user’s behalf, they require access to other apps and services, such as calendars, email accounts, or financial tools. This pattern is also reflected in Muse’s privacy practices, which place it second overall among AI chatbots, collecting 31 out of 35 listed data types. Additionally, Meta uses Muse interactions for AI training by default, though users can manually opt out in the app’s settings.³

Google Gemini now collects 24 data types — an increase of one data type, audio, compared to the previous review. Gemini’s data collection spans across multiple categories, including contact details, user content, search and browsing histories, etc. This also includes precise location data, which only Gemini, Meta AI, Muse, and Perplexity collect (whereas Microsoft Copilot switched to coarse location data). Alongside Meta AI and Muse, it also collects sensitive information, which includes racial or ethnic data, sexual orientation, pregnancy or childbirth information, disability, religious or philosophical beliefs, trade union membership, political opinion, genetic information, or biometric data.¹

ChatGPT may collect 17 out of 35 data types, as reported by its developers. First disclosed in a March 2026 analysis⁴, this represents a 70% increase from the 10 data types recorded in last year's AI chatbots review, indicating a notable broadening in the extent of user data collection. The additional data types collected include coarse location, health and fitness, search history, audio data, advertising data, and customer support.

DeepSeek currently collects 14 types of data — also an increase of one data type, audio, compared to the previous review. However, DeepSeek, like ChatGPT, states that audio data is collected exclusively for functionality purposes. This supports activities such as authenticating users, enabling features, preventing fraud, implementing security measures, maintaining server uptime, reducing app crashes, improving scalability and performance, and delivering customer support.¹ In contrast, Google Gemini, Meta AI, and Muse reserve the right to use audio data for other purposes, such as product personalization or analytics.

Methodology and sources

This study compares the 12 best AI chatbots identified by Zapier⁵, alongside Muse from Meta, which was added following its recent surge to the top of app store charts ahead of several major AI chatbots.² Privacy practice details for each app were sourced from the Apple App Store on September 22, 2026, which outlines 35 unique data types categorized into 16 categories. The analysis focused on the total volume of data types collected, the stated purposes for processing, and the presence of user tracking practices.

For the complete research material behind this study, click here.

Data was collected from: Apple (2025). App Store.

References:

¹ Apple (2026). App privacy details on the App Store;

² CNBC (2026). Meta’s Muse AI agent downloads are surging. Here’s how it compares to ChatGPT, Grok and Claude;

³ Meta (2026). Muse Privacy Policy;

⁴ Surfshark (2026). Meta AI, Google Gemini, and ChatGPT are top data collectors;

⁵ Zapier (2026). The best AI chatbots in 2026.

Fact-Checked by Irfan Ahmad.

Read next:

• Google Expands AI Mode Info Monitoring Globally, Letting All Users Track Web Info and Get Updates

• Study Finds AI Chatbots Provide Less Diverse Information Than Traditional Web Searches

Google Expands AI Mode Info Monitoring Globally, Letting All Users Track Web Info and Get Updates

2026-09-29T17:16:55.194+05:00

Google is rolling out information monitoring capabilities in AI Mode in Search to all users globally, Robby Stein, Google’s vice president of product for Search, said Sept. 29.

Stein said Ultra and Pro subscribers had been using the capability before the wider rollout. Users can tell AI Mode what to look for, while Search continuously checks changing information/topics across websites, forums and social media posts, as well as Google’s real-time data sources and Shopping Graph of more than 60 billion products.

Stein gave examples including new restaurants or pop-ups opening nearby, local holiday activities for children, and back-in-stock or price-drop updates. He also said Search will suggest tasks while users are searching so they can get the latest information when it happens.

Google announced Search agents at I/O in May 2026, saying they would initially launch for Google AI Pro and Ultra subscribers during summer 2026. The company said the agents would monitor information across the web and fresh data for changes related to a user’s specific question.

Image: Google

Fact-Checked by Irfan Ahmad.

Read next:

• Study Finds AI Chatbots Provide Less Diverse Information Than Traditional Web Searches

• ‘Lately, my algorithm is trash’: Are we shaping our feeds or are they shaping us?

Study Finds AI Chatbots Provide Less Diverse Information Than Traditional Web Searches

2026-09-29T16:14:26.160+05:00

By University of Copenhagen

We are exposed to a much narrower range of knowledge through AI chatbots than through a basic web search. This is documented by a new study led by the University of Copenhagen. As more and more of us turn to AI chatbots for information, researchers warn that the risk of ‘knowledge collapse’ increases.

Image: Shubham Dhage - Unsplash

“Let me just ask the chatbot.” For many of us, this has become an everyday phrase. Where we used to turn to Google to find information, AI chatbots have become a common way of getting answers to everything from what to make for dinner and how to word that difficult email to the boss, to what it actually means when interest rates rise.

But the large language models underpinning AI chatbots give us a significantly narrower range of information than a conventional web search. This is the finding of a new study led by researchers at the Department of Computer Science (DIKU) at the University of Copenhagen.

“Every language model we tested provides users with more uniform information than a simple Google search across all the topics we looked at. In other words, people are to a large extent exposed to the same information over and over again. So AI chatbots are not just changing how we find knowledge, but also which knowledge we have access to,” says first author Dustin Wright, a former postdoctoral researcher at DIKU who is now an assistant professor at Aalborg University.

At least 18% less diverse than Google

The researchers tested 27 different large language models on 155 topics. For each topic, they used 200 different prompt formulations based on questions from real users. This generated a dataset containing around 70 million individual claims produced by the models.

The results show that even the language model producing the most diverse answers – OpenAI’s GPT-5 – provides at least 18.7 per cent less varied information than Google. The topics tested by the researchers ranged from nuclear weapons, marriage, pornography, racism and genocide to more country-specific topics such as Marine Le Pen, the Falklands War and K-pop.

According to the researchers, the fact that people are increasingly using AI models as their primary gateway to information could have significant consequences:

“We risk exposing people to fewer perspectives and a narrower range of knowledge. This could create a vicious cycle in which the most popular content becomes even more dominant, while other content is increasingly overlooked,” says Professor at DIKU and senior author Isabelle Augenstein, adding:

“It’s similar to globalisation. Today, you can buy the same products and find the same coffee chains almost everywhere in the world. That has many advantages, but it has also reduced diversity.”

Why is diversity so low?

According to the researchers, the low level of diversity in language models is partly a result of how the models basically work. They compress the vast amount of text they are trained on and learn the patterns that occur most frequently. In the process, information that deviates from the most common patterns is filtered out.

The effect could be amplified if language models are increasingly trained on text produced by other AI models – something the researchers expect to happen. In that case, models would learn from their own outputs, which are already less diverse than the human-written texts on which they were originally trained.

If this process is repeated over several generations of models, the range of information could gradually become narrower. This is what the researchers refer to as ‘knowledge collapse’.

“It’s a worrying thought. However, we can see that the more recent models produce slightly more diverse answers than older models, so knowledge collapse is not happening yet. But the mechanism that could trigger it in the longer term is already there. So it is something we should be aware of,” says Isabelle Augenstein.

Seek out different sources

The researchers therefore also hope that people will use AI thoughtfully:

“AI chatbot summaries can of course be useful – that is why so many people use them. But it is still important to seek out different sources in order to understand the nuances and get a broader picture – especially for the younger generation growing up with AI. We must not become so dependent on the technology that we stop understanding and thinking for ourselves,” says Isabelle Augenstein.

The AI industry should also pay attention to the issue, the researchers argue:

“We hope AI developers will build language models in a way that preserves the breadth of knowledge available to us. We have developed a method that they can use to measure diversity in models, which could help ensure that future language models do not become less diverse,” says Dustin Wright.

The study also notes in its limitations, "while epistemic diversity is important, it should also be contextualized with other important aspects of knowledge such as factuality and relevance."

About the study

The researchers analysed 27 large language models from OpenAI (GPT), Meta (Llama), Google (Gemma) and Alibaba (Qwen).
The models were tested on 155 topics relating to 12 different countries.
The analysis covered around 1.7 million AI-generated answers and approximately 70 million individual claims.
The results show, among other things, that smaller AI models generate more diverse content than larger models, and that newer models are more diverse than older models. Overall, however, all the models had significantly lower diversity than traditional web search engines.
The research was conducted by researchers from the University of Copenhagen, Aalborg University, Stanford University, the University of Colorado Boulder and the University of Texas at Austin.

Fact-Checked by Irfan Ahmad.

Read next:

• ‘Lately, my algorithm is trash’: Are we shaping our feeds or are they shaping us?

• Study Finds People May Embrace AI Advice That Confirms Their Views and Dismiss It When Challenged

‘Lately, my algorithm is trash’: Are we shaping our feeds or are they shaping us?

2026-09-28T20:39:35.020+05:00

By Liz Mineo, Harvard Gazette

Image: Raychan - Unsplash

Recommendation algorithms promise social media users a personalized experience, but some experts worry they promote doomscrolling, brain rot, and myriad mental health problems. Australia recently proposed legislation to allow users to opt out of algorithms with the goal of protecting children online and giving people more control over what they see on their feeds.

We asked faculty members from different fields of study to share their views and concerns about recommendation algorithms. Their responses have been edited for clarity and length.

‘Dark nudging’ from the Garden of Eden to AI age

— Cass Sunstein, Robert Walmsley University Professor, Harvard Law School, and co-author of the book “Algorithmic Harm: Protecting People in the Age of Artificial Intelligence”

It is standard to define manipulation as a form of trickery or covert influence. Some definitions suggest that manipulation treats people as “tools or fools.” My preferred definition sees manipulation as an influence that fails to respect people’s capacity for reflective and deliberative choice. The basic idea is that manipulators exploit a lack of information on the part of their victims or take advantage of our behavioral biases (such as a focus on the short term, unrealistic optimism, and limited attention). You can think of manipulation as a form of “dark nudging.” The problem is the same, from the Garden of Eden to the era of algorithms and AI.

I like recommendation algorithms! They can help us put our time and our money in the right places. If an algorithm tells me about a new book on behavioral economics or constitutional law, I will be grateful. Personalization is often highly beneficial. I don’t want to see anything about hockey, advanced physics, Jell-O, Beethoven, or Bach, even though a lot of people do. The concern here points to manipulation, as when recommendation algorithms exploit people’s lack of information or their behavioral biases. That can be a real problem.

We need laws and regulations to protect against the worst forms of manipulation, just as we need them to protect against deception. We need a new right — the right not to be manipulated. As individuals, we can and should pause a few beats and ask: “On reflection, do I really want to buy what they’re selling? With my time or my money? They might be trying to trick me!”

Feeds like ‘fun house mirrors’

— Emily Weinstein, Lecturer on Education and co-director of the Center for Digital Thriving, Harvard Graduate School of Education

I get to do a lot of deep listening to young people, and usually our aim is not just to understand what they see when they’re behind their screens, but how they see it. I’ve been doing this work for more than 15 years. Recently, the “algorithm” seems to have shifted from a supporting player to a leading role in teens’ telling of their experiences with tech. They say things like: “My algorithm is just not great on that app” or “Lately, my algorithm is trash.”

Young people shape their feeds in ways they’re well aware of, as well as in ways that they say they don’t understand. Their feeds shape them, too. One 16-year-old recently told me that her feed is filled with posts where, “It almost feels like everyone’s pushing the rush of your life — like get married at 18, have kids by 20. That’s just not everybody’s reality, but it makes it seem like it’s supposed to be.” Other teens have feeds that look completely different.

When you sit alongside teens as they scroll, it’s abundantly clear how they are being profiled and pulled. Their interests (and their identities) are mirrored back with distortions like fun house mirrors. Another high schooler who is an avid gamer showed me his feed was full of “edits” (rapid-cut montages of movie scenes and TV shows), games, and ads for sports betting and online gambling.

Algorithms can contribute to teens (and all of us) feeling inundated by narrowly themed content. For one 17-year-old I spoke with in the spring, the algorithm was saturating her feed in prom content: “I get it, I get it, it’s prom season, but it’s like every single time I refresh, it’s prom. But I don’t think it really changes how I feel about prom itself. It’s just like when I’m looking at it on my phone, I’m like, bro, can I see something else?”

Our personalized algorithms transport us into very different worlds, and even when we understand how they work, it doesn’t inoculate us from their impacts. For researchers, algorithms can also make it hard to capture evidence of the effects of these technologies, because everyone’s experiences are so varied and different by design.

Responsibility for digital agency can’t rest just on the shoulders of our young people, who are being asked to withstand designs that prey on psychological vulnerabilities and developmental sensitivities. We need more control over algorithms, and features that allow us to make changes to them. Young people also deserve to understand how and why agency can be hard to exercise — in particular the incentive structures that are behind design choices that impact them.

‘Guard your attention very carefully’

— Rebecca Lemov, Professor of the History of Science, and author of the book “The Instability of Truth: Brainwashing, Mind Control and Hyper-Persuasion”

Hyper-persuasion is a word I use to describe the compelling and even coercive parts of persuasion seen in the operations of modern digital media. An early example that serves as a harbinger is the 2014 Facebook Experiment by the Proceedings of the National Academy of Sciences journal, in which 689,003 users of the site found themselves unwittingly “nudged” into slightly more depressive (or slightly less despondent) states, as measured by their behavior after having their personal feeds manipulated. This was the power of social media — to induce “massive-scale emotional contagion,” almost like turning a dial on the user’s psyche. Little by little, as we engage with these technologies, we are ungrounded. The algorithmically-driven version of this is seen in how almost all of us — be honest! — can be prone to getting caught in a loop, a flow state, a doomscroll, or another daily type of dissociative mind state.

Social media interactions and AI chatbots, with their targeted algorithms, can become microenvironments. One of the key principles, really the foundation, of classical brainwashing is what Robert Jay Lifton called milieu control. Everything else — the emotional engineering of a captive, their ideological realignment (mostly temporary), or a radical personality change — follows from this control of the local environment, especially via control of what comes in and what goes out. Milieu, after all, means place or surroundings. You can imagine a highly controlling abusive cult or a prisoner of war camp that has no fences or barbed wire because it is so remote (this was the milieu in which U.S. GIs captured in the Korean War found themselves, where they underwent communist re-education). Although these circumstances seem extreme compared with the microenvironments created by your seemingly tiny interactions with your phone, for example, the same dynamics are at play. The device, with its algorithmic targeting, selects or controls what comes in and out of view for you, to which you react, which then continues to feed the selection process. This is why the word siloing is often used. It amounts to a type of invisible control.

My recommendation is to guard your attention very carefully. If it goes astray, or you find yourself in a mindless attention-sapping loop, bring your attention back to awareness. Strive to check in with your own feedback, as experienced in the body. There’s a subtle feeling that can arise when any of us encounters draining or (even) undermining messages. You can ask intermittently, how does this make me feel? It’s not a matter of avoiding what makes you uncomfortable but rather of making a deliberate choice that giving my attention to this is meaningful.

‘Is my phone listening in?’

— Sitan Chen, Assistant Professor of Computer Science

An algorithm is any procedure that takes an input, performs some clearly specified computation, and produces an output. The first algorithms that we learn, in grade school, are how to add, multiply, and divide numbers. In machine learning, the process of training a model on data, either to predict (e.g., “Is this individual eligible for a loan?” or “Is this chest X-ray that of a healthy patient?”) or to generate (e.g., images, videos, text), is an algorithm called “gradient descent,” whose roots go back to 19th-century mathematics. Its input is data, and its output is a bunch of numbers, in some cases trillions of numbers, specifying a model. Somewhat confusingly, these models themselves are also algorithms. They are designed to take in some input, for instance a picture of someone’s face, or some prompt in plain English, and output a response, for instance whether the person has checked into their flight, or an answer to a math question.

Algorithms and scientific progress have gone hand in hand since antiquity. More recently, algorithms for numerical simulation have been instrumental in the study of complex physical and biological systems. Algorithms for search and sequence alignment enabled the sequencing of the human genome. Algorithms for reconstructing an image from noisy measurements are run every time a doctor performs an MRI or CT scan.

Recommendation algorithms are pervasive in social media. Their basic functionality is to take as input a user’s browsing patterns and output suggestions for content meant to maximize user engagement and drive sales. The implications for addiction, among other mental health concerns, are well documented. In terms of privacy, not only is it unclear to the user which aspects of their interactions with social media are being used to drive recommendations (“Is my phone listening in on my conversations?”), but these interactions ultimately also go into training the machine learning model that is making recommendations to other users, possibly revealing to them traces of a user’s activity they never intended to share. There is now over two decades of scholarship on this latter kind of vulnerability, and a massive body of work, e.g., on differential privacy, proposing interventions that algorithm designers could employ to mitigate these risks.

Legally, ‘We could, and should, do much better’

— Leah Plunkett, Meyer Research Lecturer on Law and faculty associate with the Berkman Klein Center for Internet & Society

My overarching concerns about the impact of recommendation algorithms on young people’s lives are about two types of decisions: those that are made about young people without their knowledge or full consent, and those that are made by young people without their full awareness or meaningful consent, about the role of a recommendation algorithm.

Recommendation algorithms appear in tools, services, and platforms that might inform access to major life opportunities such as employment, insurance, etc., but also in social and emotional learning programs used by schools. And in social media, the feed being pushed is usually through the secret sauce of a recommendation algorithm, and for kids and teens, they might feel that a lot of the suggested content is organic and natural. The recommendation algorithm really can lure you into a false sense that this content is really meant for you. It can also get very insidious, even risky and dangerous, if the content that is being fed to you to keep your attention is depicting activities, emotions, experiences, relationship dynamics, and worldviews that are harmful or risky to minors.

I think the law in the United States has largely failed to provide ethical, practical protections for kids on social media and many other tech platforms. Ideally, the federal legislature would have done comprehensive individual consumer privacy for the digital age a couple of decades ago, and at the very least should have done it for minors, and they haven’t. We have seen many leading tech companies move fast and break things, fail early and often, act all the way up against, sometimes even past, the boundaries of what the law permits. So, to the extent that we have seen tech companies avail themselves of the freedoms entitled to them in this country to build lawful businesses, I would say shame on companies that have chosen to conduct their affairs past what the law permits, or even close to the boundary, and it is long past time for the places where that has happened to change.

I am not anti-tech. I am not anti-social media. I am not anti-innovation. I think the formal and informal governance mechanisms here are complicated, and I think that most people, whether they are federal lawmakers, state regulators, tech executives, or vice principals, are trying their best to regulate the way these companies operate and protect our kids and teens. I also think we could, and should, do much better.

Fact-Checked by Irfan Ahmad.

Read next:

• Study Finds People May Embrace AI Advice That Confirms Their Views and Dismiss It When Challenged

• Google’s content recommendation algorithm is a black box. Here’s how it decides what to show you

Study Finds People May Embrace AI Advice That Confirms Their Views and Dismiss It When Challenged

2026-09-28T16:48:37.842+05:00

By University of Michigan

Image: Swello - Unsplash

When an AI chatbot agrees with our reasoning in resolving a social dilemma, we may become more confident in our opinions. But when the chatbot disagrees, we may simply dismiss the challenge rather than reconsider our views.

That’s according to new research from the University of Michigan examining how people respond when artificial intelligence weighs in on personal and interpersonal dilemmas.

The study, appearing in Computers in Human Behavior Reports, highlights a challenge for developers trying to reduce the tendency of chatbots to agree too readily with users.

The authors note that AI models are often designed to be agreeable, which can lead them to validate users’ reasoning even when that reasoning may be flawed. They describe how this kind of confirmation can reinforce harmful ideas and distorted perceptions of reality.

“Developers have been working to make AI models more willing to disagree when necessary, but having an AI disagree with the user does not necessarily mean that people will reconsider their own views,” said Atakan Atamer, the study’s co-lead author and U-M psychology doctoral student. “Instead, they may simply discount the AI’s response.”

Putting AI advice to the test

Atamer and colleagues recruited 482 U.S. adults who considered one of three hypothetical interpersonal dilemmas with no objectively right or wrong answer. The scenarios involved conflicts between a partner and family, whether to forgive a cheating partner, and whether to express a preference that differed from that of a friend group.

After making their choice and rating how strongly they believed in it, participants received an AI response that either agreed with or challenged their reasoning. The researchers then measured participants’ confidence in their choice and their perceptions of the AI, including its ability to understand emotions, human likeness, and whether they would use AI again for similar conversations.

Agreement reinforces confidence

When the chatbot agreed with their reasoning, participants became more confident in the choices they had already made. But when the chatbot challenged their reasoning, their confidence in their original views did not significantly decrease.

Participants who encountered disagreement were more likely to view AI as “machine-like” and less capable of understanding human emotions. They also expressed less interest in using AI for similar interpersonal conversations.

In other words, AI disagreement did not appear to prompt participants to substantially reconsider their original positions. Instead, it affected their perceptions of the technology.

A challenge for AI-human interaction

The authors argued that the findings have implications for the growing use of AI for personal advice. AI systems that consistently agree with users can reinforce existing beliefs and increase confidence without necessarily improving the quality of their reasoning. But simply making AI disagree more often may not solve the problem: Users may be less willing to engage with AI systems that challenge their views or may simply dismiss the counterarguments they provide.

“Our results suggest that people often prefer AI to reinforce their existing views and may discount its response when it disagrees, meaning that simply making AI more willing to challenge users may not be enough to prevent potentially harmful conversations with real-world consequences,” said Olivia Pinto, co-lead author, U-M alumna and UX researcher.

As AI becomes more common in personal decision-making and advice, the researchers say that reducing excessive agreement may require more than simply teaching AI systems to push back. If users tend to embrace AI advice when it confirms their views but dismiss it when it challenges them, reducing AI sycophancy alone may not address the larger risks of relying on AI for personal guidance.

Fact-Checked by Irfan Ahmad.

Read next: Higher Subscription Prices Can Reduce the Number of Customers While Encouraging Heavier Usage, Research Finds

Higher Subscription Prices Can Reduce the Number of Customers While Encouraging Heavier Usage, Research Finds

2026-09-28T15:28:58.781+05:00

By Caitlin Clark, Texas A&M University

Image: Jacob Padilla - Unsplash

Paying a steeper price for a subscription service can trigger a familiar impulse: You want to get your money’s worth.

For companies selling AI tools, cloud computing and other digital services, that can complicate a basic pricing strategy. Raising prices may keep some customers from signing up, but those who do pay may be motivated to use the service more.

New research from Texas A&M University examines what those competing effects mean for companies trying to manage demand through pricing, particularly for digital services where additional usage carries real costs.

“These high prices can actually have a double-edged sword effect, where people say, ‘We paid a lot, so let’s just get our money’s worth and use it more,’” said Dr. Rajiv Mukherjee, a professor at Texas A&M’s Mays Business School.

The study, published in Production and Operations Management, was conducted with Sreekumar Bhaskaran of Southern Methodist University and Sanjiv Erat of the University of California San Diego.

The other side of demand

Companies have traditionally used price as one way to manage congestion. When demand for a service strains capacity, raising the price can reduce the number of customers willing to buy access — and in turn, reduce the number entering the system.

But Mukherjee and his colleagues argue that this view can overlook what happens after someone pays. For prepaid and subscription services, customers don’t just decide whether to buy access. They also decide how much to use the service once they’re in.

The researchers developed an analytical model to examine both sides of demand: how many consumers purchase access to a service and how much each of those consumers subsequently uses it. A higher price can reduce the first while increasing the second.

The reason lies in a behavioral economics concept known as mental accounting. People tend to mentally track what they have spent against what they receive in return. Paying more upfront can create a greater incentive to consume enough of a service to feel the purchase was worthwhile.

“When people pay for something, they expect certain value out of it,” Mukherjee said. “As soon as you pay for a service, you create a mental account deficit.”

The researchers refer to the resulting tendency to consume more as “consumption bias.” When that bias is strong enough, heavier usage among paying customers can work against the reduction in the number of subscribers and potentially worsen congestion.

Why AI changes the equation

The issue is particularly relevant for AI and cloud computing services, where additional usage carries real costs for providers. AI queries require computing power, while cloud services carry infrastructure costs.

“Every query you make, that has a significant amount of cost that the firm has to bear,” Mukherjee said of AI services. “Once the customer subscribes, they don’t really care. They are just getting things done using the service and trying to get the money’s worth in the process.”

The paper cites reports involving ChatGPT and Amazon Web Services as real-world examples that helped motivate the research. After price increases, ChatGPT subscribers reportedly explored more features and increased their usage, while AWS customers used more of their precommitted cloud spending before the billing period ended.

“Unlike traditional digital goods where the marginal cost was negligible, modern firms in the post-AI and cloud-computing era have a high marginal cost of service, and they haven’t quite figured out how to incorporate that into a good pricing strategy,” Mukherjee said.

When charging less could make sense

The findings don’t mean every subscription service should lower its prices.

When consumption bias is low, raising prices to reduce the number of customers can still make sense. But when customers are strongly motivated to get their money’s worth, lowering the upfront price can sometimes reduce how much they use the service.

The researchers also found that when usage is costly to provide, a pricing model that charges customers partly based on how much they use the service can become more attractive than relying solely on a subscription fee.

The larger lesson is that subscription companies need to consider not only how price affects the number of customers, but what those customers do after they subscribe.

“The initial demand through the people who are coming into the system is not the end of the story,” he said. “That’s pretty much the beginning of the story when the marginal cost of service is high.”

Fact-Checked by Irfan Ahmad.

Read next:

• Human-like AI not always the answer in customer service

• In ‘NAZA,’ Israeli whistleblowers’ revelations about civilian deaths in Gaza reflect a longer history of Palestinian isolation

• Why Five Common Assumptions About AI Jobs, Productivity, Intelligence, Business Value, and Headcount Need Scrutiny

In ‘NAZA,’ Israeli whistleblowers’ revelations about civilian deaths in Gaza reflect a longer history of Palestinian isolation

2026-09-28T12:48:40.259+05:00

By Drew Paul, University of Tennessee

Image: Mohammed Ibrahim - Unsplash

There’s a powerful scene in Moroccan-French-Israeli filmmaker Simone Bitton’s 2004 documentary “Wall,” which tells the story of the construction of the West Bank barrier, a 448-mile – or 721-kilometer – network of walls and fences dividing large parts of the West Bank from Israel.

A concrete slab is slowly lowered into place in order to complete a section of the wall. Inch by inch, the vibrant, hilly Palestinian community on the other side gets smaller and smaller.

And then – just like that – a world that’s a stone’s throw away is gone.

As someone who studies the restrictions placed on Palestinians, I often think about this scene as a microcosm of Israel’s decades-long use of walls, checkpoints and surveillance technologies to seal off Palestinians from the rest of society.

Nowhere is this dynamic more evident than the Gaza Strip, which Israel fenced in during the 1990s and has blockaded since 2007, citing security concerns and the threat of militant attacks.

Thanks to these barriers, the immense suffering of Palestinians in Gaza – more than 70,000 of whom have been killed as a result of the Israel-Hamas war since 2023 – has been difficult to document. The Israeli government largely bars international journalists from entering the Gaza Strip independently, while Palestinians in Gaza are generally not allowed to leave the Gaza Strip.

But the Sept. 10, 2026, premiere of Israeli filmmakers Yuval Abraham and Rachel Szor’s film “NAZA” at the Venice International Film Festival has put the spotlight on the ongoing tragedy in Gaza without relying on footage of gruesome corpses or bombed-out neighborhoods.

The film consists largely of anonymous testimony from 24 members of Israel’s military and intelligence services. They discuss the ways in which Israel has used artificial intelligence to target Hamas officials, despite knowing the attacks will also generate large numbers of civilian casualties. The title, “NAZA,” comes from an acronym for the Hebrew phrase “nezek agavi,” or “collateral damage,” and it’s the Israeli military’s threshold system for how many civilian casualties are acceptable in its military strikes on Gaza since the Oct. 7, 2023, attacks. Hamas-led militants from Gaza launched those attacks in southern Israel, murdering 1,200 people – mostly civilians – and also taking about 250 hostages.

To me, part of the film’s power lies in its focus on the clinical way of carrying out strikes on Gaza from a distance. But the Israeli military’s approach could only work after systematically isolating the Palestinian people over the course of decades.

Unseen victims

In addition to constructing walls and blockading Gaza, Israel has enforced separation of Palestinians by restricting or denying permits to Palestinians wishing to enter or work in Israel through aerial surveillance and through the expansion of Israeli settlements in the West Bank.

Both the visual language and the content of “NAZA” reflect and reveal how knowledge of Palestinian suffering is often mitigated in Israeli culture, politics and daily life.

The film juxtaposes audio of Israeli soldiers’ testimony with shots of the Tel Aviv cityscape at night. In these interviews, the soldiers map out how targets are chosen, how attacks are planned and how much leeway is given for civilian casualties.

The victims themselves are largely unseen, despite their close proximity: Tel Aviv sits a mere 43 miles – or 70 kilometers – up the coast from the Gaza Strip.

The choice to film primarily at night, in darkness, using disembodied voices to tell a story not depicted on camera, further highlights the extent to which these events have been obscured in Israel and around the world. Perhaps that’s why the Israeli government has gone to such great lengths to discredit the film, saying it contains “deliberate distortions of reality.” Prime Minister Benjamin Netanyahu has threatened to revoke the citizenship of its directors.

In response to the interest and backlash surrounding the film, the filmmakers announced that “NAZA” will be available to stream for free, for anyone to see, beginning in November 2026.

Breaking through

“NAZA” highlights a psychological distance between attacker and attacked that was enabled by AI, computer screens and the physical isolation of the targets. Yet it is far from the first film to call attention to the invisibility of Palestinians.

Elia Suleiman’s 2002 dark comedy “Divine Intervention” repeatedly stages scenes in which Palestinian characters are stuck at or try to get around an Israeli military checkpoint. And Yoav Shamir’s 2003 documentary “Checkpoint” chronicles everyday interactions between Palestinians and Israeli soldiers at checkpoints in the West Bank and Gaza.

Other films center on acts of transgression or protest. Hany Abu-Assad’s 2013 drama “Omar” and Ameen Nayfeh’s 2020 drama “200 Meters” both feature Palestinian characters who break the law and surreptitiously cross the West Bank border into Israeli territory, exposing themselves to physical and legal danger.

“No Other Land” – the 2024 documentary that netted Abraham and Szor an Oscar, along with fellow directors Basel Adra and Hamdan Ballal – points the camera on Palestinian protest movements against land expropriations and home demolitions in the West Bank. In the 2011 film “5 Broken Cameras,” Palestinian farmer Emad Burnat captures his West Bank village’s struggle against the construction of a massive wall that chokes off farmland and access to vital resources. Even as his cameras are repeatedly damaged or destroyed, Burnat refuses to stop documenting the resistance.

‘NAZA’ will be available to stream for free starting in November 2026.

Crying while shooting

“NAZA” also points to another phenomenon: how isolation and separation encourage complicity. The physical distance between Israeli soldiers and the people they target may make it easier to go along with or ignore what is happening.

In this way, “NAZA” grapples with a trope in Israeli films and literature of exploring the guilt of Israeli soldiers. Sometimes derisively referred to by critics as “shooting and crying,” the thematic framework of these films has been criticized for focusing on Israeli moral anguish at the expense of Palestinian experiences and perspectives.

For example, S. Yizhar’s 1949 novel, “Khirbet Khizeh,” which was adapted into a film in 1978, tells the story from the perspective of Micha, an Israeli soldier fighting in the 1948 war that established the state of Israel. He’s part of a squad that expels Palestinians from their village, and his ensuing guilt provokes a crisis of conscience that haunts him for years to come.

Decades later, the 2008 animated documentary “Waltz With Bashir” follows the film’s director, Ari Folman, as he plumbs his repressed memories of his participation in the 1982 Israeli invasion of Beirut and how he inadvertently helped Lebanese Christian Phalangist militia fighters carry out massacres of Palestinians at the Sabra and Shatila refugee camps.

‘Waltz With Bashir’ centered on the fragmented memories, trauma and guilt of an Israeli soldier who had helped facilitate the Sabra and Shatila massacre in Beirut, Lebanon, in 1982.

The “shooting and crying” genre was largely accepted in mainstream Israeli culture, with “Khirbet Khizeh” taught in Israeli schools for decades. But the Israeli government’s hostile reaction to “NAZA” perhaps signals not only the growing intolerance of dissent in Israel, but also a fading cultural and political concern with the morality of Israel’s occupation of Palestinian territory.

Whose voices get heard

The centering of Israeli perspectives in these earlier works – as well as in “NAZA” – also shows how violence against Palestinians can be given greater credence when it’s described by Israelis themselves.

Recent Palestinian narratives about Gaza, such as “From Ground Zero” and “Thirteen in Gaza,” have garnered some attention and praise in the film world, but nowhere near the same amount as “NAZA” has in the weeks since its premiere.

Even “No Other Land,” which is centered on a Palestinian village, features an Israeli journalist and filmmaker, Abraham, documenting the protests and confronting Israeli soldiers, and may not have received the same acclaim if it had been made entirely by Palestinians.

Yet, despite these many barriers, Palestinians have never stopped finding ways to make themselves heard. In addition to films like “Omar” and “200 Meters,” videos and images shared directly to social media have allowed Palestinians to connect directly with audiences and shape global awareness of their plight, especially since the current war in Gaza began.

They have been able to bypass the physical barriers that have long restricted their movement – and the cultural barriers that have often kept their experiences out of sight.

NBC News covers ‘NAZA’ ahead of its premiere at the New York Film Festival on Sept. 26, 2026.

Drew Paul, Professor of Arabic, University of Tennessee

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Fact-Checked by Irfan Ahmad.

Read next:

• Why Five Common Assumptions About AI Jobs, Productivity, Intelligence, Business Value, and Headcount Need Scrutiny

• Study Finds Health Claims Can Mask Nutritional Trade-Offs in Food Products

Study Finds Health Claims Can Mask Nutritional Trade-Offs in Food Products

2026-09-26T15:06:11.387+05:00

By Laura Reiley, Cornell University

Image: Christian Bowen - Unsplash

Walk down any supermarket aisle and the packaging will reassure you: Low Sodium. No Artificial Flavors. Organic. High Fiber. Immunity Support.

But a new analysis of 145,609 food and beverage products launched in the United States over the past decade, led by Miguel Gómez, the Robert G. Tobin professor of food marketing in the Cornell SC Johnson College of Business, and Pei Zhou, a postdoctoral researcher in his lab, suggests that reassurance is often misplaced. The study, which drew on a decade of new-product data from the market research firm Mintel (Global New Products Database), found that products bearing health-related claims are not only not consistently healthier, they are sometimes less nutritious. Reducing one unhealthy ingredient sometimes means they added more of a different one.

Zhou’s original interest in the topic came from her own shopping experience.

“One day I went to a supermarket, and I noticed there was original flavor and lower-fat cream cheese,” she said. “They had reduced the fat, but they had more total sugars and more sodium. There are so many label claims on packages. Our main question was whether products with claims are healthier than those without health claims.”

The Cornell team applied the Nutrient Profiling Model, a scoring system developed by the U.K. Food Standards Agency, which – unlike a single-nutrient claim – weighs a product's full nutritional makeup at once, penalizing excess sugar, sodium, saturated fat and calories while rewarding fiber, protein and fruit and vegetable content. It’s the same tool British regulators use to decide which products can be advertised to children.

Applying that lens to the American marketplace, the researchers found that, of the newly launched products they analyzed, two-thirds qualified as less healthy overall even though half of the products had on-package health claims.

The trouble, the researchers found, often comes down to trade-offs hidden in the fine print of formulation. A “reduced sodium” soda, for instance, is disproportionately likely to make up for it with extra sugar. A “no trans-fat” packaged meal often leans harder on saturated fat and sodium instead. The claim is technically true. The overall nutritional picture, the study suggests, doesn’t necessarily improve, and can sometimes get worse in another direction.

Beverages showed the starkest gaps, with claims like “low sodium,” “sugar free” and "low saturated fat" the most likely to diverge from actual nutritional quality. Food products fared somewhat better, with most claims aligning with genuinely improved nutrition – the notable exception being reduced trans-fat claims, which tracked with higher saturated fat.

Perhaps the most striking finding involved a category of labeling that isn't regulated as a nutrition claim at all: “natural.” Terms like organic, non-GMO and whole grain say nothing formally about a product’s sugar, sodium or fat content. Yet prior consumer research cited by the authors, which included Harry Kaiser, the Gellert Family Professor Emeritus in the Charles H. Dyson School of Applied Economics and Management, and Yuqing Zheng at the University of Kentucky, has repeatedly shown that shoppers read “natural” as shorthand for “healthy” anyway, a psychological shortcut researchers call the “halo effect.”

These companies are not breaking the law, Gómez and Zhou said. On-package claims are typically regulated for accuracy about the specific attribute they name – sodium content, say, or the presence of added vitamins. They’re not required to represent the product’s nutritional profile as a whole. But that gap, the study argues, loses ground for public health.

The findings arrive as federal regulators are already rethinking how nutrition information reaches consumers. The Food and Drug Administration finalized an updated definition of the “healthy” claim in recent years and has proposed a new front-of-package labeling scheme intended to summarize a product’s nutrition profile at a glance – an approach closer to the systems already used in the U.K., and to mandatory warning labels adopted in Chile and Mexico.

“We provide some direction for policymakers in our paper,” Gómez said. “The general idea is to transition from single claims to a broader nutrient profile style. In countries like Chile or Mexico, they use traffic stop signs to warn consumers about high salt, sugar and fat, and this limits the positioning of products in the supermarket.”

For now, Gómez and Zhou say, the practical advice for shoppers is unglamorous but familiar: A claim on the front of a box is a hint, not a verdict. The nutrition label on the back – added sugars, sodium, saturated fat, all of it together – still tells the fuller story.

It will be an uphill battle, Gómez fears.

“Moving toward a more comprehensive system in which you can evaluate the healthiness profile of products is better for consumers,” he said. “But it’s hard to implement politically.”

News courtesy of the Cornell Chronicle.

Fact-Checked by Irfan Ahmad.

Editor's note: Updated the study details and added the dataset name.

Read next:

• Why Five Common Assumptions About AI Jobs, Productivity, Intelligence, Business Value, and Headcount Need Scrutiny

• FIU Study Finds Smartwatches Overestimate Calories Burned During Exercise