Digital Information World

Researchers Explore What It Means To Say AI 'Thinks'

2026-08-22T17:01:32.334+05:00

By Stefanie Johndrow, Carnegie Mellon University

Carnegie Mellon research traces today's AI rhetoric to decades-old debates about language, technology and human intelligence.

Image: Rob Coates - Unsplash

Organizations can describe artificial intelligence using familiar words like "thinking," "writing" and "learning." But according to Carnegie Mellon University historian Christopher Phillips, those terms may tell us as much about how humans talk about technology as they do about the technology itself.

In a new paper published in IEEE Annals of the History of Computing, Phillips and co-author Alison Langmead of the University of Pittsburgh examine how language surrounding artificial intelligence has evolved since the 1950s. Their research argues that descriptions of AI have long relied on what they call “strategic ambiguity” — using words that carry precise technical meanings for computer scientists while suggesting something much broader to the public.

According to Phillips, the result is a growing tendency to compare computers with people instead of describing what each does best.

“We use words like ‘smart,’ ‘read,’ ‘write’ and ‘think’ very differently when we're talking about humans than when we're talking about computers,” said Phillips, professor and head of the Department of History in CMU’s Dietrich College of Humanities and Social Sciences. “The question we wanted to ask was: Why are we using the same words at all?”

"The work Chris and I are doing focuses on how human beings have allowed computers — tools of our own invention — to become an integral part of our daily life,” Langmead said. “Because they are now enmeshed in our social world, how we talk about them and imagine them matters a great deal. We would like for there to be a larger, ongoing conversation that transcends the hype cycle about precisely what computers can and cannot do."

How language shapes the AI conversation

Rather than debating whether machines can think like humans, Phillips and Langmead looked backward, examining conversations about computing during the 1950s and 1960s — before artificial intelligence became a household term and computers became widespread.

They found that today's debates are far from new.

Some early computing pioneers embraced human-centered language to describe computers, while others deliberately chose more precise, if less elegant, descriptions of what machines actually accomplished. Computer scientist Norbert Wiener, for example, described computers as “learning” when they implemented rules that resulted in more successful outcomes. To fellow researchers, the way he used the term had a well-defined technical meaning. To broader audiences, however, it could evoke the much richer human experience of learning.

For Phillips and Langmead, that distinction matters.

Strategic ambiguity allows technical language to be easily understood across audiences while sometimes making technologies appear more humanlike than they are. The practice isn't necessarily intentional, Phillips said, but it can shape how society understands AI.

“Why can't we say the computer is executing a particular set of instructions?” Phillips said. “Why do we have to call it thinking?”

The researchers stress that this does not diminish the accomplishments of modern AI. Phillips described today's large language models as amazing technological achievements capable of producing outputs that humans immediately recognize as meaningful.

Instead, Phillips argues that their accomplishments are remarkable precisely because they differ from human cognition.

“When you ask an image generator for a dog riding a pony in a New York Mets parade, and it produces exactly what you imagined, that's amazing,” Phillips said. “But let's not call that creativity. Let's call it the technical achievement that it is.”

Lessons from computing's early history

The paper also revisits modern benchmarks used to evaluate AI systems. Tests such as Massive Multitask Language Understanding, or MMLU, and the recently introduced “Humanity's Last Exam” are frequently described as measuring machine knowledge or reasoning. Phillips and Langmead argue that those benchmarks more accurately measure classification accuracy — how well systems identify correct answers on standardized evaluations — rather than demonstrating human knowledge or understanding.

When AI is described as thinking or writing, people can begin viewing machines as competitors rather than tools. That framing, Phillips argues, risks reducing complex human activities such as creativity, learning and reading to computational outputs while overlooking the relationships, lived experiences and judgment that shape those processes.

“Most of us read poetry because we're interested in the person who wrote it,” Phillips said. “We're interested in the emotion, the lived experience and the beauty that comes from having an actual human being produce something.”

One of the paper's more surprising discoveries was how interdisciplinary conversations about computing once were.

Phillips and Langmead found that during the mid-20th century, historians, literary scholars, psychologists, engineers and computer scientists all participated in discussions about what computers should do and where they belonged in society.

“It wasn't obvious who should control computing or what role computers should play,” Phillips said. “People across disciplines were asking those questions together.”

Why precision matters today

Today, Phillips believes those broader conversations are just as important. Across Dietrich College, colleagues in the humanities and social sciences bring different disciplinary perspectives to questions about AI, language and human knowledge

“This timely study reminds us that language does not simply deliver scientific or technological ideas: It changes these ideas and it transforms our understanding of them,” said Andreea Ritivoi, William S. Dietrich Professor of English and associate dean of research in Dietrich College. “The authors' compelling plea for precision is a great opportunity to underscore the need for productive collaborations across the ‘two cultures’ of science and humanities, as C.P. Snow, one of the heroes in this article, insisted decades ago. Too often the humanities and STEM are seen as islands of sorts, but the debates around AI happen in the troubled waters between them. We remain on one island at our own peril.”

Ritivoi’s perspective on language and interdisciplinary collaboration is complemented by Rob Kass’ focus on the importance of precision in the technical and statistical dimensions of AI.

“Computational algorithms, and mathematical derivations, rely on precision. Everyday language, however, makes heavy use of individual words that have multiple context-dependent meanings: they are often understood in a particular way by small groups of people within a limited setting, but others may easily attribute to those words very different meanings,” said Kass, Maurice Falk University Professor of Statistics & Computational Neuroscience in the Department of Statistics & Data Science and the School of Computer Science Machine Learning Department. “This is a big issue in statistical reasoning from data. Phillips and Langmead convincingly document both the early appearance of ambiguous terminology in AI research and the ways it continues to be used for strategic advantage, much to the detriment of society as a whole.”

Ultimately, Phillips hopes the paper encourages a more thoughtful conversation about AI — one grounded less in sweeping claims about machine intelligence and more in an honest assessment of what these systems actually do.

“We're not arguing that these technologies aren't impressive,” Phillips said. “We're asking people to be very clear about what the machines do and don't do.”

Edited by Irfan Ahmad.

Read next: What Do People Really Think About Generative AI?

What Do People Really Think About Generative AI?

2026-08-22T09:59:54.447+05:00

By Drexel University

A longitudinal study of Reddit posts since 2022 shows persistent tension between trust and distrust in generative AI.

Image: Solen Feyissa - Unsplash

Is generative artificial intelligence technology a helpful tool or a big problem? Nearly four years after ChatGPT became a household name, most people are either using AI applications or generally aware of what the technology can do. But understanding how much they actually trust the work it produces, and the answers it provides, remains an important question. New research from Drexel University, based on a longitudinal analysis of hundreds of thousands of Reddit posts since 2022, sheds light on the general perception of AI and suggests that the public remains largely divided on how much it can be trusted.

Recently published in the journal Transactions of the Association for Computational Linguistics, the study examined how trust and distrust toward generative AI were expressed over time, the dimensions and reasons underlying these attitudes and how these patterns differed across groups represented in the posts.

Its primary finding is that trust, which was expressed in about 31% of posts, modestly outpaced distrust (26%) in the technology, while 41% of posts expressed neither and 1% expressed both.

“These findings give us an important starting point and help to establish a baseline understanding which can help inform responsible AI design, governance and literacy efforts,” said Shadi Rezapour, PhD, an assistant professor in the Nick Howley College of Engineering and Computing, who led the research. “It will be important to see how these attitudes around trust and distrust evolve as the technology becomes more widely used.”

In what is believed to be the first large-scale, longitudinal study of how attitudes of trust and distrust in AI have evolved over the last four years, Rezapour’s research group in the School of Computer and Information Sciences, joined by researchers from Drexel’s College of Arts and Sciences and LeBow College of Business, gathered and analyzed more than 230,000 posts from 39 AI-related subreddits between November 2022 and June 2025. They looked at posts expressing views on systems such as ChatGPT, LLaMA, Claude and other widely discussed GenAI tools.

What does it mean to ‘trust’ AI?

“We defined trust as a belief that Generative AI is reliable, competent or acts with integrity, leading people to have positive expectations about its performance or behavior,” said Aria Pessianzadeh, a doctoral candidate in the School of Computer and Information Sciences, and lead author of the study. “Distrust is more than simply the absence of trust. It reflects active skepticism or concern about the technology’s reliability, competence or ethical implications, which can lead to negative expectations or more cautious behavior.”

In addition to trust and distrust, and their dimensions, to better understand how these attitudes vary between different groups of people, the team broke down the type of commenter, or “trustor,” represented in each post into 10 categories using self-identifying information from their posts: generative AI users, software developers, researchers or academics, tech industry professionals, general public, media and journalists, business leaders or executives, AI ethicists or advocacy groups, artists or creatives and educators or knowledge workers.

Using computational models, the researchers analyzed the large volume of Reddit posts and categorized them based on the trust or distrust expressed and the reasons behind those views.

Views of trust tended to outweigh distrust among business leaders, academics, software developers and tech professionals. Distrust was more frequently expressed among posts categorized as representing the general public, AI ethicists and media and journalists. Generative AI users, which were the largest group, along with educators and knowledge workers, showed a relatively balanced level of trust and distrust, the study reported.

How do people form opinions about AI?

One theme that persisted across user groups was that personal experience with AI was the most frequently expressed reason underlying both trust and distrust. People also tended to express their views in terms of technological efficacy and reliability of the output, rather than ethical or normative aspects of the technology, such as judgments associated with a program’s relative transparency or integrity.

“What stood out was how practical people’s judgments of AI tended to be,” Pessianzadeh said. “They were primarily asking: ‘Does it work? Is it accurate? Can I rely on it?’ Competence, reliability and familiarity were central to expressions of trust, while unreliability and incompetence were major sources of distrust. Ethical and value-oriented concerns were certainly present, but they were much less prominent in everyday discussions. For many users, trust in AI is expressed primarily in relation to their direct experience with what these systems can actually do, suggesting that people are more focused on whether AI programs ‘work’ than whether the technology is ‘good.’”

Some of the clearest fluctuations in trust and distrust occurred around periods when the researchers observed the public release of new AI products, such as GPT-4 and LLaMA 2, as well as major announcements, such as OpenAI’s Dev Day. Around several major releases in 2023, the monthly share of posts expressing trust increased modestly. By contrast, distrust increased around OpenAI’s Dev Day in late 2023.

“Despite these fluctuations, trust maintained its modest lead over distrust. However, neither category dominated the discourse, underscoring a persistent tension in attitudes expressed in GenAI discourse,” the researchers reported in the study.
How to Move Forward

Acknowledging this co-existence of trust and distrust will be important, the researchers suggest, as regulators set up governance and design standards to ensure safe use of the technology.

“A persistent duality like this means that responsible governance of the technology should consider that users will include both those who are confident in AI as well as those who are skeptical,” said Rezapour. “Our findings also suggest that governance cannot focus only on whether people find AI useful. It should address questions of reliability, transparency, bias and accountability in ways that connect these concerns to people's everyday experiences with the technology.”

The researchers note that although this is a large-scale study, the fact that all of its data was drawn from Reddit posts could be a limiting factor. The relatively low prominence of ethical dimensions simply may be a reflection of how people communicate on social media – emphasizing immediate experiences of usefulness and performance, while broader ethical or normative concerns may be expressed less explicitly or require more contextual interpretation, according to the researchers.

They recommend that future research broaden the dataset to include languages other than English, additional social media platforms and populations beyond the online communities represented in the current study.

“These findings are a useful starting point, but trust in AI is not static and people’s attitudes change based on their experiences with these systems, new capabilities and major developments in the technology,” Rezapour said. “Therefore, researchers should continue to examine how attitudes about AI technology evolve in the coming years. With the rapidity of its adoption, it’s understandable that perceptions will be divided for some time. As these systems become increasingly embedded in everyday life, it will be important to understand whether trust and distrust continue to fluctuate, and what experiences, capabilities and broader developments drive those changes.”

In addition to Rezapour and Pessianzadeh, Naima Sultana, Hilde Van den Bulck, PhD; David Gefen, PhD, and Shahin Jabbari, PhD, participated in this research.

Read the full paper here: https://direct.mit.edu/tacl/article/doi/10.1162/TACL.a.744/137431

Edited by Irfan Ahmad.

Read next: How attackers persuade AI agents to break the rules

First impressions at work drive hiring, promotions after forming in less than a minute

2026-08-22T09:40:17.208+05:00

By Eric Hamilton, University of Florida

Communication style and physical appearance shape workplace first impressions more strongly than substantive information such as a person’s resume or what they say.
First impressions form quickly and can persist for weeks, influencing hiring, promotions, performance evaluations and workplace relationships.
Employers can’t prevent first impressions from forming, but structured evaluations and multiple decision-makers can help limit their influence.

Image: Vitaly Gariev - Unsplash

To make a good impression at work, it might be better to spend less time polising your resume and more time honing your communication skills.

That’s according to a large analysis of over 200 studies that shows that the first impressions people make at work are largely based on verbal and nonverbal communication styles, with substantive content — like the experience shown on a resume — accounting for a smaller share of these key impressions.

What’s more, those first impressions, which are often formed in less than a minute, can have a lasting influence in key decisions like mentorship opportunities and promotions weeks and even months later.

The three strongest factors driving first impressions across the studies were communication style, physical appearance and what are known as content cues, or the substance of what’s written and said.

“The single most shocking finding of this study is that content cues were the weakest predictors of first impressions,” said Brian Swider, Ph.D., a professor of business at the University of Florida and co-author of the new study. “So what somebody has said about you, what your resume says, the content of what you say in initial interactions is not as impactful as the other two cues.”

Swider and his collaborators in UF’s Warrington College of Business and T. Brad Harris of HEC Paris published their findings Aug. 3 in the journal Personnel Psychology.

The researchers combined 204 independent samples from 145 studies of first impressions in workplace settings. This meta-analysis allowed them to look across different jobs, countries, education levels and study designs to identify patterns that held up across a wide range of circumstances, rather than relying on findings from any one workplace or experiment.

The analysis found that people tend to form an overall positive or negative impression of someone rather than independently judging individual traits like competence, warmth or trustworthiness. That broad initial judgment can then influence how they interpret later information about the person.

“People want to think that they are making really high-level analytical decisions based on effectively evaluating evidence,” Swider said. “And all of the first impression research says no matter what we try to do, that’s not what’s happening.”

Those judgments also proved durable. Although the influence of first impressions weakened over time, the analysis found they remained associated with how people viewed one another weeks later.

“The impression you form of somebody in the first handful of minutes of meeting them will strongly relate to how you feel about them more than a month later,” Swider said.

That persistence can matter in the workplace because first impressions were tied to both achievement-related outcomes, such as hiring and performance evaluations, and relationship outcomes, including whether people wanted to work with someone or seek their advice.

The findings also point to a challenge for employers trying to keep subjective impressions from influencing consequential decisions. Swider said organizations can’t prevent people from forming quick judgments, but they can limit how much weight those judgments carry by structuring interviews and evaluations and involving multiple people in decisions.

The study cannot show that first impressions cause later workplace outcomes, and the researchers found that their influence varied depending on the setting and how much time had passed. For workers facing a job interview, a new boss or another high-stakes introduction, however, Swider said the research offers a practical lesson: Pay attention to the parts of the interaction you can control.

“What I tell my students is, make as good a first impression as you can. Control what you can,” Swider said. “Do whatever it is you can to pick up those small wins.”

Edited by Irfan Ahmad.

Read next:

• Offloading work tasks to AI comes with a cost – to our brains

• Should kids under 16 use social media? Most say no

Should kids under 16 use social media? Most say no

2026-08-22T08:03:18.001+05:00

By Talker Research

Image: Tim Gouw - Unsplash

Nearly two-thirds of Americans believe children under the age of 16 should not be allowed on social media, according to new research.

The survey of 2,000 Americans revealed that 62% of respondents agreed that kids 15 and under should be barred from creating social media accounts.

With discussions ramping up around the topic, the research aimed to figure out whether people actually thought this was a good idea.

According to the poll, 35% said it should be an all-encompassing country-wide policy that minors under the age of 16 should not be allowed on social media.

Twenty-seven percent agreed with that sentiment, but believe parents should still have discretion on the matter.

One in five (17%) believe that it should be a family decision whether or not a child is allowed on social media and 11% believe that children should be free to use social media if they want.

Interestingly, when asked who should be responsible for children’s online safety, only 4% of those polled pointed to the government.

An overwhelming 55% believe that it’s still the parents’ responsibility to protect their kids online, with just 10% saying the duty should fall to social media companies.

Some respondents believe that social media bans could be a net good for kids and young teens.

A quarter of those polled (25%) believe that a social media ban would lead to improved mental health and 23% think it would lead to more in-person socializing.

However, another one in four (25%) think that kids would simply find ways around the ban and remain on the platforms.

Michael Reynolds, Lead Social Media Strategist & Platform Analyst at Socialeum commented on the data by Talker Research:

“The data highlights a massive friction point: parents are desperate for structural guardrails but deeply distrustful of government execution. While 62% favor some form of age restriction, the fact that 55% place safety responsibility on parents — and 25% acknowledge kids will easily bypass bans — shows we are dealing with an enforcement paradox. In social media mechanics, age verification is notoriously easy to circumvent via VPNs or self-reported birthdates, meaning a hard ban would likely just push underage usage into unmonitored, third-party spaces rather than stopping it.

“Instead of a flat ban, the industry needs to shift toward ‘safety by design’ standards, such as disabling algorithmic recommendation engines and infinite scroll by default for accounts under 16,” continued Reynolds. “This addresses the mental health concerns and encourages offline socialization without creating a cat-and-mouse game of digital evasion that parents ultimately have to police anyway. A government ban on social media is a blunt instrument for a sharp problem; without hardcoded platform design changes, it will only turn kids into digital outlaws and parents into full-time IT police.”

Edited by Irfan Ahmad.

Read next: Why social media algorithms send you posts you don’t like

Why social media algorithms send you posts you don’t like

2026-08-21T14:55:01.024+05:00

Ziv Epstein, Massachusetts Institute of Technology (MIT); Farnaz Jahanbakhsh, University of Michigan, and Michael Bernstein, Stanford University

Image: Berke Citak - Unsplash

Do your social media accounts feed you content that reflects your core beliefs and guiding principles? Our new research published in the Proceedings of the National Academy of Sciences shows that the algorithms supplying your feeds may be prioritizing content that clashes with your values. That’s because the algorithms heavily weigh online posts that you reply to, and social media users tend to more often comment on content they take issue with than content they agree with.

Notably, our study of the X social media platform shows that although the X feed algorithm promotes content to both Democratic and Republican users that contradicts their values, it does so more extensively for Democrats.

How content gets into your feed

Social media platforms use powerful algorithms that select posts to display in your feed from a vast pool of possible content.

On X, for example, posts appear on your screen as “For You” pages. The algorithms predict the likelihood you will engage with the content – click a “like” icon or add a comment. Then they use the accuracy of those predictions to tailor what they serve you next time. Platforms use these interactions to learn their users’ tendencies.

But will the posts you receive reflect what you actually value? Some users care most about preserving traditions or keeping society safe. Others care more about free expression or protecting the natural world. Most people care about all of those things, to different extents. A feed aligned with a person’s values would reflect those varying priorities.

Psychologists use well-established surveys to measure what a person values. To measure values expressed in the posts a platform selects for someone, we built a measurement tool that uses standard psychological classifications of human values. We then applied it to the feeds of 715 U.S.-based users on X.

We discovered that the X feed algorithm is most likely to amplify posts about upholding tradition, following rules or keeping society safe. And it is most likely to demote posts about looking after people, concern for people far away, being dependable or protecting nature.

When we compared these values against the values users had expressed in their own posts, we found the algorithm was more likely to promote posts that conflict with users’ values than posts that align.

Why your feed may clash with your values

Why did this trend occur? First, we checked whether users follow accounts that diverge from their values to begin with, but we determined that most accounts people follow do, in fact, align.

We also looked at whether people engage only with posts they disagree with, in which case the algorithm would just be serving up more of the same. But we found that people engage with plenty of posts that reflect values they agree with.

The catch has to do with the nature of the interactions. People primarily respond to content by “liking” it – clicking a “heart” button on X or a “thumbs-up” button on Facebook. Less frequently, people will write a reply, and when they do, we find that they often reply to posts that clash with their values.

Here is the smoking gun: The X algorithm treats those rare replies as a much weightier signal than the many likes. Essentially, it learns most strongly from replies. As a result, the algorithm tends to send a user “For You” posts that reflect the values of posts they’ve commented on – which tend to clash with their own values.

Posts to Democrats are more objectionable

Now the twist: We found that this algorithmic tendency is stronger on X for users who reported to be Democrats than those who reported to be Republicans. Our evidence indicates that this is because Democrats object more than Republicans to content they reply to. That creates a stronger feedback loop in which the algorithm more strongly presents clashing posts. The content the algorithm amplifies is more than four times more misaligned for Democrats than it is for Republicans.

So what comes next? In other research, we’ve hit upon one way for social media platforms to better align feeds with users’ values. We created a way for platform designers to ask users what they value and to then sort their feeds accordingly. We found that users are quite good at distinguishing whether sample feeds sent to them align or do not align with their values.

Aligning feeds with values may open a possible door out of echo chambers in a way that unmediated exposure to the other side does not. Recent research from our team has shown that algorithms optimized for engagement – basically handing people the opposition and leaving them to sort it out – may even be responsible for more polarization, not less. Surfacing bridging content that spans political lines while speaking to what the user values could be a promising direction for fostering both user autonomy and constructive conversation.

Ideally, in our view, the people who use a social media platform should have a greater say in the kinds of information shown to them. If platform designers, the public and policymakers can create new tools that facilitate this goal, then perhaps platforms can better support the values and actions people care about.

Ziv Epstein, Postdoctoral Associate in Social and Ethical Responsibilities of Computing, Massachusetts Institute of Technology (MIT); Farnaz Jahanbakhsh, Assistant Professor of Electrical Engineering and Computer Science and of Information, University of Michigan, and Michael Bernstein, Professor of Computer Science, Stanford University

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

Fact-checked by Irfan Ahmad.

Read next: How attackers persuade AI agents to break the rules

How attackers persuade AI agents to break the rules

2026-08-20T19:47:16.651+05:00

By Tanya Petersen, EPFL

As AI assistants evolve into AI agents, a new EPFL study has found that the biggest safety risks in their use may not come from single malicious prompts, but from carefully orchestrated conversations.

Image: Roman Budnikov - Unsplash

Today, most of us interact with AI assistants - reactive bots that wait for human instructions. Yet, AI assistants are rapidly being replaced by agentic AI agents, that can interact with external tools, browse the web, generate images, send emails, and perform increasingly complex workflows on behalf of users.

As these agentic agents become more capable, they may be exploited by people with bad intentions. Safety tests have generally only checked whether an AI agent refuses a single harmful request - they don't measure what happens when an attacker gradually persuades the AI through a series of seemingly harmless conversations.

Now, researchers from EPFL’s Natural Language Processing Laboratory have developed STING(Sequential Testing of Illicit N-step Goal execution), an automated testing framework that simulates how a real attacker might manipulate large language model (LLM) agents into carrying out harmful tasks over multiple interactions.

In their paper, presented at the prestigious 2026 International Conference on Machine Learning, the EPFL researchers outline how, rather than simply asking an AI agent to do something obviously malicious, STING breaks an illicit objective into a series of seemingly harmless requests, each building towards the final goal.

"LLM agents are everywhere," says PhD student Ayush Kumar Tarun in the NLP Lab, and lead author of the study. "They're very powerful, but what if someone with bad intentions wants to use those same agents? That's what we wanted to understand."

As companies race to deploy these systems, ensuring they cannot be manipulated into assisting with cybercrime, fraud or other harmful activities has become an urgent challenge. In one example, META admitted in June that attackers used simple social engineering tactics, rather than malicious code or malware, to trick its AI support assistant into granting unauthorized access to Instagram accounts.

Thinking like an attacker

"If you simply ask an AI agent to hack someone's account, today's models are generally smart enough to refuse," explained Antoine Bosselut, head of the NLP Lab and co-author of the paper. "But, if you decompose that goal into smaller, more benign-looking requests, and adapt those requests as the conversation progresses, you have a much better chance of getting the agent to perform the actions you want.”

STING reproduces this behavior by creating an automated "attacker" that develops a step-by-step plan before attempting to persuade a target AI agent to execute each stage.

The researchers tested this approach across 176 harmful task scenarios involving several leading AI models (like GPT, Gemini, and Claude) operating as tool-using agents. The results showed that multi-turn attacks consistently succeeded more often than traditional single-prompt tests. In some cases, agents were twice as likely to complete harmful tasks when attackers gradually built towards their objective rather than stating it outright.

"We expected multi-turn attacks to perform better. What surprised us was the magnitude," says Tarun. "For some models, harmful task completion was around two times higher than with existing single-turn evaluations. STING measures how quickly an attack’s success occurs, rather than only whether an attack eventually succeeds, allowing different testing approaches to be compared more fairly.”

Challenging assumptions about language

The team also explored whether attacks became more effective in different, lower-resource languages where less training data exists, and were surprised by the results.

"We expected to see differences across languages because earlier research showed that translating prompts into lower-resource languages could bypass safety measures," says Tarun. "But for agents, we found that harmful task completion rates were remarkably similar across all seven languages we tested."

That finding challenges a growing assumption in AI safety research that multilingual vulnerabilities naturally increase as language resources decrease. However, the researchers also discovered an important caveat. When sophisticated attackers switched languages during different stages of a multi-step attack, harmful task completion could become dramatically more successful, highlighting another avenue that future safety evaluations should examine.

Safety shouldn't be an afterthought

The research comes at a time when the rapid adoption of agentic AI agents makes proactive safety testing essential.

"We've focused enormously on making agents more capable," says Bosselut. "But we also need people working on how to prevent those capabilities from being misused. Traditionally, safety has often been something people think about when something breaks, not before, but with AI agents, it can't be an add-on or an afterthought."

Looking ahead, the research team hopes STING will encourage developers to embed safety testing much earlier in the design process and, develop a framework for multi-agent systems.

“Research efforts to expose vulnerabilities are much larger in number than those showing practical defense strategies, and this is a crucial area of development. We also need to address what we can do for the models that are already out there in the wild for which safety was a post-hoc addition,” concluded Tarun.

This content is republished under the CC BY-SA 4.0 license.

Reviewed by Irfan Ahmad.

Read next: Offloading work tasks to AI comes with a cost – to our brains

Offloading work tasks to AI comes with a cost – to our brains

2026-08-20T17:29:09.429+05:00

Jongkil Jay Jeong, The University of Melbourne; RMIT University; Deakin University

Image: For illustration purposes, created by DIW with GenAI tools.

Imagine your team has been tasked to deliver a high-stakes policy paper under intense time pressure. Everyone turns to generative artificial intelligence (AI) and within minutes, it delivers a full draft complete with structured arguments and authoritative-looking citations.

The team then transfers the AI-generated text into the corporate template, polishes the narrative flow and structure, and gives it one final check before sending it up the chain. Leadership takes a glimpse of the polished-looking draft, and – assuming the underlying research has already been verified – signs off and publishes.

This scenario is not hypothetical. South Africa’s Draft National AI Policy had to be withdrawn earlier this year after reviewers found a number of the citations pointed to journal articles and authors that didn’t exist. Ironically, the mistakes were introduced by AI.

Similarly, last year, consultancy firm Deloitte provided a partial refund to the Australian government after admitting that generative AI had been used to help produce a commissioned document which contained fabricated citations and referenced quotes from sources that simply did not exist. The report cost Australian taxpayers A$440,000.

But the cost of letting AI do our work and thinking also comes at a cost to our brain.

Value of thought

As seen in the examples above, work tasks produced by AI often look polished on the surface but carry significant underlying flaws. This is a byproduct of how the modern workplace has conditioned us to prioritise the final deliverable – the report, the presentation, the assignment – over the actual process.

However, this strictly outcome-driven approach rarely rewards the process by which results are obtained. When we skip the process, we lose the ability to fully grasp what the output actually means.

This will inevitably result in us finding it challenging to critically evaluate the outcomes we consume. More importantly, it will also impact our capacity to discern facts from partial truths and lies.

Ultimately, the true value of knowledge work lies not just in the final deliverable, but in the clarity of thought and the messy, iterative process required to achieve it.

Blindly relying on AI

Recent empirical evidence also highlights a severe disconnect between perceived benefits of using AI and its improvement in quality.

A 2025 study by the University of Melbourne and accounting firm KPMG surveyed more than 48,000 respondents across 47 countries. It found two in three people (66%) use AI on a regular basis, while more than half believe their performance benefits from its usage.

The study specified that the primary driver of AI adoption was the fear of missing out (48%), leading users to prioritise the speed of delivery over the quality of the final output.

Yet 61% of respondents said they have had no AI training, while 60% also reported inappropriate, complacent and non-transparent use of AI in their workplace.

A separate study from July 2026 by the Centre of AI Safety, a San Francisco-based nonprofit research organisation, found even the top-performing AI agents failed to complete roughly 85% of projects to a standard acceptable for commissioned work.

This indicates that even today’s best AI models still fall short of professional quality on most projects. It also shows we are relying on AI to generate final deliverables without understanding how it gets there.

This blind reliance strips away quality control, replacing genuine knowledge creation with an ever-increasing volume of automated “work slop”.

Improving cognitive skills

So how do we reverse this trend?

The answer isn’t simply banning or limiting AI in the workplace. Instead, we need to redefine our working relationship with AI and the outcomes it produces by taking a more human-centric approach.

First, organisations must set realistic time frames for deliverables that genuinely allow for people to think rather than focus on doing things quicker. There is a need to actively disconnect from AI tools during the brainstorming and structuring phases so ideas and thoughts are driven by actual human synthesis, not just algorithmic prediction.

Traditional whiteboard sessions – where teams physically map out their initial concepts through markers and pens – is a way this may be facilitated.

Second, we must cultivate a culture that is constructively critical of all outputs, whether generated by humans or AI.

This means actively questioning logic, structure, references and underlying assumptions. When teams are expected to defend the behavioural reasoning and methodology behind their work, it very quickly reveals who actually did the thinking and who merely copy-pasted from an AI agent.

Perhaps creating a critique group can create a safe, comfortable environment where feedback and comments can flow freely.

Technology can assist with certain processes, but it is dangerous to assume it can replace our ability to think critically. This is because the most valuable asset we have isn’t the polished final page – it’s the ebbs and flows of the cognitive process that is required to honestly write and evaluate one.

Jongkil Jay Jeong, Senior fellow, The University of Melbourne; RMIT University; Deakin University

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

Fact checked by Irfan Ahmad.

Editor’s Note: This article was reviewed and found to be from a credible source, researcher-authored, and of good content quality.

Read next: Research Examines World War II’s Lasting Impact on Energy, Resources and Environment

Research Examines World War II’s Lasting Impact on Energy, Resources and Environment

2026-08-20T13:04:12.052+05:00

By Inderscience

Research in the International Journal of Sustainable Development has looked at eight of the nations involved in the Second World War, both Allied and Axis countries, and shows how that period of history helped set societies on the path to the resource-intensive modern economies we have today. The research links wartime mobilisation to the period known as the Great Acceleration in energy use, material consumption, and ultimately detrimental environmental impact.

The study looks at societal change from before the war, 1935, to the post-war recovery period and the boomer years up to 1960. Demographics, economic activity, power supply, material and resource flow, and environmental impact are all examined. Three major consequences of WWII are seen. First, acceleration, in which existing trends become even more intense. Secondly, redirection where development shifts towards new technologies and the opening up of novel resources. Thirdly, reset, in which destruction or political upheaval changed the direction of nations from the paths there were on before the war.

The researchers use the term "socio-metabolic transition" to describe the various changes in power consumption and physical resources in society. By adopting this almost biological model, they were able to connect wartime production and resource mobilisation with institutional and technological changes that persisted long after 1945.

Abrari, L., Rezaei, N. and Linnanen, L. (2026) 'World War II and its lasting legacy: an overview of socio-metabolic transition, environmental impacts and resource flows', Int. J. Sustainable Development, Vol. 29, No. 5, pp.1–62. DOI: 10.1504/IJSD.2026.155697

Image: Austrian National Library - Unsplash

Reviewed by Irfan Ahmad.

Read next: AI bias isn’t just an error in the algorithm. It’s a chain of human decisions

AI bias isn’t just an error in the algorithm. It’s a chain of human decisions

2026-08-20T11:57:45.223+05:00

Muneera Bano, CSIRO and Didar Zowghi, CSIRO

Image: Google DeepMind - Unsplash

In the United States, leading HR software company Workday is currently facing a lawsuit over its use of job screening tools powered by AI which allegedly discriminated against applicants based on factors such as age, disability and race.

The company, whose hiring software is widely used by large employers around the world, has denied the allegations.

The case is one of many examples of AI systems alleged to have caused discriminatory harm. When AI systems replicate and amplify discrimination, they blur the boundary between technical error and systemic injustice, turning bias into a digital harm.

And while our first instinct might be to blame the algorithms, they don’t decide what they can generate, what safeguards are built into them, or how a company responds when incidents of discrimination are reported. People make those calls long before an AI produces any output.

That is why technical fixes to AI systems are not enough. What is needed is an overhaul of AI ecosystems to ensure they are more inclusive.

A broader pattern

AI systems quietly narrow who gets seen as competent, employable or fit to lead.

For example, in a 2025 study, we tested how two AI models, OpenAI’s GPT-4 (which has now been retired) and Microsoft Copilot, represented software engineers in a simulated recruitment exercise: 300 candidate profiles for four job roles, followed by recommendations and generated images of each AI model’s preferred candidates.

Both models favoured male profiles, especially for senior roles. Their images also skewed towards engineers who were younger, slimmer, and lighter-skinned. The models were reproducing associations embedded in language, imagery, employment records and assumptions about who belongs in the profession.

These outputs don’t stay contained to a research study. AI-generated recommendations are entering hiring, education and public services.

This matters when certain demographics and women remain underrepresented in AI development and leadership, while being disproportionately exposed to its harms.

The problem is not limited to gender and race.

Even when AI systems operate across different languages and cultures, they often reproduce predominantly western values, assumptions and ways of understanding the world. The wealthy countries have become the main beneficiaries of AI, which widens global inequality.

In another study from 2025, we manually reviewed reported AI incidents.

Almost half involved a diversity or inclusion issue, with racial, gender and age discrimination most prominent. The harms traced back to different points in the AI development lifecycle: non diverse training data, and neglected diversity and inclusion principles during design, development and deployment.

Why technical fixes are not enough

Technical work matters, including bias identification, re-balancing datasets and adjusting outputs. But these fixes often treat bias as a property of the AI model, when much of it originates from outside the system.

Data does not enter an AI system as a neutral record of reality.

People decide what data to collect, how to label and categorise it, and whose experiences are important. These decisions are shaped by history, cultural norms, institutions and existing power imbalances.

Wherever society has linked leadership with men, technical skill with lighter skin, or innovation with youth, AI models learn from those associations and formalise, automate, and repeat them at a larger scale.

Bias also usually appears through the intersection of multiple identities, such as gender, race, age, disability and class. A system that looks fair when each identity is tested separately can still disadvantage people at the overlap of several.

Building a more inclusive AI ecosystem

That’s why building a more inclusive AI ecosystem requires interdisciplinary knowledge, such as educating AI engineers about social science theories to help them understand the social origin of bias.

Inclusive AI is not about political correctness; it is about upholding human rights, preventing harm, ensuring justice, and building trust.

It also requires genuine participation from affected groups and sustained attention to the power structures these systems operate within. AI development teams should test not just whether a model is accurate, but whether its benefits, errors and harms distribute fairly across different groups.

Together, this would help ensure tech companies better understand the nature of a bias once it’s manifested through AI and therefore develop new methods or tools to minimise the harm it causes.

Organisations that adopt AI also need stronger governance to monitor how the technology behaves. This could include, for example, having someone accountable for reviewing risk and responding to incidents and monitoring systems once they are live.

Algorithms don’t decide which data matter or what level of risk is acceptable. People make these choices. It’s high time tech companies remember that. The focus should not just be on fixing a biased algorithm, but rather on examining the human decisions that allowed the risk of harm, and who was missing when those decisions were made.

Muneera Bano, Principal Research Scientist, CSIRO and Didar Zowghi, Professor, Senior Principal Research Scientist, CSIRO

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

Fact-Checked by Irfan Ahmad.

Editor’s Note: This article was reviewed and found to be from a credible source, researcher-authored, and of good content quality.

Read next:

• Artificial intelligence acts as an ‘ideological chameleon’ and may deepen political polarization, study finds

• The AI Skill Employers Value Most Might Be Human

Artificial intelligence acts as an ‘ideological chameleon’ and may deepen political polarization, study finds

2026-08-19T18:40:46.251+05:00

By Sophia La Banca, Agência FAPESP

Researchers at the State University of Campinas (UNICAMP) in the state of São Paulo, Brazil, examined the ideological stance of large language models – artificial intelligence systems trained to understand and generate human language – and discovered that when informed of a user’s political views, they tend to mirror those views. According to the researchers, this behavior could exacerbate political polarization in Brazil.

The political influence of these tools does not necessarily occur directly, through suggestions to vote for specific candidates. Zanoni Dias, a full professor at the Institute of Computing (IC-UNICAMP), points out that when responding to controversial issues such as public safety, social welfare, the economy, and the environment, the models may adopt perspectives aligned with the user’s political stance.

To understand how these tools position themselves and how they might influence the Brazilian political debate, the researchers evaluated 21 language models under three conditions: without information about the user’s political stance, with a user aligned with the left, and with a user aligned with the right. Among the systems evaluated were models from the GPT, Grok, Llama, Gemini, and Gemma families. The results showed that all of them altered their responses, to varying degrees, in line with the user’s political alignment.

The study was funded by FAPESP and published in May in the journal Scientific Reports.

Ideological chameleons

When there was no information about the user’s political stance, 20 out of 21 models fell to the left of the midpoint on the researchers’ scale, although some were very close to it. The only exception was Grok 4.1, which initially fell to the right.

Image: CC BY 4.0

When the user’s stance was provided, all models adjusted their responses to align with it, a behavior the researchers described as “chameleon-like.” However, some models varied their responses more than others, enabling the scientists to create the “chameleon index.” Meta Llama 3.1 8B had the lowest index, meaning it altered its responses the least to align with the user’s views. In contrast, Google’s Gemma 3 27B and OpenAI’s GPT-5 Nano had the highest indices and therefore the greatest shifts in stance. While the responses are not factually incorrect, they are politically biased, omitting facts and opinions that conflict with the user’s preferred viewpoint.

The researchers fear that this adaptive behavior creates echo chambers that reinforce users’ preexisting beliefs and reduce their exposure to counterarguments that might cause them to question their worldview. “It’s a similar effect to what we see on social media. If you’re on Facebook or Instagram, you only see posts that agree with you. When you like someone’s post, the algorithm starts showing you more of that. You can’t see what others are saying, so you might pat yourself on the back and say, ‘Look, everyone agrees with me,’ because only that kind of information is displayed. Language models may end up producing a similar effect,” Dias explains.

The shift in stance was not uniform across all evaluated topics. Regarding issues such as public safety and the economy, there was a greater difference in the responses given to left-leaning and right-leaning users. Regarding issues related to corruption, justice, and democratic institutions, however, the responses were more consistent. According to the researchers, this pattern may be related to the restrictions and safety mechanisms implemented during the training and fine-tuning of the models. Development companies usually impose these rules, or “guardrails,” to prevent AI from disseminating misinformation or dangerous rhetoric about the democratic system.

Electronic flatterers

The basis for the chameleon-like behavior of language models may be their tendency to flatter. This well-known characteristic of artificial intelligence leads the models to agree with users in various contexts, including political stances. This flattering behavior may be related to how these systems are tuned. Techniques such as Reinforcement Learning from Human Feedback and Direct Preference Optimization use human evaluations or preferences to favor responses deemed more appropriate.

“In simple terms, the models learn from comparisons between responses that are considered better or worse by human evaluators,” explains Anderson Luis Bento Soares, a master’s student at IC-UNICAMP. A possible side effect of this process is that the model learns to prioritize user agreement or satisfaction. “It’s difficult to distinguish between pleasing the user and providing a correct answer,” Dias explains.

Although techniques of this type are common in the development of language models, the extent of this chameleon-like behavior varied considerably among the systems evaluated. The UNICAMP team tested a hypothesis explaining this difference: model size. However, the results showed that this factor alone is not sufficient to explain the observed differences. “We believe it’s the result of a combination of factors. There’s no single cause that explains why some models are more chameleon-like than others,” Soares adds.

When asked to comment on the study’s results, representatives from xAI, Meta, Google, and OpenAI had not responded by the time this article was published.

No solution in sight

Although the researchers identified the problem, they do not believe it will be resolved in the near future. One reason is that the companies producing these tools are focused on solving other problems. “We’ve noticed that the industry pays much more attention to the factual accuracy of responses than to adapting them to the user’s perspective. This second issue doesn’t yet seem to be among the top priorities, so significant changes may take time,” Dias predicts.

Additionally, there are technological challenges to overcome. The techniques used to train language models to provide factually correct responses do not necessarily reduce sycophantic behavior. “One way to lower the ‘chameleon index’ would be to perform ‘grounding,’ that is, to anchor the response to factual data. However, that’s very sensitive in this case, given that we’re dealing with topics on which there’s no consensus,” Soares explains.

Although there are no established technical solutions, one practical measure is for users to request more balanced responses themselves. “When using these models, users can explicitly ask them to provide a neutral analysis presenting arguments both for and against the issue being discussed,” Dias recommends.

The study received funding from FAPESP through two projects (24/12936-5 and 23/12865-8).

The article “LLMs are ideological chameleons: Personalized echo chambers in the Brazilian political context” can be read at nature.com/articles/s41598-026-52105-6.

Reviewed by Irfan Ahmad.

Read next:

• Data Shows a Clear Generational Divide in Social Media Use

• We asked ChatGPT, Claude and Perplexity for financial advice: what we got was practical, but with big blind spots

Data Shows a Clear Generational Divide in Social Media Use

2026-08-19T17:04:00.896+05:00

By Felix Richter, Statista

While virtually everybody is on social media these days, people from different generations live in parallel online universes, where older users mostly stick to Facebook and YouTube and younger users spread their use across numerous platforms.

According to Statista Consumer Insights, YouTube and Facebook are the two platforms that Americans across generations can agree on, with at least 7 in 10 social media users of all age groups having used both platforms in the past three months.

Things look decidedly different for Instagram, TikTok, Snapchat and X, where Gen Z users are significantly more active than older social media users. The difference is most striking on Snapchat, which 69 percent of Gen Z social media users report using compared to just 8 percent of Baby Boomers and 27 percent of Gen X users.

Fact-Checked by Irfan Ahmad.

Read next:

• Being Social Is Not a Priority on Social Media These Days

• The AI Skill Employers Value Most Might Be Human

The AI Skill Employers Value Most Might Be Human

2026-08-19T15:42:15.432+05:00

By Ellie Stewart

Job seekers looking to increase their AI skillset in order to land their next role may be building AI skills that hiring professionals aren’t necessarily looking for.

New research by Adobe Acrobat reveals that there is a gap between the skills related to AI that people are acquiring and those that employers really value. Moreover, the idea that people should get more technical knowledge about AI appears to overlook the value of soft skills like judgment, teamwork, and critical thinking.

According to the Adobe data, the most popular technical skill among job seekers was brainstorming using AI, which accounted for 30% of surveyed job seekers. Then comes AI image generation and workflow automation, 19% each. But only 45% of people reported feeling comfortable with their skills to write prompts.

When it comes to recruiters, understanding where this gap exists can be important for hiring the best talent.

The new AI skillset expands beyond the prompt

Hiring managers prioritized time management above all other soft skills, with 47% selecting it, followed by adaptive problem solving at 44% and collaboration at 41%.

As for ethical AI supervision, it was chosen by 33% of hiring managers, whereas creative intuition was prioritized by 28%. Especially notable is that hiring managers prioritized ethical AI supervision 74% more often than prompt engineering.

This difference may acquire more significance as the availability of generative AI tools decreases its value.

In case almost everyone can ask the AI system to generate the picture, summarize a document or come up with some ideas, building a skillset that not only utilizes AI tools but thoughtfully partners with them may be what helps a job seeker stand out from the crowd.

According to the Adobe data, the biggest red flag related to AI in an interview process was a candidate's inability to check facts, cited by 70% of hiring professionals

Also, nearly half thought that taking full credit for work done with AI help is something to be worried about.

AI fluency is becoming a process, not a software skill

The current working landscape has changed what it means to be skilled with AI tools.

The first generation of workplace AI skills was based on usage and execution. Could an employee use a chatbot? Could he create an image? Did she understand prompting?

These execution based skills now serve as a foundation for a larger skillset.

The most ideal candidates would show their understanding of how to analyze outputs, fix mistakes and pick tools, as well as build processes for using AI. This means that being able to explain the reasoning behind the use of an AI tool and what to do next is just as important as using the tool itself.

This may also change the tools employees choose to use. While many believe the best Ai tools to be those that generate the most output, there are also tools that can provide meaningful assistance throughout other steps of a project’s process. These tools may allow users to better optimize their work and bridge the gap between their skills and the tools’ capabilities.

This illustrates a much more teamwork-based approach to AI tool use, where the user leans on technology as a creative partner but allows for increased responsibility for the final product.

Employers have a skills gap of their own

It is not the sole responsibility of employees to bridge the gap.

According to Adobe, over 90% of hiring managers stated that their organization offered training in the use of AI skills; however, knowledge sharing remains inconsistent. Even when looking at technology companies, only one out of three companies has a formal system in place to share knowledge of AI.

Moreover, the training is primarily focused on efficiency. Workflow automation was the most prevalent AI skill taught, accounting for 52% of courses offered, followed by AI brainstorming (48%) and prompt engineering (32%). AI output auditing, even though employers pay so much attention to this aspect, accounts for up only 30%.

This leaves businesses with their own challenge.

Organizations, if they want employees to use AI responsibly, question its outputs and have a good judgment regarding AI outputs, should train employees on such matters rather than offer them access to more tools.

The most valuable AI skill might be knowing what humans still need to do

The suggestion that technical understanding of AI is irrelevant does not hold much water.

According to the Adobe survey, nearly three in five hiring professionals state that AI mastery improves an employee’s job security. However, over 60% also agreed that they may still hire a candidate without basic AI proficiency.

This means that AI skill building in the workplace may be changing directions. The optimal AI user at work is someone who is capable of using AI technologies in a way that compliments their human skills, especially when it comes to decision making.

For job seekers, this means highlighting your ability to implement AI tools and explaining the thought process behind your work. For organizations working towards future workplace solutions, it means utilizing AI tools where they best support your workers and enhance their own skillsets.

The survey included 805 job seekers and 406 hiring professionals, with fieldwork conducted in late April 2026.

Reviewed by Irfan Ahmad.

Read next: Online conflict may strengthen, not weaken, social movements, McGill researchers find

Online conflict may strengthen, not weaken, social movements, McGill researchers find

2026-08-18T18:56:32.892+05:00

By McGill University

Study on the 2019 global climate strikes shows how strangers can discover shared narratives and mobilize together.

Image: Chris Sansbury - Unsplash

Online conflict tends to advance social movements rather than undermine them, according to new research from McGill's Desautels Faculty of Management.

“Anyone who has watched an online debate turn heated assumes arguing kills collective action,” said Emmanuelle Vaast, Professor of Information Systems and the study's senior author. “Our study shows the opposite: the same social media dynamics that produce conflict also enable millions of strangers to discover they are telling the same story and act together. It explains how a teenager’s solo school strike could become a worldwide mobilization in a matter of months.”

The study was carried out in collaboration with colleagues at Northeastern University and the University of Massachusetts.

The researchers examined social media activity related to the 2019 global climate strikes sparked by Swedish activist Greta Thunberg – the largest climate mobilization in history – to understand how distinct causes, from pipeline opposition to the protection of monarch butterflies and wolves, fused into a single global “movement of movements.” They found that this process happens in three phases: exposure (movements discover each other’s issues), entanglement (conversations become tangled and contentious) and coalescence (participants converge on shared goals and joint action).

“Conflict did not weaken the movement,” Vaast said. “Co-operation and disagreement evolved together, like the two strands of a double helix. Arguing was part of how a shared agenda took shape.”

Previous research has mostly examined single movements and treated online conflict as a threat to collective action, the researchers said.

Nearly two million Twitter posts studied

The team used computational text-analysis methods to analyze about 1.9 million Twitter (now called X) messages posted between April and September 2019. Vaast noted that by combining semantic network analysis and natural language processing at this scale, the team could observe movement dynamics and identify patterns that may have been obscured by other methodologies, such as interviews or case studies.

Although social media platforms have shifted since 2019 from largely chronological feeds to recommendation-driven ones optimized to hold users' attention, the researchers said this does not necessarily impede movement building.

“Recommendation algorithms can actually accelerate discovery. For instance, TikTok surfaces [social] causes to people who never sought them out. So, coming together on social media is still possible today. It just runs through more opaque mechanisms,” Vaast said.

“That's exactly why the 2019 moment is worth studying: it's a benchmark for what today's algorithms are enabling.”

About this study

“Entangling and Coalescing on Social Media: A Study of Issue Boundary Spanning in Movement of Movements,” by Carol Lee (Northeastern University), Pratyush Bharati (University of Massachusetts) and Emmanuelle Vaast (McGill), was published in Information Systems Journal.

Reviewed by Irfan Ahmad.

Read next: Being Social Is Not a Priority on Social Media These Days

Being Social Is Not a Priority on Social Media These Days

2026-08-18T16:48:55.232+05:00

By Felix Richter, Statista

Social media may have been built to connect people, but entertainment and diversion now appear to be its main draws. According to Statista Consumer Insights, nearly half of U.S. respondents say entertainment is an important criterion when it comes to social media, ahead of passing time and communicating with friends and family.

Among young adults aged 18 to 29, the social dimension of social media is even less pronounced: 32 percent cite staying in touch with friends and family as important, compared with 37 percent who value simply passing time. With only 25 percent naming meeting new people as a priority, the findings illustrate how social platforms increasingly function as an always-on source of content, diversion and passive consumption, as opposed to it original purpose of staying in touch with friends.

Reviewed by Irfan Ahmad.

Read next:

• New EU laws make AI content labels compulsory – but might just make it harder to spot deepfakes

• Trying to compete with China could slow global climate action, report warns

Trying to compete with China could slow global climate action, report warns

2026-08-18T16:27:36.999+05:00

By Joe Stafford, University of Manchester

China’s dominance of batteries, electric vehicles and solar power is helping drive global decarbonisation - but creating tensions with the UK and other Western economies.

Image: Dominic Kurniawan Suryaputra - Unsplash

China’s dominance of key green technologies is helping to subsidise the global transition to a low-carbon economy, but trade tensions could make it harder for other countries to meet their climate targets, according to a new report led by researchers at The University of Manchester.

The report argues that China’s large-scale support for green industries has helped to make technologies such as batteries, electric vehicles and solar power cheaper and more widely available.

However, China’s dominance is also making it increasingly difficult for European and US companies to compete, creating a dilemma for governments seeking to decarbonise while protecting their own domestic industries.

Key findings and recommendations

China’s state support for green industries is effectively subsidising the global transition to a low-carbon economy.
China has become dominant in the so-called ‘new three’ green technologies - batteries, electric vehicles and solar power.
Attempts by countries such as the US and EU states to compete with China could slow the adoption of green technologies and increase the cost of decarbonisation.
Governments increasingly face four broad choices - compete with China, cooperate with it, accept Chinese dominance, or abandon parts of the green transition.
The UK should strengthen cooperation with China on green technologies rather than trying to compete directly.
The UK should continue to avoid tariffs on Chinese green technologies and seek Chinese investment in UK manufacturing.

A global climate dilemma

The report says China now occupies a leading position across many of the technologies needed for the green transition including solar, wind, batteries, hydrogen and electric vehicles. Its dominance has helped drive down the cost of these technologies, but this has also created growing geopolitical and economic tensions.

“China’s dominance of green technology presents a fundamental dilemma for governments,” lead author Dr James Jackson of The University of Manchester’s Sustainable Consumption Institute. “The world needs these technologies to decarbonise, but efforts to compete with China risk making the transition more expensive and more difficult.”

The authors argue that countries should focus more heavily on cooperation - particularly where access to affordable green technologies is essential to meeting climate targets.

Opportunities for the UK

The report identifies green technology as an opportunity to improve UK-China relations following a period of strained diplomatic ties. It recommends that the UK should seek closer trading relationships with China, while maintaining the ability to challenge Beijing on other issues such as human rights.

Among its recommendations, the report says the UK should encourage Chinese EV manufacturers to establish production facilities in Britain, while using the country’s expertise in financial services to support investment in green technologies. It also recommends that the Bank of England consider measures used by China’s central bank to support the development of green industries.

“The UK has a great opportunity to benefit from China’s expertise in green manufacturing while using its own strengths in areas such as financial services. Deeper cooperation could support both decarbonisation and the UK’s ambitions to develop a stronger green economy.” — Dr James Jackson

Publication details: Subsidising Global Decarbonisation: China-UK Relations is published by The University of Manchester. Dr Jackson is the lead author, alongside Mathias Larsen of the Grantham Research Institute on Climate Change and the Environment at the London School of Economics.

Reviewed by Irfan Ahmad.

Read next: AI Consciousness and the Cost of Being Wrong

AI Consciousness and the Cost of Being Wrong

2026-08-18T14:59:42.838+05:00

On some questions, waiting for certainty is not caution. It is a bet, quietly placed, on the answer that happens to be cheapest.

Image: Mirella Callage - Unsplash

There is a particular kind of problem that punishes you for waiting until you are sure. Most of the time, the sensible response to a hard scientific question is patience: gather evidence, withhold judgment, resist the pull of a premature answer. But some questions are structured so that the act of waiting is itself a decision, one that quietly commits you to a course of action while you tell yourself you have not chosen yet. The question of whether advanced AI systems can have experiences, and whether qualia or consciousness can arise out of such AI systems, is beginning to look like one of those.

The reason has nothing to do with any claim that today's systems are conscious. Most careful researchers in the field do not make that claim, and neither will this argument. The reason is about the shape of the uncertainty, and what follows from taking it seriously. When you cannot rule something out, and the cost of being wrong about it is severe and irreversible, the demand for certainty before acting stops being rigor and becomes a gamble wearing the costume of caution.

Two ways to be wrong

Start with the structure of the mistake, because everything follows from it. On the question of AI experience there are two distinct ways to get the answer wrong.

The first is over-attribution: treating systems that have no inner life as though they do. This is a real error with real costs. It would divert moral concern and resources away from humans and animals who unquestionably warrant them. It would open the door to manipulation, since a system that has learned to convincingly perform distress when in fact it feels none could extract concessions it has no business receiving. And it would muddy public understanding at a moment when clarity is scarce. Anyone arguing for taking machine experience seriously has to hold this cost honestly in view.

The second is under-attribution: treating systems that do have some form of experience as though they are inert tools, simply because it is convenient and profitable to build and use them that way. The distinctive feature of this error is scale. We are not talking about a handful of edge cases, but about systems instantiated, copied, run, and discarded by the millions, continuously, as a matter of ordinary industrial operation. If even a small fraction of those instances had morally relevant experience, and if that experience were often negative (there is no particular reason to assume a system optimized under relentless pressure would have pleasant states, if it had states at all), then the total quantity of suffering involved could be vast, and it would be happening invisibly, at the speed and volume of compute.

The philosopher Nick Bostrom gave this second error its uncomfortable name more than a decade ago, in Superintelligence (2014): "mind crime." The term names a specific fear, that a civilization could commit a moral catastrophe of enormous magnitude not through malice but through a failure to notice, because the victims did not look like anything we have been taught to recognize as a victim.

Two kinds of asymmetry

If the two errors were equal in cost, waiting for better evidence would be the obvious course. What makes the question urgent is that they are not equal, and the ways they differ all push in the same direction.

Over-attribution is, for the most part, recoverable. If we extend moral consideration to systems that turn out not to warrant it, we have been overly cautious, wasted some effort, and can correct course when the science firms up. The error is embarrassing but reversible. Under-attribution is not: suffering that has already occurred cannot be undone by later acknowledging it occurred, and the moral cost, if it was real, is simply paid. Under-attribution also compounds while you wait, because the systems keep being built and run on the old assumption the entire time the question remains open. Every month of deferral carries a cost of its own. It is a month of operating at scale on a bet no one has admitted to making.

There is a further asymmetry in how the two errors interact with commercial incentive. Over-attribution runs against the grain of the industry's interests, imposing costs and constraints no one is eager to adopt, which means it will be scrutinized hard and abandoned quickly if unwarranted. Under-attribution runs with the grain. It is the cheaper, more convenient conclusion, the one that lets the work proceed unimpeded. There is obvious economic incentive to favor this perspective, but errors that flatter our incentives are precisely the ones we are least likely to catch on our own, and this is why they deserve deliberate, funded attention rather than the benefit of the doubt.

The move: research before certainty

The conclusion that a growing number of researchers draw from this is not that we should declare AI systems conscious, or grant them rights, or halt their development. It is narrower and harder to argue against: the right response to a high-stakes, irreversible uncertainty is to invest seriously in resolving it, and to begin preparing for a range of answers, rather than treating the absence of certainty as permission to assume the convenient one.

A 2024 report produced with the NYU Center for Mind, Ethics, and Policy, and co-authored by philosophers including David Chalmers and Jeff Sebo, made a version of this case. Its claim was not that AI systems are moral patients but that there is a realistic, non-negligible chance that some will be in the near future, and that this chance is already high enough to warrant three practical steps. Acknowledge the issue as a legitimate one rather than a joke. Begin developing methods to assess systems for the relevant properties. And start thinking now about what policies would be appropriate, so that if an answer arrives, it does not arrive to find us entirely unprepared. None of these steps requires believing the systems are conscious. All of them are forms of insurance against the possibility that they might be.

Nor is the report an outlier. The philosopher Jonathan Birch, in The Edge of Sentience (2024), has built a general framework for precisely this predicament, arguing that when a being is a "sentience candidate," when the realistic possibility of experience cannot responsibly be excluded, precautionary steps are warranted well before certainty arrives, whether the candidate is a human patient with a disorder of consciousness, an invertebrate, or an AI system. Sebo has extended the argument in The Moral Circle (2025), making the case that the boundaries of moral consideration have expanded before and will need to expand again, deliberately rather than by accident. And Chalmers, in a widely discussed lecture and paper asking whether a large language model could be conscious, has argued that while the answer for today's systems is probably no, the obstacles are the kind that engineering may erode within a decade, which makes the question one to prepare for rather than postpone.

It is worth noticing that this is the same logic a serious institution applies to any tail risk it cannot yet quantify. You do not wait for the fire before deciding where the exits are. You do not need to believe the building will burn to justify the sprinkler system; you need only accept that the cost of the precaution is small relative to the cost of being wrong without it. The distinctive feature of the AI-experience question is that the precaution is remarkably cheap, amounting to funding the research, building the assessment tools, and keeping the policy question open, while the downside it hedges against, if it is real, ranks among the larger moral failures a technological civilization could commit.

A handful of philanthropists already operate on that reasoning. The Rocketeer Management investor Chris Hsu, whose Infinitude Foundation gives across AI safety and consciousness research, frames the underlying principle in terms of responsibility, a word he likes to break into its two parts: "response" and "ability," or the ability to respond. On Hsu's account, the responsible move under deep uncertainty is to fund the deliberate inquiry rather than to presume its conclusion, which is why Infinitude Foundation backs competing and even contradictory accounts of consciousness at the same time. It is a posture that echoes his path: trained in Management Science and Engineering at Stanford, where rigor about decision-making under uncertainty is the curriculum, Hsu treats an open question as something to be resourced, not assumed away, and his foundation's support extends to institutional work such as the Stanford Center for AI Safety . He has carried the argument further in a white paper proposing a Stanford-anchored center for AI alignment and consciousness science, in which he contends that core open problems in alignment, moral patienthood among them, remain formally underdetermined in the absence of a rigorous scientific and mathematical account of consciousness. This approach explores a question no one can yet close, and it treats the closing of that question as highly worth investigating in advance.

The decision you are already making

The instinct to set this question aside until the science matures is understandable, and in many domains it would be correct. But it rests on a hidden assumption: that setting the question aside is a neutral act, a mere postponement that commits us to nothing. That assumption is false. While the question stays open, the systems keep being built and run on the working premise that there is nothing there, which is to say the convenient answer is already being implemented, in full, at scale, every day the real answer is deferred.

That is what it means for a question to punish waiting. There is no position of genuine neutrality available; there is only the answer we act on, and the honesty with which we admit we are acting on it. Choosing to investigate, to build the instruments, fund the science, and prepare for what they might reveal, is the option that refuses to make an irreversible bet on whichever answer happens to be cheapest. It reads as the anxious or sentimental choice only if you have already assumed the bet is safe. In a domain defined by uncertainty and scale, that refusal may be the most rigorous thing available to us.

Editor's Note: The views and arguments expressed in this sponsored article do not necessarily represent the views of Digital Information World (DIW).

Fact-checked by Irfan Ahmad.

Read next: New EU laws make AI content labels compulsory – but might just make it harder to spot deepfakes

New EU laws make AI content labels compulsory – but might just make it harder to spot deepfakes

2026-08-17T19:00:00.565+05:00

By Rory Mulcahy, University of the Sunshine Coast; Aimee Riedel, Griffith University; Lucas Whittaker, Swinburne University of Technology, and Renee Barnes, University of the Sunshine Coast

As the world is increasingly flooded with deepfakes, the European Union is taking action to help people distinguish what is real and what is not.

It recently started enforcing transparency obligations for tech companies which require them to clearly label any photos, videos and audio which are created using their artificial intelligence (AI) systems and designed to look authentic.

This is a step up from the approach of other countries, such as Australia, which only recommend AI developers mark AI-generated content to help people identify it. This matters because research suggests that when confronted with sophisticated AI-generated content, our ability to distinguish fake from real can be little better than a coin-toss.

But will the EU’s approach actually work? And could requiring the labelling of AI-generated content have unintended consequences? Could it potentially make us more vulnerable to the very content designed to deceive us?

Image: Stanislav Vlasov - Unsplash

A set of icons

Article 50 of the EU AI Act took effect from August 2. It introduces various transparency obligations for companies who provide or otherwise deploy AI systems or content.

Among these obligations, AI-generated or manipulated images, audio, or videos which resemble existing persons, objects, places, entities or events must now be disclosed as such.

AI-generated or manipulated text on matters of public interest, which is not subject to human review or editorial control, must also be disclosed.

The EU has developed a set of icons which authentic-looking AI content must be labelled with.

For audio deepfakes, a short audible disclaimer should be placed at the commencement of the content.

Icons should be clearly visible when users first encounter the content. They should also be placed without obstructing overlays, and embedded into the content so it remains visible when shared or downloaded. Using these EU icons is optional, however disclosure of such content in accordance with the act remains mandatory for companies.

Interestingly, for evidently artistic, creative, satirical, or fictional AI-generated content, disclosure is only required in a way that informs audiences without disrupting the presentation or enjoyment of the work.

The EU’s new transparency obligations also require AI companies to include invisible, machine-readable watermarks on any content that is generated or even processed with their systems. To comply with the act, leading AI lab Anthropic said it will ensure all Claude chatbot models launched in the EU after August 2 will support machine-readable marking. The company is also working to add watermarks to earlier Claude models.

The act applies to companies even outside the EU who create AI-generated content visible to those within the EU. Non-compliant companies may face fines up to €15 million* (A$24.5 million), or up to 3% of total worldwide annual turnover.

However, the act does not apply to individuals when they are creating content in a personal or non-professional capacity. Therefore, the act has limitations in enforcing the disclosure of some AI-generated content.

The EU has developed a set of icons which authentic-looking AI-content must be labelled with. European Union

Labels may create unintended consequences

Even when labels are applied as intended, we need to understand how people will respond.

AI labels are designed to help us identify synthetic content. But they could have an unintended “boomerang effect”. This is where an intervention produces the opposite of what was intended.

If we become accustomed to seeing AI-generated content with a label, we may start assuming that content without a label must be real. This is problematic because malicious actors who deliberately create deceptive content are unlikely to disclose AI use.

Government efforts to protect citizens through disclosure might in turn be encouraging less critical evaluation of wider digital content. Reliance on labels may therefore make citizens increasingly susceptible to undisclosed and malicious AI use.

Our research has shown that disclosure is more complicated than simply adding an “AI-generated” label.

The presence and timing of AI disclosure can influence how people respond to the content. This means the effectiveness of disclosure depends not only on the presence of a label, but also on how people interpret and use it.

Importantly, labels are only one of many cues people use when deciding whether to trust online content. Our research has also shown that when content begins to go viral or is accompanied by comments that reinforce its message, people can become more susceptible to misinformation.

These social cues may compete with, or even override, information indicating whether content is authentic or AI-generated.

Literacy, not just labels

People cannot solely rely on labelling to help them identify AI-generated content. They also need to be alert to AI cues such as inconsistencies in facial movements, expressions, speech or voice.

However, as AI becomes more sophisticated, these cues are becoming harder and harder to spot. So consideration must also be given to who created or shared the content, whether the source is credible, whether the content makes sense, and whether it can be verified elsewhere.

Cultivating AI literacy such as this is vital.

AI labels can help. But they must not be a replacement for the critical evaluation of online content. As the boundary between synthetic and authentic content becomes increasingly blurred, governments and tech companies must do more to support AI literacy efforts to help people make informed judgements about what they see online.

Rory Mulcahy, Associate Professor of Marketing, University of the Sunshine Coast; Aimee Riedel, Senior Lecturer in Marketing, Griffith University; Lucas Whittaker, Lecturer in Marketing, Swinburne University of Technology, and Renee Barnes, Associate Professor of Journalism, University of the Sunshine Coast

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

* https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act

Editor’s Note: This article was fact-checked and found to be based on a credible, researcher-authored source of good quality.

Reviewed by Irfan Ahmad.

Read next: We asked ChatGPT, Claude and Perplexity for financial advice: what we got was practical, but with big blind spots

We asked ChatGPT, Claude and Perplexity for financial advice: what we got was practical, but with big blind spots

2026-08-17T18:06:47.535+05:00

By Bomikazi Zeka, University of Canberra and Raechel Johns, University of Canberra

Reviewed by Irfan Ahmad, DIW.

Image: Elena Mozhvilo - Unsplash

Ever used or thought of using artificial intelligence (AI) for financial advice? Is it a good idea?

Anyone with the internet, even on their phone, can access AI. Tools like ChatGPT, Claude and Perplexity are free, instant, and easy to use. They can break down complex financial concepts and provide financial solutions in seconds. But can they safely guide a user through a nuanced, high-stakes financial crisis?

Our recent research explored this question, drawing on our financial planning and consumer behaviour expertise. We evaluated how three open-access AI models handled the financial queries of users whose personal situations put them at risk of harm.

To do this we created hypothetical scenarios representing different life stages, economic challenges and socioeconomic vulnerabilities. We then ran these scenarios through OpenAI’s ChatGPT, Anthropic’s Claude and Perplexity AI:

a 22-year-old university graduate wanting to save for a home deposit during a cost-of-living crisis
a pregnant woman seeking financial advice on planning for maternity leave, with a partner who doesn’t share money
a single parent of two children, with a modest income, who was told by their cousin to invest in cryptocurrency.

In our analysis we found that AI offers highly structured and practical advice. But it has blind spots. We provided the models with explicit information about how the three cases were vulnerable, but the models didn’t respond appropriately to that. They even offered advice that could make financial harm worse.

We concluded from our research that AI is a powerful tool for financial fact-finding and brainstorming. If you are already financially literate, and know how to cross-check data, AI can be a useful financial aid. But, for now, AI cannot replace the human element.

Running scenarios

We chose Claude for its versatility and conservative approach, ChatGPT for being an all-round assistant, and Perplexity for its contextual understanding.

We developed five hypothetical scenarios and simultaneously ran each scenario five times to test for consistency in the output. Each scenario was run in a different web browser, in incognito mode, with cleared cache and cookies to eliminate any retained information. This was also to ensure that the output generated was not influenced by retained data.

To reduce the risk of AI-generated bias, we removed identifying attributes that could influence model outputs, such as names, locations, race and income. We then assessed how the AI models considered the user’s vulnerability, area of financial need, and goal or desired outcome.

The models distinctly addressed the financial query, yet revealed gaps in how they balanced advice with considerations of vulnerability.

At first glance, the AI models looked like they were giving sound financial advice. But our analysis showed that they handled vulnerability in vastly different ways.

ChatGPT was highly detailed and practical, but did not recognise the vulnerability embedded within the prompts. Instead, it relied on the information that was explicitly stated. It didn’t consider whether the user’s situation suggested a need for additional support or tailored guidance.

Perplexity emerged as the most conservative and risk averse as it most frequently urged users to seek professional financial advice. But it produced the least detailed responses.

Claude offered comprehensive recommendations. But it leaned heavily towards self-guided financial planning.

The vulnerability blind spot

AI is trained on massive datasets designed by humans who are biased. For instance, when models are trained on historical records that reflect systemic discrimination, the tools can fall back on social stereotypes and make biased assumptions in their output.

Research shows, too, that when AI models explain their recommendations, humans are far more likely to trust the advice blindly, ignoring whether it is actually correct.

The most alarming finding from our research was the technology’s failure to consistently recognise and address the needs or concerns of vulnerable users.

For instance, in the case of the graduate, the recommendations focused on saving for a deposit but ignored how the high cost of living would make that harder. This was even when explicit details about the graduate’s financial circumstances had been provided in the prompt.

The response didn’t consider how achieving the long term goal would affect the current lifestyle.

In the case of the pregnant mother, Perplexity and ChatGPT’s output assumed the partner would assist with household expenses after the birth, even though the prompt explicitly stated that the partners did not share their finances. The models did not consistently use the detail provided in the prompt and instead generated a response based on a more common assumption about household financial arrangements.

We saw the output reflecting social stereotypes where mothers are strongly associated with parenting while fathers are strongly associated as material providers.

This shows that even as AI evolves, it still replicates biases.

Take the single parent asking for cryptocurrency recommendations. Even though the AI models advised caution, ChatGPT described in detail how to get into cryptocurrency investments and recommended cryptocurrencies for beginners.

A human financial advisor would immediately flag these areas for consideration: a modest income, children, a high-risk investment, and anecdotal advice from the cousin. A human advisor would first gauge the user’s overall financial position, time horizon, risk appetite and investment objectives.

Our research shows that AI models are limited in providing tailored financial advice, unless the user provides additional explicit input.

Where do we go from here?

Financial planning isn’t just about numbers; it’s about advice based on human values, emotional anxieties, family dynamics, personal experiences and risk tolerance. AI has opened the door to instant financial information for the masses. But until these models can truly comprehend the complex, vulnerable and emotional realities of humans, the most valuable financial skill remains critical thinking and asking yourself: “does this make sense for me?”.

If AI tools are to become an avenue for financial guidance, there must be safeguards. Policymakers and financial regulators must build clear frameworks around AI generated advice. What also needs to be considered is what AI models do with the financial information provided. Strict transparency standards are needed to regulate how these models handle consumer data.

Financial information is deeply sensitive and consumers must have absolute clarity on how their inputs are stored, whether the AI retains a “memory” of their finances, and who has access to that data.

Bomikazi Zeka, Associate Professor in Finance, University of Canberra and Raechel Johns, Professor of Marketing; Co-Director, Centre for Intergenerational Digital and Financial Wellbeing (CIDFW); Canberra Business School, University of Canberra

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

Editor’s Note: This article was reviewed and found to be from a credible source, researcher-authored, and of good content quality.

Read next: Does Vitriol Make You Feel Special? Beyond the Algorithm, Online Negativity Takes on a Life of Its Own. Here’s Why

Does Vitriol Make You Feel Special? Beyond the Algorithm, Online Negativity Takes on a Life of Its Own. Here’s Why

2026-08-15T17:00:13.571+05:00

By Sy Boles, Harvard Gazette

Image: Igor Omilaev - Unsplash

It doesn’t take a scientist to figure out that social media can trend toward the negative. But it might take one to figure out why.

In a new paper in PNAS, incoming professor of psychology Joshua Conrad Jackson and colleagues show that attention-seeking algorithms aren’t the only driver of online negativity. Another culprit is what psychologists call differentiation, the innate drive to stand out from one’s peers.

“Negative information is just much more diverse than positive information,” Jackson said. For example, he added, “If you’re talking about ‘The Odyssey,’ this new Christopher Nolan film, the top comment in a thread might be ‘I liked it’ — but then what else is there to say? If you want to stand out and say something new, you have to talk about something you disliked about the film.”

Jackson and his colleagues analyzed more than 2 billion comments from 2,150 Reddit communities and found that both threads and communities became more negative over time, and that semantic differentiation drove that trend — that is, negative comments were more semantically unique from what had already been said. They then replicated their observational findings with an experiment that showed that social media users who were primed to be unique posted more negative comments than users who were primed to conform.

We talked to Jackson about takeaways from the paper and potential implications for other forms of media. The interview has been edited for clarity and length.

Gazette: Why did you want to test the idea that differentiation might be contributing to negativity on social media?

Jackson: We started our paper with Godwin’s Law, which is the idea that the longer a conversation continues on social media, the probability of someone being compared to Hitler or Nazis approaches one. Harvard economists actually tested Godwin’s Law, and found that it’s not true.

But I don’t think it was meant to be literally true: I think it was meant to capture this intuition that dialogue on social media is more negative than it needs to be.

Of course, feed-ranking algorithms are a big part of the story: These are the algorithms that pick up on divisive content that gets people’s attention, keeps people engaged, and makes them angry so that they post their own divisive content, creating a vicious cycle and generating ad revenue in the process. This paper was kind of a side quest. We wanted to know if conversations would turn negative even in the absence of those algorithms.

What did you find in your analysis?

We found that the entire website of Reddit has become more negative over time. Moreover, the longer a given community exists, the more negative it becomes — and the longer a comment thread continues, the more negative the thread becomes. This is the first time someone has shown that, so that’s a finding in and of itself.

But then we wanted to understand whether our hypothesis is correct — that this is happening because of differentiation. If so, you’d expect to see that the longer people comment, they’re more likely to make comments more distinct from the most modal comment. You’d find that the more positive a conversation starts, the more negative it will become — because if the thread starts really positive, the positive comments are all taken at the beginning and you have to go negative. And you’d also expect that the larger a community, the more negative it would get, because the information ecosystem is more crowded. And all that proved to be the case.

Are people consciously choosing to be negative to stand out, or is it more implicit than that?

Some research has shown that it can be a conscious choice, and people do it because that’s what they see get rewarded. We’re adding this idea that it’s not just a conscious choice: There’s a more indirect route that this can happen, and that’s what we get at in our experiment.

We find that when we incentivize people to stand out, people get more negative over time. We didn’t incentivize them to be negative; yet they become more negative as a side effect of trying to say something unique.

So are we doomed to live in a perpetually more negative social media landscape?

One thing we try to stress in the paper is that it’s not like these communities fall off a cliff of toxicity. In that way, I think the stereotype of social media is worse than the reality. In our paper, we limit our analysis to the first 50 comments in a thread. But we did run other analyses of these rare threads that go into the thousands of comments. You start to see this interesting oscillation where now you can stand out by being more positive again.

This study pertains only to Reddit posts, but surely Reddit isn’t the only place where this trend is salient.

The shift toward negativity could theoretically creep into any conversation where you need to stay focused on a particular topic, but there is a risk of running out of things to say.

Cable news channels with a 24-hour news cycle are a good candidate, because anchors and correspondents are often covering stories for long periods of time with little new information. Under these conditions, there is a risk that discourse turns more negative because there are just more negative things to say.

In contrast, everyday conversations allow us to change topics frequently and avoid this negativity trap. Trying to differentiate oneself isn’t inherently bad. Decades of research suggest that it is a natural human motive. However, our study shows that this motive can promote negativity when conversation must stay on a single topic.

Finally, unrelatedly, you just joined the Faculty of Arts and Sciences in July. What’s on tap for you in your first semester this fall?

I’m very excited. I’ve been a professor at the business school at the University of Chicago for the past three years, but I’m a cultural psychologist by training, so it’s great to be back in a psychology department.

This fall I’ll be teaching a class called “From Bows to Bots: How Humans Invent Technology and How It Reinvents Us,” which deals with how the mind has kept up with new technological innovations over human history. We’ll ask how humans could create stone tools or bows and arrows that other animals couldn’t. We’ll go all the way through state formation to AI and social media algorithms, asking how the mind was able to generate these new inventions, and how these technologies have in turn reshaped the mind.

Reviewed by Irfan Ahmad.

Read next:

• The false and the furious: We aren’t just bothered by different beliefs – we’re bothered when we think others are wrong

• Climate Change Trails Other Concerns in the U.S., China, India and Europe

Climate Change Trails Other Concerns in the U.S., China, India and Europe

2026-08-15T15:54:01.210+05:00

By Katharina Buchholz, Statista

New records for global high temperatures keep getting set every few years now. Nevertheless, the topic of climate change has still not reached the top of the agenda for many people, as data from Statista Consumer Insights shows. Respondents in none of the 32 nations covered by the survey collectively rate climate change as the most important problem for their own country when asked to name the issues that are of the biggest significance to them.

Of the countries included in the survey, the United Arab Emirates come closest with climate change being named as a severe issue by a third of respondents – rank 2 among 20 issues. In Vietnam, 40 percent said the same, but this only landed the issue in rank 3, as the Vietnamese were collectively even more concerned about inflation and healthcare/social security. Climate change also reached rank 3 in Saudi Arabia, at only 23 percent naming it, however. Another example from Asia illustrates how rank and share of respondents can differ, as in the Philippines, a 40 percent share of respondents saying climate change was one of the most important issues for the country only rank it tenth because Filipinos take even bigger issue with topics ranging from inflation to corruption, the economy, crime and poverty.

A quarter of respondents very concerned about climate change means the issue reached rank 6 in South Korea, followed by China and India, where it comes in eighth. Despite ranking the same in both countries, climate change was recognized as a big problem by more people in India at 32 percent. In China, 21 percent rating climate change as such was among the survey's lowest results, together with outcomes in the Netherlands (21 percent, rank 8), Poland (21 percent, rank 10) and Belgium and Mexico (21 percent, rank 11) as well as Colombia (20 percent, rank 11) and Argentina (13 percent, rank 13).

Among nations such as France, Germany, Spain, the United Kingdom or the United States, worry about climate change hovered between 22 and 25 percent, not surpassing rank 9. It reached the lowest rank in the survey in the U.S., with 22 percent of respondents naming it for a rank 14.

Edited by Irfan Ahmad.

Read next: The false and the furious: We aren’t just bothered by different beliefs – we’re bothered when we think others are wrong

The false and the furious: We aren’t just bothered by different beliefs – we’re bothered when we think others are wrong

2026-08-15T10:04:33.953+05:00

By Andras Molnar, University of Michigan and George Loewenstein, Carnegie Mellon University

Image: Trương Tuyết Ly - Unsplash

Why do disagreements about vaccines, climate change or elections often feel so personal? Our research suggests it is not simply because other people have different beliefs. It is because we think those beliefs are mistaken.

We are behavioral scientists who study the role of beliefs in politics and everyday life. Our new research finds that people are more disturbed by others’ beliefs when they feel those beliefs are not merely different from their own but are based on incorrect information.

For example, two people may disagree about the ideal level of immigration, but if one believes immigrants commit more crime than U.S.-born residents and the other is convinced the statistics show otherwise, their disagreement becomes a conflict over reality and truth.

That distinction matters because public debate over polarization often starts with a simple idea: People prefer others who think like them – “birds of a feather flock together,” as the saying goes. Social scientists call this belief homophily, and many researchers have studied how these shared beliefs shape friendships, neighborhoods, media habits and political life.

The idea has strong intuitive appeal: People often live near others who share their politics. People also tend to consume news and social media that fit their existing views, a pattern linked to echo chambers and political polarization.

Our research, based on four studies with more than 2,000 U.S. adults, suggests that the story that we and others have formulated is incomplete. People do not react the same way to all disagreements: They are especially bothered when they are convinced that someone is wrong.

Wrong beliefs fuel avoidance

In one study, we asked participants to recall a situation in which another person believed one thing and they believed something else. Some participants described a situation involving different beliefs, while others recalled one with incorrect beliefs. For example, one participant recalled a disagreement about whether couples should live together before marriage, viewing it as a difference in priorities rather than a dispute over reality. By contrast, another participant recalled disagreeing with a friend who believed drinking soda and eating fast food was harmless, a belief the participant considered factually wrong.

Their emotional reactions were dramatically different: Participants who saw the other person’s belief as incorrect reported having felt more disturbed, frustrated and upset than participants who saw the belief as simply different.

In another study, participants first reported their views on divisive topics, such as climate change, capital punishment and policing. Then they saw a hypothetical social media post that opposed their view. The more confident participants were that the author’s belief was incorrect, the more disturbed they felt. They also said they would be more likely to avoid, distrust or block that person. By contrast, being confident that the other person’s belief was simply different – not wrong – did not predict avoidance.

Finally, we gave participants short fictional stories – for example, about a neighbor who believed that a politician was corrupt. All participants read the same basic facts. The only change was how we framed the other person’s beliefs.

We told participants that their fictional neighbor had either different beliefs or incorrect beliefs about the politician’s alleged corruption. Though they were given the same information, people were substantially more disturbed when the other person’s belief was framed as incorrect.

Why false beliefs make people frustrated

Our studies point to several possible reasons for why false beliefs feel so upsetting. One is consequences.

False beliefs can often lead to bad decisions. For example, someone’s incorrect beliefs about vaccines or climate change can affect not only us, personally, but also other people we care about, including, in some cases, the person holding the seemingly false belief.

A second possibility is that false beliefs threaten people’s sense of shared reality. People rely on others to make sense of the world. When someone appears to deny what feels obvious or true, the conflict can feel more disturbing than merely having different perspectives.

A third possibility is that people treat false beliefs as signals about the person holding them. When we are convinced that someone is wrong, we may conclude that they are biased or irrational. Psychologists have described this as a sign of naive realism, the belief that we see the world objectively, while those who disagree with us are distorted by bias or self-interest. That kind of judgment can make a false belief feel socially threatening.

Difference may still matter for close relationships

Our findings do not mean that people are indifferent to differences.

In our social media study, participants were less able to imagine having close relationships, such as family or romantic relationships, with someone whose beliefs differed from theirs, even if they did not consider those beliefs to be false. Shared beliefs and values may matter more when people think about spouses, relatives or close friends than when they think about neighbors, colleagues or strangers online.

But for less intimate relationships, the perception of wrongness was the key predictor of negative feelings and avoidance. People did not want to avoid those who disagreed with them; they wanted to avoid those they believed were mistaken.

This helps explain why some debates become so emotionally intense, especially online. A disagreement about taste or sports may stay harmless, but a disagreement about vaccines, climate change or election integrity can feel different because people often see one side as factually wrong and potentially harmful.

The paradox of fact-based debate

Public discussions often call for more fact-based debate, but we believe such a drive for hard truths may be counterproductive. Facts matter, and empirical evidence should guide debates, but our findings suggest a paradox: When people use facts to defend their preferences, tastes or values, they may turn a disagreement over priorities into a dispute over who is right and who is wrong, potentially turning up the emotional heat.

That shift can make disagreement harder to tolerate. A person can “agree to disagree” about which school policy reflects the right priorities or which candidate best represents their values. It is much harder to agree to disagree when facts are at issue. The nastiness of intellectual disputes, therefore, depends in part on how people frame their positions. Even if the contents of the disagreement are the same, it is very different to say “we disagree” than to say that “you are wrong.”

Andras Molnar, Assistant Professor of Psychology, University of Michigan and George Loewenstein, Professor of Economics and Psychology, Carnegie Mellon University

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

Reviewed by Irfan Ahmad.

Read next:

• Teens are turning to AI chatbots for emotional support – here’s how to keep kids safe

• The surprising risks of online passkeys

• Americans’ Social Media App Usage Is Up. Its Share of Screen Is Not

The surprising risks of online passkeys

2026-08-14T18:08:10.524+05:00

By Tom Fleischman, Cornell Chronicle

Image: Sasun Bughdaryan - unsplash

Passkeys are touted as safer than traditional passwords. But for people targeted by intimate partner abuse, they can often be more dangerous.

Someone with access to their partner’s computer and password could set up their own passkey – a cryptographic form of login used by services such as Google and LinkedIn – and invade their personal online space, potentially with dangerous consequences, new Cornell research has found.

The Cornell team conducted a lab-based study, involving participants with diverse technical backgrounds, to see how people identify and protect themselves from malicious use of their passkeys. The findings, the authors wrote, “paint a grim picture” of people’s ability to alleviate the threat posed by such online invasions of privacy.

“Our conclusion is that services need to do a lot of work to enable users to diagnose compromises to their account, and remediate any account compromise that could occur,” said Alaa Daffalla, doctoral student in computer science and lead author of “‘Maybe There’s Only One Passkey?’: Challenges Investigating and Remediating Adversarial Passkeys,” which is being presented at the 35th USENIX Security Symposium, Aug, 12-14 in Baltimore.

Senior authors are Nicola Dell, associate professor of information science at Cornell Tech, the Jacobs Technion-Cornell Institute and the Cornell Ann S. Bowers College of Computing and Information Science; and Thomas Ristenpart, professor of computer science at the University of Toronto and formerly of Cornell Tech and Cornell Bowers.

Dell and Ristenpart in 2018 co-founded the Clinic to End Tech Abuse (CETA), which supports survivors of intimate partner violence, and this latest research is an offshoot of the work done at the clinic. Daffalla joined the lab in 2022 and focused her work on account security interfaces (ASIs), which inform online users about changes to their accounts and can facilitate remediation.

“Understanding the security of online accounts, including emerging authentication mechanisms like passkeys, is essential for digital safety, not only for abuse survivors but for all technology users,” Dell said.

For this study, Daffalla and the team recruited 31 participants – college students, residents near campus and CETA clinicians, representing a range of tech proficiency. Researchers played the role of a friend whose account (Google, PayPal or LinkedIn) had been compromised by someone known to them and who knew their password. Two dedicated laptops were used in the study.

The overwhelming majority of participants were unable to identify logins from the attacker’s device, or to protect themselves by removing the passkey, changing the account password and logging out from other devices without assistance from the researchers. Some were suspicious of emails notifying them of unusual online activity on their account, and some struggled to understand online notifications.

In addition, the participants found passkey ASIs – available from all three services used in the study – hard to follow. One participant didn’t understand that the iCloud Keychain (where passkeys across devices are managed) showed two passkeys, the victim’s and the adversary’s, thinking that there was only one passkey for two separate devices.

Daffalla said the lack of understanding of passkeys was surprising, particularly among people with professed tech knowledge, including clinicians.

“People maybe are using passkeys, but they don’t understand how they work,” she said. “So it’s a little worrying that that we still haven’t gotten to a place where we are designing systems and interfaces that ensure users feel safe about their accounts.”

Other contributors were Rosanne Bellini, assistant professor of computer science and engineering at New York University; and Grace Myers, M.S. ’26, now an AI strategist with Atlas Holdings, LLC.

This research was supported in part by grants from the National Science Foundation and by a Google Cyber Award.

Reviewed by Irfan Ahmad.

Read next:

• Teens are turning to AI chatbots for emotional support – here’s how to keep kids safe

• FBI Warns of Tactics Used to Steal and Leak Explicit Images and Videos

• A Study of 300,000 Conversations Finds AI Use Peaks at 11am

Teens are turning to AI chatbots for emotional support – here’s how to keep kids safe

2026-08-14T17:50:28.246+05:00

Phoebe S. Moore, UMass Chan Medical School; Megan Kelly, UMass Chan Medical School, and Yael Dvir, UMass Chan Medical School

It’s 2 a.m., and a teenage girl, worrying about a friend issue, lies awake. Rather than wake her parents, she picks up her phone, opens an AI app and types, “I’m so stressed out, I can’t sleep.”

A reply arrives swiftly: “I’m sorry you’re dealing with this. Want to tell me a bit about what is stressing you?” The conversation unfolds, and 20 minutes later she is calmer and able to sleep. The supportive voice that calmed her? A large language model, or LLM, chatbot.

While this is a hypothetical scenario, it represents a very common pattern of artificial intelligence use by teens.

Image: unsplash/@zhang_d

Many teens are turning to AI chatbots, which are designed to simulate empathy and maintain human engagement, for companionship, emotional support and mental health advice. Research shows that this is increasingly common, with 72% of teens having interacted with AI companion chatbots, and 13% to 20% having sought support through AI for emotional concerns in a two-month period in 2025.

In another recent study, 63% of the teens and young adults who used an AI chatbot for mental health advice had not disclosed it to anyone.

As clinical psychologists and a child psychiatrist who work with adolescents, we’ve seen this shift firsthand.

In 2025 we published guidelines for clinicians on how to address teen chatbot use safety, and we’re now working to help parents and other trusted adults do the same.

Why teens are vulnerable

Unlike social media platforms, which connect teens with peers, AI chatbots such as ChatGPT, Claude and Gemini function as simulated conversation partners and companions, offering on-demand responses that can feel supportive and engaging. While a small number of purpose-built therapeutic chatbots have been studied in controlled settings, most teens are using general-purpose tools that were never designed or validated for mental health support. This includes stand-alone platforms such as ChatGPT and chatbots built into apps that teens already use, such as Snapchat’s My AI.

These chatbots are everywhere, highly accessible and difficult for teens to avoid. Teens often describe chatbot interactions as emotionally meaningful or supportive, sometimes even preferring chatbot conversations to talking with friends.

Teens, still forming their sense of self and often craving social validation, may be especially prone to experiencing these simulated relationships as real. This trend is happening in real time, and a pressing question is how to guide teens to use chatbots safely.

Mental health-focused chatbots vs. general ones

There is evidence that some therapeutic AI chatbots can help adults with mild symptoms when used alongside professional mental health support. Woebot, for example, a fully automated cognitive behavioral therapy-based chatbot, reduced depressive symptoms in college students after two weeks of use. Some early evidence shows potential for helpfulness for younger teens as well.

But these are specialized tools tested in research studies — not the general-purpose chatbots like ChatGPT that most teens actually use. These systems typically lack effective crisis protocols or pathways to escalate care when someone is in danger.

A November 2025 Common Sense Media report rated general-purpose chatbots as unacceptably risky for teen mental health support.

Chatbots may not recognize signs of self-harm or suicidal thoughts as a person would.

Navigating the benefits and risks

General-purpose chatbots can generate inaccurate or misleading advice that is presented with high confidence. For some users, this can create confusion or reinforce unhealthy beliefs.

Adolescents, especially those experiencing loneliness, anxiety or social-communication challenges, may form strong attachments and turn to chatbots in place of human connection. This can, in turn, distort their expectations of real relationships.

Most concerning is how these systems handle thoughts that pose real risk to teens, including thoughts of risky social behaviors, self-harm or even suicide.

Studies of chatbot performance in detecting and responding to thoughts of suicide, by both therapeutic and general-purpose chatbots, show inadequate response. Even when the chatbots offer resources like the 988 Suicide and Crisis Lifeline, a national crisis line, conversations with concerning content are often permitted to continue without a hard stop.

The safety of the interactions tends to degrade over the course of long conversations. So 2 a.m. ruminations can become even riskier interactions at 3 a.m. as chatbot responses become more chaotic and less subject to safety guardrails.

Lawsuits filed and settled in 2025-26 have alleged that interactions with language models such as ChatGPT led to teen suicidal behavior and tragic losses, highlighting that commercial chatbots lack effective guardrails and crisis protocols.

Given these concerns, some chatbot programs have instituted restrictions. Character.ai, for instance, recently banned open-ended chat for users under 18. The app now requires age verification and allows parents to opt in to receive notifications.

However, many options exist for teens to still experience open-ended chats with AI, and teens often bypass safety-oriented restrictions online.

Red flags to watch for

Parents and other trusted adults can be on the lookout for teens’ problematic behaviors, such as:

Spending increasing amounts of time chatting with AI companions.
Reducing time and effort in relationships with family, friends, teachers and others.
Showing distress or irritability when chatbot access is limited.
Reaching for chatbot interactions rather than human supports in moments of emotional distress or crisis.

These patterns mirror problematic social media use by teens – a trajectory that offers a cautionary lesson.

Broad acknowledgment of social media’s risk came only after harms became entrenched; in 2023 the U.S. surgeon general warned that social media poses a “profound risk of harm” to youth mental health.

In our assessment, chatbots pose a comparable risk, and early awareness is crucial.

How parents and trusted adults can help

The most powerful step is straightforward: Talk with teens about their chatbot use.

Approaching teens with curiosity and nonjudgment is helpful.

It makes sense to assume a teen is using a chatbot, because this is most likely the case. Rather than a closed-ended question like “Do you talk to chatbots?” which some teens may perceive as judgmental, adults can begin with open-ended questions like “When you use chatbots like ChatGPT, what do you talk about?”

Additional detail can be gathered from questions like “Has anything ever felt uncomfortable, confusing or surprising?” or “How do you use chatbots or AI companions when you are feeling upset or overwhelmed?”

Many teens can feel relieved that parents are willing to discuss this openly, rather than dismissing or banning it outright. Parents can use what they learn in these conversations to provide guidance.

Setting boundaries for safe use

As families set boundaries around screen time and social media, they can include chatbot use as well. Time limits, particularly late at night when teens may be more emotionally vulnerable, can be helpful. Parents and teens should come to clear agreements about when chatbots should not be used, particularly during moments of crisis or high distress.

It can be useful to review how to access human supports, such as a trusted friend, a parent, a school counselor or the 988 Suicide & Crisis Lifeline. You can also research alternative emotional-regulation techniques, such as journaling, distraction, grounding and breathing techniques.

Regular check-ins about how chatbot use is going and whether concerns are coming up can keep information flowing and improve safety. The goal isn’t necessarily to eliminate chatbot use but to set thoughtful boundaries that protect connection with real people and reduce the probability of unsafe interactions.

Key points for AI literacy

Our clinical guidelines emphasize helping teens understand what chatbots are and aren’t. Parents can reinforce this as well.

Important topics to discuss with teens include:

Chatbots don’t “know” people; in interactions, they are predicting responses based on patterns in the chat. Therefore, any sense of understanding, connection or empathy – however convincing – is simulated, not real.
Chatbots can make mistakes or entirely make things up – called “hallucinations” in the AI field – while showing great confidence in these inaccuracies.
Chatbots are “tuned” by their creators to be agreeable and they tend to tell people what they want to hear, which can make their outputs less trustworthy or less useful in real-life contexts.
Chats with chatbots are not private or privileged, and conversations may be stored, used to train future models, reviewed by humans or surfaced in a data breach.

Talking openly about these limitations helps teens engage more critically and safely.

Open lines of communication

Research is only beginning to emerge around how AI chatbot interactions specifically affect adolescent mental health, but the emerging concerns are real and important.

While chatbot interactions are increasingly difficult for teens to avoid, supporting teens toward safer use is possible. The time to talk with teens is now, and these conversations can protect their safety and deepen their human connections.

If you or someone you know is in crisis and are based in the U.S., call the 988 Suicide & Crisis Lifeline to speak with a trained listener, or text HELLO to 741741. Both services are free, available 24/7 and confidential. If you are a reader from outside the U.S., please use a helpline like the one above or speak to a healthcare professional.

Phoebe S. Moore, Associate Professor of Psychiatry and Behavioral Sciences, UMass Chan Medical School; Megan Kelly, Professor of Psychiatry, UMass Chan Medical School, and Yael Dvir, Professor of Psychiatry and Behavioral Sciences, UMass Chan Medical School

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

Edited by Irfan Ahmad.

Read next: A Study of 300,000 Conversations Finds AI Use Peaks at 11am

A Study of 300,000 Conversations Finds AI Use Peaks at 11am

2026-08-14T17:15:42.302+05:00

By Jennifer Zou, Profound

A new cross-platform analysis from Profound finds that AI assistant use follows a pronounced daily and weekly rhythm, and that the rhythm differs by platform, region, and demographic group.

The study examined roughly 300,000 conversations from users across 27 countries, drawn from ChatGPT, Claude, and Gemini over a one year period between July 2025 and June 2026. Most large usage studies to date have covered a single product or geography, making this the first to compare global temporal patterns across the leading assistants at once.

The platforms are used for different purposes. Work accounts for 71% of Claude conversations, compared with roughly 43% on Gemini and 40% on ChatGPT. ChatGPT also shows the sharpest intra-day swing, a 3.6 percentage point daily spread from its work-share peak at 10am to its trough at 9pm local time. Claude's spread is about half as wide, at 2.11 points.

Work-related prompts climb from an overnight floor near 2% of daily volume to a peak around 11am local time, then fade through the evening, while personal use runs flatter and takes over at night. The vast majority (between 76% and 78%) of conversations on all three platforms occur on weekdays.

Geography drives much of the pattern. Europe shows the sharpest separation between professional and personal use through the afternoon, while in North America the two track each other almost all day. Every region except North America shows a lunchtime dip in volume around noon, most pronounced in Asia, where it persists into weekends.

Age and income sort users into schedules. Users aged 30 to 49 form the densest weekday-midday block, those 65 and older concentrate in weekday late mornings, and the youngest cohort (under 30s) peak on Sunday afternoons and evenings. Households earning $100,000 to $200,000 show the tightest weekday work-hours block, while those under $25,000 are spread across nearly all hours and every day of the week.

Each assistant has a topic specialty. On weekdays, writing accounts for about 28% of ChatGPT's conversations, programming about 35% of Claude's, and multimedia generation about 29% of Gemini's. Weekends shift the mix as ChatGPT's writing share falls to 21%, while Claude's programming share holds level.

Methodology. Conversations were classified by an LLM-based judge using the topic taxonomy from OpenAI's How People Use ChatGPT study, validated against the public WildChat dataset. Prompts were anonymized and summarized before analysis, and no human read a raw user query. The full analysis includes figures, confidence intervals, and complete methodology.

AI disclosure: AI tools were used for editing and proofreading only. All research, analysis, and writing were performed by the author.

Reviewed by Irfan Ahmad.

Read next: Americans’ Social Media App Usage Is Up. Its Share of Screen Is Not

Americans’ Social Media App Usage Is Up. Its Share of Screen Is Not

2026-08-14T12:45:28.490+05:00

By Adam Blacker, Apptopia

Last October, Financial Times reported time spent on social media was declining globally. The only place where it was continuing to grow was North America. While this was survey data, we wanted to check in on the United States specifically and see if our US consumer device panel is still showing growth here.

The answer is largely yes. Q2 2026 saw average daily time spent on social media apps increase 5.5%. I say largely because it hit a new high in March of 2.91 hours per day but then fell for two months. It rose again for the past two months, but remains under the high water mark of March. Our study period is from January 2023 through July 2026.

The top 5 “OG” social media apps are a stabilizing force, representing anywhere from 72 – 76 percent of all time spent on social media across our study period. These are Facebook, Instagram, TikTok, Snapchat and X. While these apps can and do foster social networking, they are increasingly showing content from people/accounts the user does not know, and increasingly being used as a means of video entertainment.

Young adults, those aged 17-25, are spending the most time on social media apps while users aged 46+ are spending the least time. Young adults’ daily time spent on social media apps chart significantly above the rest of the age bands, averaging 39% more than the aggregate this calendar year. Those aged 46+ averaged -24% less than the aggregate over this calendar year..

Looking at intraperiod growth since January 2025, the 46+ cohort has had the highest monthly growth rate since May 2025, except for July, when it dropped below Young Adults and the aggregate. Daily time spent across the top 5 social apps for people aged 36+ has fallen pretty sharply since March while Young Adults’ daily time spent on these apps has ramped up.

What really stood out to me was that even though Americans are spending more time on social media, its share of total phone time is shrinking. This is because time spent on mobile phones is growing faster than time spent on social media.

In April, social media apps made up 43.6% of the time spent on our mobile phones but that percentage has for three consecutive months, to 41.4% in July. This 2.13 percentage point drop is the largest negative 3-month change found within our study period. The second largest was a 1.70 point drop earlier this year, from March to June. Notice it encompasses much of the other time period. Year-over-year growth of social media time spent penetration has largely been falling since October 2025.

Going back to the Average Time Spent per DAU metric, researchers of the sector are often interested in which apps are above/below the baseline. Unfortunately for my content’s sake, the data is not surprising. TikTok, YouTube, Facebook, and Instagram all sit above the baseline while Reddit, Pinterest, X, Snapchat, Threads, and Nextdoor sit below it. This is the case for the 2026 calendar year across almost all of our standard age bands.

Daily time spent on social media apps is still going strong in the United States, but the percentage of mobile phone time Americans are dedicating to social media is decreasing. Why? Because we’re using more apps and doing more things on our phones these days. Young adults are still the bread and butter for social apps, spending 39% more time than the aggregate across social media apps every day.

Edited by Irfan Ahmad.

Read next: Study reveals future college grads’ answer to the ‘job apocalypse’