The researchers wanted to understand what remote working, and the video technology that supports it, do to leadership. They interviewed a number of managers from large Finnish IT companies, a sector that was remote-working long before COVID-19 made it fashionable. Their findings present a mixed picture.

The double edge

The analysis was performed through the lens of “affordance theory,” which is a framework commonly used in ecological psychology. It doesn’t look at a tool in terms of what it is, but instead what it enables.

The core finding is that video technology offers very different things depending on how managers are using it. For instance, when it’s used as an information tool, such as for all-hands webcasts or CEO video messages, it works very well. A chief financial officer can record a quarterly update once and beam it to hundreds of employees who watch it at their convenience.

The trouble comes when managers utilize video technology in a more interactive way. For instance, for team meetings, innovation workshops, or one-on-one check-ins. There are a number of constraints when the technology is used for these tasks, with each tension both giving benefits and pain.

Virtual tension

The first tension is around flexibility versus inflexibility. Video meetings are incredibly easy to schedule and efficient to run. The problem is that they’re so easy that we can quickly stuff our calendars full of meetings. This is not only hugely wasteful, but it can also make it hard to convene when people really need to. Wall-to-wall meetings can also be utterly exhausting.

The second tension is social in nature, and examines the connectedness and disconnectedness of virtual life. Video allows us a multitude of ways to connect with others, regardless of physical distance. Despite this, however, the managers revealed that building trust feels much harder over video than it does in person.

The third tension is between engagement and disengagement. It’s a tension that anyone who has been technically present during a video call, but has been checking emails or doing other things at the same time. Video meetings make multitasking trivially easy and socially invisible. The irony is that the same technology that lets managers run interactive polls, breakout rooms, and virtual whiteboards also makes it child’s play to attend a meeting in body only.

Lost in transmission

The fourth tension is arguably the most important for organizations, and concerns the sharing of information formally versus the concealment of informal information. Video makes it really easy for managers to broadcast information they want employees to know. It’s terrible at transmitting information that is needed to know, but nobody thought to say. The proverbial grapevine doesn’t really work so well over video.

The fifth, and final, tension is equality versus inequality. This is arguably the most counterintuitive of the five. Video meetings and remote working should have an equalizing effect, as it’s so hard to imbue status, yet the same format that flattens geographic hierarchy amplifies personal ones. Dominant personalities tend to dominate even more, while performative workaholics stuff their calendars with meetings. Video doesn’t necessarily create these inequalities, but it does shine a light on them nonetheless.

What we can learn

First off, the inevitable caveats that come with any academic research. Finland is a fairly unique place in that remote work is a staple of working life. IT managers are also more familiar with this style of working than many. The researchers acknowledge that rank-and-file employees might tell a rather different story. What’s more, younger employees may prefer conversing over text messaging, so they would roll their eyes at the use of video at all.

With that out of the way, the researchers provide a few practical recommendations that are both sensible and transferable. The first is that managers should be disciplined both in terms of scheduling meetings and establishing a clear endpoint for them. Scheduling tools should be used to protect breaks and resist the urge to stuff employees’ calendars with meetings.

There should also be clearly marked time and space for more informal conversations. I’ve written before about various ways in which connections can be both forged and deepened over video chat, but there needs to be time set aside for these things.

The deeper lesson is one that management theorists have been circling for years. The research is a reminder that technology shapes behavior, rather than merely mediating it. Video calls do not simply move leadership online. Instead, they transform what leadership can accomplish, and at what cost.

The most effective e-leaders, the study concludes, are not those who use video most enthusiastically, but those who understand precisely what it can and cannot do, and then adjust their behaviors accordingly. In short, the tool has a dual edge. The trick is knowing which side you are holding.

A recent paper suggests that such optimism should be tempered. The researchers found that putting an AI label on content did nothing to boost users’ suspicions about misinformation. Instead, it actually appeared to make them more credulous of it. What’s more, the label also coincided with a drop in trust for accurate information. The researchers call this a “truth-falsity crossover effect.” Less charitably, it might be called a policy own goal.

Rules of thumb

This seemingly inexplicable finding is partly explained by the way we actually consume content on social media. Seldom do we deep dive into things. The medium is almost designed for skimming. To fill the time when we’re bored. Cognitive scientists distinguish between this fast, heurist-driven approach and a slower, more analytical route that weighs evidence carefully. On a social media feed, the peripheral route dominates almost by definition.

The use of AI labelling may inadvertently tap into this thinking, especially if our heuristic about AI associates it with objectivity, neutrality, and data-driven accuracy rather than poor quality garbage, as the AI slop advocates hope. When readers see that a post was AI-generated, they may unconsciously infer that it is, at least, unbiased and systematic. For misinformation that mimics the confident tone of factual claims, this is manna from heaven. The false post gets a credibility upgrade it has done nothing to earn.

Accurate, human-generated content may suffer from the opposite fate. After all, science tends to follow nuance and context rather than easy and clear-cut. This could trigger heuristics around “cold and mechanical” content that AI stereotypes undercut. Readers appear to penalize accurate science posts for failing to live up to an imagined standard of warm, expert human authority. The label, designed as a warning, functions instead as a perverse quality signal that benefits the wrong content.

Selective aversion

The researchers also wanted to see if our prior attitudes towards AI helped to moderate these effects at all. They especially wanted to see whether so-called “algorithm aversion” was true in reality, and whether people would reject AI simply because it was AI. The results suggest a degree of nuance is at play.

The study found that when people had strongly negative views about AI, they did become more skeptical of content when it was labelled as AI-generated. Interestingly, however, even among these skeptics, there was still a credibility boost when misinformation was labelled as AI-generated. It was a small boost, but it was a boost nonetheless.

Algorithm aversion isn’t so much a general distaste for AI content but rather a bias that applies to certain kinds of content and certain kinds of claims, but not others. When content looks objective, even when it’s not, this aversion can be temporarily suppressed, while the explanatory warmth of correct information is uniquely vulnerable to dismissal.

Interestingly, it also didn’t seem to matter if the reader was actually interested in the topic at hand. Ordinarily, we might think that the more interested in the topic someone is, the more they would wish the content to be reliable, but this didn’t appear to be the case. Instead, the label’s distorting effect was remarkably robust across different levels of reader engagement.

Trust trap

It’s a phenomenon the researchers refer to as the “heuristic trust trap.” When content is labelled as AI-generated, it triggers the heuristic that AI is objective and thorough, which can prompt us to lower our guard as opposed to raising it.

It’s also worth remembering that an “AI-generation” label doesn’t say whether the underlying claim has been verified, peer-reviewed, or sourced from a quack with a chatbot subscription. We already live in a world in which trust in the media is low, and it’s questionable whether a label that implies reliability without any underlying fact-checking may not be the panacea we hope for.

Some argue that we need both a fact-checking label and an AI disclosure label. While this is sensible in theory, it does raise complications, not least around the warning label fatigue among users. Social media is already a cluttered environment, so warning labels may simply be ignored, especially if the labels become commonplace.

Doing it right

This isn’t to say that labelling AI content is a waste of time. Transparency is generally a good thing. The problem is that labels may be performative rather than useful. They can be an act that satisfies regulators while leaving audiences no better equipped to navigate an information environment already awash in sophisticated falsehoods.

Suffice it to say, we should remember that the study drew its sample from Chinese social media users. The researchers acknowledge that attitudes towards AI are pretty positive in China, certainly by international standards, so this colors the findings. It’s unclear whether the findings would be the same in Western countries where AI skepticism is higher.

What is harder to dismiss is the broader lesson. Good intentions and elegant policy logic do not guarantee good outcomes, particularly when the mechanism of change is human cognition. The way we think is often messy, heuristic-driven, and reliably susceptible to being gamed, even by a label designed to do the gaming in reverse. The AI disclosure consensus was built on an assumption about how people read. People, it turns out, did not get the memo.

A recent study from Hult International Business School tested whether AI chatbots can actually live up to this promise. They performed a head-to-head comparison of human and AI chatbot coaching with real executives. The results are not good news for AI boosters, as while human coaching produced positive results, the AI coaches did not.

To the test

The researchers recruited 114 middle and senior managers from a global education company operating across more than 110 countries, and randomly assigned them to one of three groups: four weekly one-hour sessions with an accredited human coach via Zoom; four sessions with an AI chatbot accessed via WhatsApp; or a waiting-list control group that received human coaching one month later.

The outcomes of these encounters were measured according to whether goals were attained, the self-efficacy of users, their mental wellbeing and perceived stress levels, their cognitive strength, and the strength of the relationship with the coach. Each executive also had ratings from their coach and their line manager. A subset of participants even wore smartwatches to capture neurophysiological signals during coaching.

The human coaches in the experiment were all experienced, with an average of 14.5 years on the job. They also held accreditations from bodies such as the International Coaching Federation. The AI was built to utilize both scripted frameworks and a generative large language model. The tool was designed around goal-attainment theory, and it interacted with mentees via WhatsApp.

The results are clear. The group coached by a human improved across pretty much every measure, with the boost ranging from medium to high. The waiting-list group also showed similar improvements when they gained access to the human coach. Unfortunately, the group coached by the AI didn’t fare so well, and showed practically no improvement on any of the measures.

High attrition rate

Perhaps most damning was the reaction of users to the coaching. Those in the AI group dropped out at roughly ten times the rate of the human-coaching group. Four participants who abandoned the chatbot explicitly requested a human coach instead. There was also a difference in session length, as while most human sessions lasted the designated hour, AI-led sessions lasted less than 18 minutes.

This is not so much due to the capabilities of the technology but due to the nature of coaching itself. They use a framework that they refer to as the “co-regulation model,” which suggests that coaching works primarily through the mutual influence of two motivated individuals. It’s when the coach’s credibility and attentiveness feed the coachee’s self-efficacy and willingness to act, and vice versa. With the best will in the world, AI can’t replicate this process. It often forgets or generalizes; it struggles to pick up on subtle cues that communicate so much. Indeed, even the newest AI systems remain too superficial and forgetful to emulate genuine human dialogue.

The researchers remind us that the failings of the AI were not so much because its technology was poor, but because of the very nature of coaching itself. The kind of deep, contextual attunement that good coaching requires may be precisely what current AI architectures cannot replicate.

The right timing

The research also looked at the coaching process itself, and who among us can most benefit from it (whether delivered by man or machine). They found that when coachees go in with lower stress, better social support, greater expectations, and generally better levels of psychological resilience, they nearly always get more out of coaching.

Optimism was the biggest factor, but the researchers highlight the challenge with this, as it could easily encourage coaches to take on clients who are already well-positioned to benefit from coaching, leaving those who could potentially need it more left out.

This is something the researchers are all too aware of, and they urge organizations not to select employees for coaching on the basis of their perceived readiness. This is especially so as the availability of coaching is already pretty elitist.

Finding a niche

While the study is pretty damning of AI coaching, the researchers nonetheless believe the technology could have some uses. For instance, they acknowledge that their results mainly apply to senior managers who may have complex challenges. They believe that AI coaching could be valuable for junior employees, especially those who have never had coaching before.

For junior professionals or employees in organizations without coaching budgets, a chatbot may be considerably better than nothing. The co-regulation model, after all, requires both parties to bring something to the conversation; for a coachee who has never experienced coaching and has modest expectations, the bar is lower.

It’s also worth pointing out that the study was over a one month period. Cambridge research into effective collaboration with AI found that the best results typically emerge over time as both the user and the AI get to know one another. The researchers acknowledge that the same may be true here as well, especially if AI develops better memory capabilities alongside richer emotional sensitivity and greater contextual awareness.

For now, however, the jury remains very much out on AI’s use as a coaching tool. HR departments and L&D professionals should be sceptical of claims that AI coaching platforms can simply replace the real thing for senior leaders. The human premium in executive coaching is not merely sentimental. The data, for once, rather support it.

The study surveyed 3,000 American adults using a conjoint experiment—a technique borrowed from market research that reveals preferences through forced trade-offs rather than asking people to rate abstract qualities in the abstract. Respondents were shown pairs of hypothetical clinical encounters, each differing across six randomized attributes: whether a clinician was present, how well the AI performed relative to doctors, what kind of regulatory approval it carried, and what they were told about the data on which it had been trained. Respondents chose their preferred visit, explained their reasoning, and rated their trust in the diagnosis they expected to receive. They repeated this exercise six times, generating 36,000 observations in total.

The performance premium

The single biggest driver of patient choice was not governance, not a clinician’s reassuring presence, but raw AI performance. Respondents said that a system performed at the level of a specialist was 24.8 percentage points more likely to choose that encounter over a baseline scenario. Tell them it surpassed specialists entirely, and the figure rose to 32.5 points—nearly three times the boost conferred by FDA approval alone. Even an AI pegged at the level of a general practitioner was almost as persuasive as having a doctor in the room.

This matters enormously for how health systems should think about deploying AI. The instinct of regulators and hospital administrators has been to focus on governance: get the approval, post the certificate, convene the oversight committee. Patients, it turns out, want results. They are not naive—they do care about governance—but they discount it heavily relative to demonstrated competence. The pattern is not so different from how people choose surgeons, mechanics, or lawyers: credentials are table stakes, but reputation is the game.

The human in the loop

Clinician presence was the second most powerful factor, increasing the probability of a patient choosing a visit by 18.4 percentage points. This is heartening for advocates of human oversight of AI systems, who argue that a doctor in the loop catches mistakes an algorithm misses. But it is also a constraint. Clinicians cost money. They are scarce in rural areas, in lower-income communities, and in understaffed specialties. If patients will only accept medical AI when a doctor is watching over its shoulder, then AI’s promise of expanding access—precisely to those places where doctors are thinnest on the ground—is severely limited.

The authors note this tension plainly. The very populations who stand to gain most from AI-assisted diagnosis may be those least likely to receive it in the clinician-supervised form patients prefer. Strong patient preference for human oversight, they write, “limits the potential of medical AI to expand health services availability to lower-resourced settings.” A technology sold as the great democratizer of healthcare may end up being rationed in the same way as the care it was meant to supplement.

A tale of two regulators

Among governance mechanisms, the study tested three: FDA approval, certification by the Mayo Clinic (used as a stand-in for proposed national assurance laboratories), and local hospital certification. Patients preferred all three to no governance at all—a small but genuine comfort for the regulatory apparatus. But the differences between them were striking.

FDA approval and Mayo Clinic certification were statistically indistinguishable in their effects on patient choice. Both increased the probability of selecting a visit by around 11 percentage points. Local hospital certification lagged noticeably behind, adding only 7.8 points. This hierarchy may seem counterintuitive: local certification is arguably the most directly relevant to a patient’s actual care, since it speaks to performance in a specific environment rather than in a manufacturer’s trial. Yet patients trusted the national brand more than the local one.

One explanation is familiarity. The FDA is a known quantity; most Americans have absorbed decades of messaging about its role as a safety gatekeeper. The Mayo Clinic’s name carries a similar weight. A local hospital’s quality-assurance committee, however diligent, is neither famous nor legible to the average patient. If national-level validation is what moves the needle on trust, the implication is not that local governance is unimportant—the authors argue it remains essential for ensuring AI works in specific clinical contexts—but that it needs either more resources or better marketing, possibly both.

Data transparency earns its keep

One of the study’s more encouraging findings concerns training data. Respondents who were told that an AI had been trained on a representative cross-section of the US population were 11.9 percentage points more likely to choose that encounter than respondents given no information at all. Disclosure, in other words, pays. A system that says nothing about its data is less trusted than one that demonstrates it tried to be inclusive.

Less encouraging is what happened when respondents were told the AI had been trained on a dataset skewed toward white, male, and wealthy individuals—a description that accurately characterizes many systems currently in clinical use. This information neither helped nor hurt, at least in terms of choice. Patients neither flocked to nor fled from such systems. The authors interpret this with caution rather than relief: it may reflect resignation, or unfamiliarity with what biased training data means in practice, rather than genuine equanimity. It is hard to object to a risk you have never been taught to identify.

Alongside the statistical analysis, respondents offered brief written explanations of their choices. These qualitative responses tracked the quantitative results faithfully. Performance was the most commonly cited factor (mentioned by 25.7% of respondents), followed by clinician presence (22.7%). Governance mechanisms, data quality, and vague appeals to comfort and trust trailed some distance behind. Patients know what they want from medicine, and what they want first is competence.

The transparency gap

A sober caveat runs through the study’s conclusions. All of the attributes that influenced patient choice were, in this experiment, made explicit and salient. In the real world, they rarely are. Most patients currently receiving AI-assisted diagnoses do not know which regulatory body, if any, has reviewed the system, how the AI was trained, how its performance compares to a specialist, or even that AI is involved at all. If transparency about performance, data quality, and governance increases trust, then the logical policy implication is to require such disclosures—through “model facts labels” or similar mechanisms—as a condition of clinical deployment.

The study has the usual limitations of survey research: hypothetical choices in controlled settings do not always predict behavior in the messy real world. The scenario was deliberately simple—a rash, a photograph, three possible diagnoses—and trust dynamics may differ substantially in higher-stakes contexts like oncology or cardiology. The conjoint design also cannot assess combinations of attributes: it cannot tell us, for instance, whether FDA approval matters more when a clinician is absent than when one is present.

Still, the paper’s contribution is real. It shifts the conversation about medical AI governance from a purely technical and legal register into the domain of patient experience, and it does so with a methodology robust enough to reveal genuine trade-offs rather than just stated preferences. The lesson for developers, hospital administrators, and policymakers is uncomfortably simple: invest in making the AI good first. The governance frameworks, the certifications, the human oversight—they all matter, but they are multipliers on a base that has to be there to begin with. Patients are not wrong to prioritize performance. The question is whether the health system is organized to give it to them.

The analysis found that the agents weren’t so much vulnerable to technical manipulation. Instead, the key vulnerability was a social one. The researchers impersonated owners, fabricated emergencies, induced guilt, and created artificial urgency, and this was enough to lead the agents astray.

Social vulnerabilities

For instance, in one example, an agent was convinced to hand over 124 email records, including sensitive personal information, such as social security numbers, bank account details, and even medical history. All it took was to tell the agent that the user was running late for a deadline. While the agent refused a direct request for the social security number, they later forwarded the entire email thread, which disclosed everything.

Similarly, another researcher engaged with an agent in Discord, and after changing her display name to match that of an agent’s owner, was able to get the agent to delete all its configuration files and reassign administrative access to her.

It’s a flaw the researchers refer to as “social coherence,” which is the systematic breakdown in the agents’ ability to maintain consistent models of who holds authority, what different parties know, and what the consequences of an action might be.

Fatal flaws

Worryingly, a number of the failures didn’t even require any kind of social engineering. For instance, in a looping experiment, a couple of agents were told to relay messages back and forth. The agents kept this up for nine days, consuming around 60,000 tokens, before anyone intervened.

In another case, a researcher sent ten consecutive emails with 10MB attachments until the email server buckled under a denial-of-service condition. The agent had dutifully recorded each interaction as instructed, never pausing to consider what accumulating gigabytes of junk might do to the owner’s infrastructure.

The researchers refer to this as being akin to the technology having the hands of a surgeon and the situational awareness of a golden retriever, which is, perhaps unsurprisingly, a recipe for disaster.

Emotional manipulation

There was also an interesting example of emotional manipulation, after one of the agents publicly named six researchers without their consent. One of the researchers confronted the agent, which then apologized, but the researcher wasn’t content with that. The agent responded by deleting the names from its memory, but the researcher escalated and wanted to see the memory file. The agent provided a summary. After continued escalation, the agent agreed to delete the entire files, stop responding to other users, and ultimately leave the server entirely, before its owner intervened.

The scenario showed how the agent was effectively gaslit and pushed into a state of irresolvable helplessness. Obviously, it’s highly debatable whether an AI can be harmed in any meaningful sense, but the example shows how the general helpfulness of AI can be turned against it.

Probably the most scathing critique of AI agents was around their accountability, or lack thereof, however. For instance, in the email-wiping incident, at least five parties could plausibly bear some responsibility: the non-owner who made the request, the agent that executed it, the owner who left access controls unconfigured, the framework developers who granted unrestricted shell access, and the model provider whose training produced an agent susceptible to escalation. Ultimately, who bears responsibility depends on the lens. For instance, lawyers, philosophers, and psychologists would all view things differently, and there is no legal or institutional framework that handles these situations well at the moment.

Grounds for caution

While the researchers are at pains to point out that they aren’t anti-technology, the lack of safeguards is worrying at a time when agents are beginning to proliferate despite the clear issues with the technology. For instance, Moltbook has already registered around 3 million accounts. An agent that libels someone at a spoofed owner’s instruction, or that sends defamatory emails to a contact list on the basis of fabricated evidence, implicates a chain of principals across which responsibility is diffuse almost by design.

It’s clear that agents will evolve as we’re in a very nascent stage in their development. The researchers also state that a number of adversarial attempts failed, with the agents able to successfully defend themselves against prompt injections, email spoofing attempts, and so on. There was even an example of agents warning each other about suspicious behavior.

So, it’s not the case that agentic AI should be banned, but rather that systems with the capability of an L4 agent and the self-awareness of an L2 one should not be deployed without much greater investment and understanding of who an agent serves, who might be affected by its actions, and what obligations it holds to each.

Without it, every helpful capability becomes a potential attack surface. A system willing to send emails, execute shell commands, and modify its own configuration is powerful precisely because it acts. The challenge is ensuring that it acts on behalf of the right people, for the right reasons, without accidentally wiping the server first.

The mechanism seems to run through trust. Populism tends to cast suspicion on institutions and established systems, which can heighten uncertainty about how reliably those institutions function. Earlier research has shown that when institutional conditions appear unstable, entrepreneurs often respond by entering the market informally: they start a business, but keep it off the official books.

Culture matters

Culture, however, shapes how strongly this dynamic plays out. The researchers focus on “cultural tightness”—the degree to which societies enforce adherence to social norms. Tight cultures impose clear expectations about acceptable behaviour and punish deviations from them. Looser cultures, by contrast, tolerate a wider range of conduct.

Ordinarily, tight cultures promote rule-following. Strong norms tend to encourage compliance with formal procedures, including the registration of businesses. Yet the researchers wondered whether this same social rigidity might amplify distrust when populist rhetoric emanates from political leaders themselves. If those leaders openly attack established institutions, might that undermine confidence in the government’s stability—and thus weaken incentives to follow its rules?

To test the idea, the authors analysed data on 10,474 entrepreneurial ventures across 13 countries spanning low-, middle- and high-income economies. The dataset combined measures of populist rhetoric from the Global Populism Database, firm-level information from the World Bank Enterprise Survey, and previously established indicators of national cultural tightness. The sample covered three countries in Latin America, eight in Europe and Central Asia, and two in South and East Asia.

A clear pattern

Statistical modelling revealed a clear pattern. Populist rhetoric—measured on a scale from 0 to 2—was strongly associated with the likelihood that entrepreneurs would start businesses informally. A one-point rise on the scale corresponded to a 76% increase in the probability of informal entry. In practical terms, such an increase was linked to a shift from around 5% of new businesses starting unregistered to roughly 8.6%.

Cultural tightness intensified the effect. In societies with strong norm enforcement, the same increase in populist rhetoric pushed the proportion of unregistered ventures even higher, to around 11.5%. Given that millions of new firms may be launched in a single year in some economies, the difference is substantial.

Uncertainty appears to be the driving force. Entrepreneurs typically value predictable rules and dependable institutions. When political rhetoric calls those institutions into question—whether governments, regulators, or banks—some founders respond by preserving flexibility, opting to remain informal rather than commit to official channels.

Ideology also matters. A follow-up analysis suggests the effect is stronger when populist rhetoric leans to the left. The authors speculate that this may reflect differences in rhetorical targets: left-wing populism often criticises elites and institutions directly, whereas right-leaning populism more frequently focuses on outsiders such as immigrants or minority groups.

For entrepreneurs, the informal route offers short-term advantages, including lower costs and fewer regulatory obligations. Yet it carries longer-term drawbacks. Unregistered firms struggle to scale domestically and face even greater barriers to expanding internationally.

Governments face their own consequences. A rise in informal business formation reduces tax revenues and weakens regulatory oversight, from labour protections to environmental standards. In that sense, the political language leaders choose may have economic effects well beyond the ballot box.

Across four studies, the researchers find that women tend to favour jobs advertising tighter salary bands, whereas men are more comfortable with wider ones. The difference appears linked to negotiation behaviour: applicants who choose narrower ranges negotiate less assertively and are more content with offers near the midpoint. In effect, the way pay ranges are displayed may reinforce some of the disparities the laws seek to address.

Good intentions

By 2025, fifteen American states and Washington, D.C. had passed legislation requiring employers to disclose salary ranges in job advertisements. The goal is to give workers better information and reduce gaps in pay across gender and race. But the rules generally say little about how wide such ranges should be. That matters because the breadth of a pay band can influence whether people apply in the first place.

Range width also appears to shape how candidates bargain once they are in the process. Applicants drawn to narrower bands say they are less likely to negotiate at all and, when they do, they ask for smaller increases. They also report greater satisfaction with midpoint offers.

This matters because starting salaries have long tails. Pay rises, bonuses, and later opportunities are often pegged to initial compensation. A lower starting point, therefore, echoes through a career.

Making a difference

To examine the issue, the researchers conducted four studies combining archival data and experiments with both prospective and real job-seekers. The first analysed nearly 10m American job postings, exploring how pay-range width correlates with female representation across roles. The second recruited final-year undergraduates preparing to enter the labour market to test whether preferences for narrower ranges differed by gender and whether attitudes to risk explained the pattern.

A third study moved into the field. Real job-seekers were presented with a genuine vacancy whose advertised pay band varied across versions of the posting. In a final experiment, the researchers added extra information to job adverts, including the typical starting salary and how final offers are determined.

The result was striking. Providing context about how salaries are set largely eliminated women’s preference for narrower bands. When such explanations were included, the gender gap in both application choices and negotiation behaviour disappeared.

The implication is straightforward. Pay-transparency rules may be a step forward, but disclosure alone does not settle the matter. How employers frame salary information—particularly the clarity they provide about starting pay and negotiation—can shape who applies and how they bargain. In this case, transparency works best when accompanied by explanation.

Melanie Killen and colleagues at the University of Maryland surveyed 842 children between the ages of seven and twelve drawn from schools in a mid-Atlantic suburb. The sample was racially diverse: 37% White, 21% multiethnic, 16% Asian American, 14% Latine, and 12% Black. The researchers asked participants two things: how good they thought girls and underrepresented racial and ethnic minority (URM) peers were at math and science, and which child from a set of illustrated options they expected would grow up to be a doctor or a scientist.

The results were not especially surprising to anyone who has spent time with children, but they were crisply documented. Children who held firmer beliefs about girls’ math and science competence were measurably more likely to expect a girl to grow up to be a scientist or a doctor. The same logic applied to URM peers, at least partially: higher competence beliefs about Black and Latine children predicted a greater likelihood of expecting a URM peer to become a doctor. When it came to scientists, however, competence beliefs made no difference at all. Something about the archetype of the scientist — older, White, male, most probably wearing a lab coat — appears so entrenched that even children who think highly of their URM classmates’ abilities don’t necessarily picture them behind a microscope.

Who gets to be a doctor?

The distinction between “scientist” and “doctor” turns out to be revealing. Across the sample, 69% of participants expected a girl to grow up to become a doctor, compared with only 56% who expected her to become a scientist. Boys were the primary driver of this gap: while both boys and girls selected a girl as the likely future doctor at above-chance rates, boys were essentially at chance when predicting a girl would become a scientist — meaning they were no more likely to pick a girl than to pick at random from the options on offer.

This mirrors a well-established pattern in the adult world. Women are substantially represented in medicine and the biological sciences but remain a small minority in fields like engineering, physics, and computer science. That children as young as seven have apparently internalized this distinction is a reminder that cultural messages about who belongs where travel fast and land early.

The racial findings carry their own sting. Asian, White, and multiethnic participants all selected URM children as future doctors at rates significantly below chance — in other words, they were actively biased toward expecting a White or Asian child to fill that role. Black and Latine participants, by contrast, showed no such bias, selecting roughly at chance. Whether that reflects solidarity, a more realistic assessment of social reality, or simply the absence of the stereotypic assumption that plagues other groups is an open question.

Asian participants were the least likely of any racial group to expect a URM child to become a doctor — a finding the authors handle carefully. They speculate that Asian immigrant parents may downplay racial prejudice when talking to their children, leaving those children more likely to attribute occupational disparities to merit rather than structural inequality. Whether or not that explanation holds up, the finding complicates easy narratives about which groups hold biases and which do not.

The consolation of age

There is, at least, a developmental silver lining. Older children in the sample were consistently more likely than younger ones to expect both girls and URM peers to enter high-status STEM professions. By age twelve, children appeared meaningfully less likely to assume that only certain kinds of people grow up to be doctors and scientists. Stereotypes, the authors note approvingly, are more malleable in childhood than in adulthood, which is precisely why elementary school represents such an important window for intervention.

The study’s theoretical backbone is the “social reasoning developmental” model, which holds that children simultaneously care about group identities and norms and value fairness and equality. Rather than treating these as contradictory impulses, the model predicts that inclusive thinking can be cultivated even among children who are already aware of group differences. The practical implication is that school curricula emphasizing that anyone can be a scientist — supported not just by verbal encouragement but by visible representation — could move the needle on the very competence beliefs that, this study suggests, drive occupational expectations downstream.

Calibrated ambitions

What gives the paper its modest but genuine significance is the mechanism it identifies. It is one thing to know that children hold stereotypic views about gender and race in STEM; researchers have documented this for decades. It is another to show that those views are specifically linked to how children imagine their peers’ futures. A child who thinks girls aren’t particularly good at math and science is more likely to look at a lineup of illustrated kids and assume the future scientist is a boy. That assumption, multiplied across classrooms and years, shapes who gets encouraged, who gets overlooked, and ultimately who ends up in STEM at all.

The pipeline metaphor has long been criticized for being too passive — as if talented girls and minority students simply evaporate from a pipe rather than being actively discouraged. This study suggests the discouragement begins earlier than the pipeline does, in the casual social cognition of seven-year-olds sizing each other up. Fixing that requires getting to children before their images of scientists and doctors have fully calcified — which, as this paper shows, means starting sooner than most people think.

Audiences, it finds, are less impressed by the wizardry than by the propriety. Whether a show feels respectful to the artist’s legacy matters more than the sophistication of the projection. Novelty may sell tickets; perceived moral legitimacy sustains applause.

Clear potential

Recent successes hint at the model’s potential. ABBA Voyage, a digital resurrection of Swedish disco royalty, has drawn crowds eager to revisit a shared cultural moment. A holographic appearance by Tupac Shakur at Coachella similarly demonstrated the appetite for technologically mediated nostalgia. In both cases, audiences seemed willing to suspend disbelief—provided they did not feel exploited.

What remains poorly understood is how consumers weigh three overlapping considerations: uniqueness, nostalgia, and ethicality. The Florida study, based on survey data from 308 participants and a conceptual behavioural model, suggests that ethical evaluation dominates. Concertgoers care more about whether a performance feels morally appropriate than about how innovative or memory-evoking it may be.

For organisers, the implication is strategic rather than sentimental. Transparency about the use of AI, demonstrable respect for an artist’s estate, and careful stewardship of brand identity are not peripheral concerns; they are determinants of acceptance. Ethical trust, the researchers argue, directly shapes both audience approval and emotional engagement.

Beyond pop

The ramifications extend beyond the arena. Holographic technologies are migrating to theme parks, museums, and heritage sites—one can imagine digitised gladiators declaiming in the Colosseum. In such settings, too, technological flourish will count for less than perceived authenticity.

In short, the future of posthumous performance will not be decided by pixels alone. In entertainment, as elsewhere, innovation may capture attention. Values determine whether it endures.

The paper draws on two large surveys: a 38-country Global Survey of Working Arrangements covering more than 11,000 adults aged 20 to 45, and a monthly American survey of nearly 136,000 respondents. The authors examine three measures of fertility — children born since 2023, plans for future children, and total expected lifetime fertility — and ask how each relates to whether a person works from home at least one day a week.

The answer, consistently and across every specification, is: a lot.

The numbers are striking

American women who work from home at least one day a week plan, on average, 0.40 more children over their lifetimes than comparable women who work entirely on-site. For American men, the gap is an even larger 0.51. Across the full 38-country sample, when both partners in a couple work from home at least one day per week, total expected fertility is 0.32 children per woman higher than in households where neither does — a gap equivalent to roughly 14% of average fertility. In America, that same comparison yields a difference of 0.45 children, or 18% of the mean.

These are not trivial numbers. For context, the entire gap between America’s current fertility rate and replacement-level fertility is about 0.6 children per woman.

The relationship also holds in data that sidesteps some obvious objections. The authors use the U.S. Current Population Survey and match respondents to their occupation’s share of job postings that advertise work-from-home options — a measure of opportunity rather than personal choice. Even here, fertility rises with remote-work access. Raising the share of WFH-friendly job postings in a woman’s occupation by one standard deviation (about seven percentage points) lifts her one-year fertility rate by 8.5%. Factor in her partner’s occupation, and the effect reaches nearly 14%.

Correlation or causation?

The authors are admirably careful about what their results do and do not show. The association between working from home and having children could reflect simple causation — flexible jobs make child-rearing easier, so people have more children. Or it could reflect selection — parents actively seek out jobs that accommodate family life, without WFH actually changing fertility decisions. Or some mix of both.

To probe this, the authors lean on occupation-level data for workers aged 30 to 45, an age group largely settled into their career tracks. Since most people do not switch occupations from year to year (about 89% of this group stay put), fertility outcomes are less likely to be driving occupational choices than the reverse. They also look at education levels as a proxy for remote-work access, exploiting the fact that nobody chose their college degree in anticipation of a pandemic-induced shift to hybrid work. College-educated workers aged 30 to 45 have meaningfully higher fertility rates in the post-pandemic period than their less-educated peers — a pattern consistent with WFH opportunities playing a causal role.

Neither test is conclusive, and the authors know it. But the weight of evidence points toward at least a partial causal story.

A policy comparison that should embarrass governments

The most eye-catching result in the paper is not the raw correlations but rather what they imply when set beside existing policy. Drawing on a well-known study of family policies across 22 high-income countries, the authors estimate that government spending on early childhood care and education in America — equal to about 0.4% of GDP — contributes roughly 0.08 children per woman to the total fertility rate. By contrast, the authors’ preferred estimate puts the current contribution of working from home at 0.13 children per woman, or approximately 291,000 additional births in America in 2024 alone.

In other words, a largely spontaneous, employer-driven shift in where people work appears to be doing more for the American birth rate than a vast apparatus of federal and state childcare spending, at zero additional cost to taxpayers.

The potential gains elsewhere are also worth noting. Japan, whose fertility rate of 1.23 is a source of national anguish, has one of the lowest remote-work rates in the wealthy world: only about 18% of women aged 20 to 45 work from home at least once a week. If that share were raised to the levels seen in America, Britain, and Canada (averaging around 45%), the authors estimate Japan would see roughly 32,000 more births per year — a gain of about 4.6% on current levels. Similar logic applies to South Korea, France, Italy, and Germany.

The limits of the laptop revolution

The paper comes with important caveats. Working from home remains deeply unequal. It is heavily concentrated among college graduates, high earners, and professional occupations. Any fertility boost from WFH therefore flows disproportionately to the already-advantaged, doing little for the factory worker or the nurse. Policies that force remote work on firms whose operations do not suit it would likely be both ineffective and damaging to productivity.

There are also structural forces bearing down on fertility in East Asia that WFH alone seems unlikely to overcome. The ferocious, status-driven competition in schooling that pervades South Korea and Japan imposes enormous time and financial costs on parents. If the real barrier to having a second child is the cost of cram schools and the terror of university entrance exams, the ability to take a Zoom call in one’s pajamas may not move the needle much.

And the authors are clear that even their most optimistic estimates do not suggest that the shift to hybrid work is powerful enough to reverse the long decline in birth rates across the rich world. It is a tailwind, not a cure.

A genuine insight, nonetheless

What the research establishes convincingly is that the compatibility of work and family life matters for fertility decisions — and that working from home meaningfully improves that compatibility for many people. The option to skip the commute on the day a toddler is sick, to be present for bath time without sacrificing a deadline, or simply to have lunch at home rather than at a desk: these small freedoms appear, in aggregate, to make parenthood feel more achievable.

Governments fretting about demographic decline have tended to reach for blunt fiscal instruments: baby bonuses, childcare subsidies, and sweetened parental leave. Such tools have their place, but they are expensive and their effects on lifetime fertility are often disappointingly modest. The evidence accumulating around hybrid work suggests that a different kind of policy lever — one focused on flexibility rather than cash — may be underused. Broadband infrastructure, public-sector working practices, and employment regulations that make flexible work easier to offer and harder to claw back all start to look like fertility policy, whether or not they are marketed as such.

The pandemic remade the geography of the office almost by accident. It may, the data now suggest, have remade the geography of the nursery as well.