A new study from Vanderbilt University examined when, exactly, children begin to make this inferential leap. The findings, drawn from a sample of 115 Americans aged 4 to 10, along with a cohort of adults, suggest that the ability to detect social bias in others undergoes a striking transformation somewhere around the age of seven—with meaningful implications for how societies might approach the teaching of fairness and equity.

Clever setup

To avoid either reinforcing existing prejudices or letting participants rely on prior knowledge, the researchers introduced children to two entirely fictitious social groups—the Zarpies, who favoured yellow clothing and green triangles, and the Flurps, partial to purple and red pearls. Participants then observed a protagonist named Ann, whose own group affiliation was deliberately left ambiguous, behave consistently kindly toward one group while ignoring, dismissing, or rudely rebuffing members of the other. After each negative interaction, participants were asked simply: why did Ann do that?

The design was methodologically tidy. Novel groups meant no baggage; the question was broad enough to capture spontaneous reasoning rather than prompted conclusions; and by asking after each of six interactions, the researchers could pinpoint the precise moment at which any given participant first cottoned on to the pattern.

Neat sorting

 Among younger children—those between roughly four and six years old—fewer than half ever identified Ann as biased against the targeted group, even across six repeated demonstrations of differential treatment. When asked directly how Ann feels about Flurps at the end of the study, these children responded as though the question had no obvious answer. Many instead explained Ann’s behaviour with situational reasons: she was busy; the restaurant was closing; she was in a rush. Their explanations were coherent but, in a sense, wilfully incurious about the pattern beneath them.

Older children and adults behaved strikingly differently. Among those aged seven and above, 90% eventually concluded that Ann was biased, and every single adult did so. More tellingly, most of those who reached this conclusion did so quickly—56% after just one interaction, and nearly 90% within three. Once the threshold for bias detection had been crossed, the evidence required to cross it was minimal.

Two possibilities

What causes this? The authors consider two possibilities. The first is conceptual: children below school age may simply lack the framework to understand that a person can harbour a systematic prejudice against a category of people, as opposed to merely disliking a particular individual or being temporarily out of sorts. The second is attentional: younger children may not track statistical correspondence between an actor’s behaviour and the group membership of those on the receiving end of it, even when the pattern is glaring to adult eyes.

Crucially, the youngest children who never detected a bias were not simply tongue-tied or shy. A full 41% of them offered situational explanations rather than pleading ignorance—they were reasoning, just not in the direction of group-based inference. What they lacked was not fluency but, apparently, the conceptual toolkit to consider group-based animus as a plausible explanation for another person’s conduct.

There was a further nuance worth noting. Among those young children who did eventually detect a bias, they did so almost as quickly as older children and adults. This suggests the younger cohort splits into two distinct sub-groups: those who have not yet developed the concept of group-based bias, and those who have—the latter matching their elders in the speed of detection once the cognitive machinery is in place.

Sentiment matters

Observing Ann’s behaviour made participants, across all ages, like her considerably less and feel less inclined to befriend her. This decline was steepest among adults and older children, who were also most likely to explicitly cite her treatment of the targeted group as their reason for the cold shoulder. Notably, however, participants’ opinions of the Flurps or Zarpies themselves did not change. This stands in mild contrast to earlier work suggesting that watching a group be mistreated leads observers to think worse of them—as though the mistreatment itself implies something unflattering about its recipients. Here, the opposite dynamic seemed to prevail: blaming Ann may have left less room for blaming the Flurps.

There is a tidy implication lurking here. If identifying a biased actor spares the targeted group from the observer’s judgment, then teaching children to recognise bias—rather than leaving them to absorb unexamined patterns from their environment—may carry protective benefits for minority groups, not just corrective ones for perpetrators.

The paper is careful about its limits. The sample, recruited from Nashville, was predominantly White and upper-middle-class, and the upper age boundary was ten. Adolescence, when personal experiences of discrimination tend to spike—particularly for Black and Latino youth—remains uncharted in this framework. Nor does the study address whether children who have themselves experienced identity-based discrimination prove quicker to detect it in others. The researchers sketch two competing hypotheses here: “group-specific bias detection,” in which personal experience sharpens sensitivity only to biases against one’s own group; and “generalised bias detection,” in which it heightens vigilance across the board. These remain open questions for future research to unpick.

The broader lesson, however, is already legible. Identifying prejudice is a skill, and like most skills it develops. In an ideal world, adults would recognise this and scaffold the process—naming bias when it appears, explaining why it matters, and resisting the temptation to assume that children will figure it out on their own. The seven-year-olds in this study were already doing so, with commendable efficiency. The four-year-olds, for now, were still looking for another explanation.

That, at least, is the striking conclusion of a paper from DePaul University in Chicago. Their insight is deceptively simple: rather than lumping all college-educated immigrants together, the authors distinguish between those whose skills overlap with native workers — “competing” immigrants — and those whose skills do not — “non-competing” ones. The tool they use to draw this line is the college major. An immigrant computer science graduate, they reason, competes directly with a native computer science graduate. But that same immigrant likely complements a native education major, whose skills lie in an entirely different occupational territory.

To the test

The authors test this framework using eight years of American Community Survey data covering college graduates in the 40 largest metropolitan areas. They group the 171 majors the survey tracks into seven broad categories — business, engineering, computer science, science, health, education, and liberal arts — and measure how the occupational employment of native and immigrant graduates in each category overlaps. An “index of congruence,” borrowed from an older labour economics toolkit, quantifies the similarity. The results are clear: natives and immigrants with the same major overwhelmingly work in the same types of jobs, while those with different majors largely work in different ones.

Armed with these definitions, the authors then run the numbers. Inflows of competing immigrants reduce both full-time employment and wages for natives in the same field. The earnings elasticity is a sobering ?1.04: a ten-percentage-point increase in the competing immigrant share reduces native earnings by around ten percent. Yet inflows of non-competing immigrants push in the opposite direction, raising employment and wages for those in different fields. The good news for most American graduates is that the positive forces have dominated during the study period. Between 2012 and 2019, college-educated immigrants were disproportionately engineers and computer scientists — categories that are largely non-competing with the far larger pools of business, education, and liberal arts graduates who make up the native workforce.

This creates an important asymmetry in who gains and who loses. Native education majors are particularly fortunate: they face almost no competing inflows but benefit handsomely from the flood of STEM immigrants who need accountants, teachers, doctors, and lawyers. Native computer science majors sit at the other end of the spectrum, facing heavy competition. The study finds that computer science graduates in tech-heavy cities like San Jose experienced meaningful declines in full-time employment as a result.

Crossing the divide

The paper also sheds light on a methodological dispute that has long divided economists. Previous work using occupational employment to measure competition tends to find muddier, less significant results. The authors show why: immigrants often take jobs below their skill level on arrival, and natives sometimes move to different occupations in response to immigrant inflows — both of which corrupt an occupation-based analysis. College major, fixed before anyone entered the labour market, is immune to these distortions.

Younger workers bear the brunt of the costs and reap the largest gains. The effects on those aged 25–40 are roughly three times as large — in both directions — as those on workers aged 41–60, suggesting that immigrants and recent graduates are closer substitutes and complements than immigrants and seasoned professionals. The authors also find that the non-competing benefits flow primarily through industries where immigrants are themselves most heavily employed — professional services, information technology, healthcare, and finance — rather than through broader local consumer spending. This points to production complementarities, not a simple “more people buying more things” story, as the dominant mechanism.

The policy implications are pointed. Immigration debates tend to revolve around numbers: how many visas to issue, how high to set the cap. This paper suggests that is the wrong question. What matters is the match between the skills of arriving immigrants and those of the existing workforce in the cities where they settle. A city already saturated with software engineers gains little and loses something when more software engineers arrive; the same city may gain considerably when nurses, educators, or scientists show up instead. Designing immigration policy at the level of national quotas, without attention to local skill compositions, is a blunt instrument for a precise problem. The brain gain from high-skill immigration is real — but it is unevenly distributed, and whether any given native worker ends up on the winning or losing side depends less on politics than on what they chose to study two decades earlier.

Research from Tel Aviv University explores whether that’s actually the case. The researchers half suspected that what passes for informed public opinion on AI governance might be something rather shakier, and consist of a cocktail of vague techno-anxiety, general predispositions about institutions, and whatever happened to cross someone’s screen that morning.

Shaping decisions

The researchers worked with a random sample of over 1,500 American workers to understand how their exposure to AI in either a practical or informational sense changed their attitude towards its use in public life.

Volunteers were offered tasks and told that either a human or an algorithm would decide which task they were assigned. The tasks were designed to invoke a positive or negative reaction to AI. Each person was then sent an unrelated survey a few days later from a different source to ask them their views on using predictive AI on issues such as parole decisions or the provision of food stamps.

The tasks were cleverly designed, with the lower-paid one invoking a sense of grievance, and the higher-paid task a sense of satisfaction. Crucially, whether either outcome was attributed to an algorithm or a person was entirely random. The volunteers knew who had assigned them each task, however, and this was a common source of discontent, with almost half of those assigned to the algorithmic condition specifically mentioning the algorithm in complaints they were invited to write.

Making a change

The algorithm did appear to change things, albeit not the kind of things that matter most for democratic governance. For instance, those who were managed by AI were significantly less happy with their work, which resulted in putting in less time and effort, and also getting more tasks wrong. This is consistent with a growing body of research suggesting that algorithmic management breeds a peculiar form of alienation. What’s interesting, however, is that this didn’t change people’s opinion on whether AI should be used in government. This trend was consistent across the board.

People’s exposure to information about AI was more influential, however. Workers who were exposed to expert assessments of AI’s societal implications did update their views, and the effect persisted for days after the task was completed. For instance, when people read about AI’s potential for consistent, unbiased decision-making, they became more supportive of AI’s involvement in government. However, when they read about racial bias and accountability, they became far less enthusiastic. Interestingly, the positive information moved opinions more than the negative information.

Crucially, the updating was not merely a case of confirmation bias at work. Workers who were initially sceptical of AI, and then read positive accounts of it, shifted in the direction of that positive information, not away from it. The kind of reasoning that has made scientific communication on climate change and vaccines so frustrating was largely absent here. The researchers believe this might be because AI has largely been an unregulated environment so far.

Not all things are equal

Interestingly, there were certain policy domains whereby people were much more open to changing their minds. For instance, the volunteers were most open to changing their minds about AI’s role in distributing resources, such as welfare or the provision of public goods. They were generally far less likely to change their minds about AI being used in areas involving the punishment of individuals, such as parole decisions or restraining orders. The researchers believe that this is likely to be because AI is already unpopular in these domains.

What does this mean for calls to have more public input into the regulation of AI? In many ways, it’s quite reassuring. The fear that public opinion on AI is merely an expression of general techno-scepticism or tribal instinct seems overblown, at least for now. People do appear to be somewhat able to learn and update their opinions.

There are also some more worrying implications, of course. I’ve written before about research showing how exposure to decisions makes us more likely to oppose AI’s involvement. For instance, if we’re a recipient of welfare, we are less likely to want AI involved in welfare-related decisions. This study suggests personal exposure to AI has a limited impact on wider policy beliefs, but does impact our opinion on AI in our particular domain.

Millions of people now interact with AI in consequential settings. If those interactions do not, on their own, translate into political engagement with the governance of AI, then the feedback mechanism that democracies normally rely on is misfiring. The gap between the managed and the managers may widen without anyone quite noticing.

There remains a window of persuadability, however. AI regulation has not yet calcified into a culture-war issue, unlike climate change or vaccines before it. History suggests, however, that this window won’t remain open indefinitely. Once the issue becomes politicized, as we’re already seeing with tech companies lobbying for light-touch regulation, citizens may stop weighing evidence and start choosing sides. This study shows us that democratic deliberation has a clock running against it.

Leaders’ emotions, it finds, rarely remain contained. They spill over, indirectly shaping employees’ well-being, attitudes, and conduct. Yet the same behaviours are not interpreted symmetrically. When strain manifests as irritability or withdrawal, male leaders are granted a wider margin of tolerance; female leaders, by contrast, are held to a more exacting standard of emotional attentiveness.

Different values

The asymmetry extends to what employees value. Female staff tend to place greater weight on relational and emotional support from female leaders, while male employees prioritise task orientation and goal alignment from male ones. Across the board, however, participants proved more forgiving of male leaders who offered limited emotional support—an indulgence less readily afforded to their female counterparts.

Such double standards have tangible consequences. Negative emotional displays correlate with diminished trust, disengagement and higher turnover intentions. Positive behaviours, unsurprisingly, produce the opposite: stronger motivation, engagement and organisational commitment. Leaders who demonstrate emotional acuity—irrespective of gender—are more adept at cultivating supportive environments.

There is, moreover, a feedback loop at work. Employees who consistently receive support are likelier to extend empathy when leaders themselves show signs of strain. Regular check-ins and visible concern for staff well-being can therefore shape not only morale but also how leadership is perceived.

The prescription is straightforward, if not always easy to execute. Leaders must attend to their own psychological maintenance as carefully as to their teams’, investing in recovery and stress management. Organisations, for their part, would do well to train leaders to recognise and regulate their emotional states. Without such efforts, gendered expectations will continue to distort both leadership behaviour and the cultures that emerge from it.

The work was often dull and menial, but that was arguably the whole point. Apprentices weren’t just earning a living, they were paying via their labor for the expertise being quietly dripped into them. Now, this bargain is being disrupted by AI technology that can do much of that menial work in a fraction of the time and for a fraction of the cost.

A new world

A recent paper from the London School of Economics and Northwestern’s Kellogg School looks at how this new world might affect anyone in a knowledge-based industry where there’s an inherent interest in ensuring a healthy pipeline of new talent coming through.

The researchers illustrate what many have already feared. They show that when AI does the menial tasks, young employees have nothing left to trade for their education.

The challenge harks back to the work of Gary Becker in the 1960s. When the main benefit of workers is what’s in their heads, these skills are usually both general and portable. When employers invest in those skills, it’s very difficult to lock them into the firm. So the training must be paid for upfront, not in cash (juniors rarely have any) but in toil. The novice accepts years of below-market wages in exchange for a steady transfer of knowledge.

It’s a bargain that has worked for generations, but it’s one that AI is destabilising. When generative AI can do many of the things that junior employees could previously do, the “currency” of apprenticeships gets minted by the technology instead. We’ve already seen entry-level hiring drop, especially in occupations most exposed to AI, in a clear sign that the pipeline has sprung a leak at the bottom.

Leveraging expertise

The researchers refer to this as the “expertise leverage ratio,” and it refers to the ratio of what a fully trained graduate can produce in an AI-augmented world to what AI can produce on its own at an entry-level standard. The numerator in this equation is the value of a seasoned veteran, with the denominator the value AI brings without any human support. The ratio, therefore, tells you much saleable knowledge remains between those two poles.

They believe that apprentices will still have value so long as R is at or above Euler’s number, e, approximately 2.718. When AI automates a lot of the entry-level tasks, it essentially raises the floor and erodes R. When AI expands the capabilities of seasoned veterans, it raises the ceiling and expands it.

Whether apprenticeships survive is simply a race between these two forces, and R keeps score. In a world where AI makes a seasoned consultant five times more productive but can also handle the tasks a first-year analyst once did, the question is whether the fivefold gain outpaces whatever the floor has become.

Unsettling news

It’s an unsettling prospect for many professions. For instance, in a scenario where senior work remains largely untouched but entry-level work gets automated, the ratio falls considerably. Juniors might gain briefly, but the floor rises too quickly, and the ratio falls below e, making the whole edifice crumble. Even when the boost to senior people is mild, the same phenomenon largely occurs, as the top doesn’t grow quickly enough to keep pace with a rising floor.

The researchers suggest two approaches that could change things for the better. Firstly, firms could engineer the top prize upward by rebranding their services to emphasise human judgment, accountability, and the kind of trust that clients will pay a premium for.

The second option is a redefinition of just what apprentices do. This would involve pushing them toward tasks involving personal contact or contextual judgment that AI still handles poorly, thereby keeping their contribution above the AI floor even at the outset.

Neither is without costs or challenges. If firms have to rebuild their training systems around client-facing and judgment-heavy work, they will probably have to redo job descriptions entirely. What’s more, requiring a premium for “human certified” work supposes that clients exist to pay such a premium.

The research clearly identifies the circular nature of the risk. If apprenticeships collapse because AI does the entry work, the pipeline producing future experts dries up. A world of powerful AI tools then arrives, populated by too few people who know how to use them wisely. The masters who might have trained the next generation never found the bargain worth striking. Euler’s number, of all things, may turn out to be the dividing line between a future with human expertise and one without it.

It turns out that ambition is harder to realise than it sounds.

A new paper by Nilanjana Dutt of Bocconi University and Sarah Kaplan of the University of Toronto, published in the Strategic Management Journal, offers a sobering assessment of what women-focused accelerator programmes actually deliver. Drawing on data from 1,417 ventures that applied to 33 social-innovation accelerators across 65 countries between 2013 and 2015, they find that the headline result — acceleration improves venture performance — masks a more complicated story underneath.

On average, ventures that were accepted and participated in programmes reported substantially higher revenues a year later compared with those that were rejected: participation was associated with roughly a 112% boost in revenue change. But when the authors dug into the data by gender, the picture fractured. Women-led ventures did not appear to benefit to the same degree as their male-run counterparts. The coefficient capturing the additional benefit for women was negative, though imprecisely estimated, suggesting wide variation in the sample rather than a clean story.

Where you stand depends on where you sit

What drove that variation? The authors pursued an “abductive” approach — a method of reasoning that starts from anomalies and works backwards to explanations — and identified two factors that mattered enormously: the institutional environment of the country where a venture was based, and the design of the accelerator programme itself.

The institutional environment story is striking. In more gender-egalitarian countries — measured using the World Economic Forum’s gender-gap index — women-led ventures that participated in accelerators saw meaningfully stronger revenue growth than women who did not participate. The advantage was particularly pronounced when programmes were explicitly designed to support women entrepreneurs. In less egalitarian countries, however, the pattern flipped. Participation offered no measurable benefit, and for women-led ventures in women’s empowerment programmes in these environments, the association with post-acceleration performance was actually negative.

To put it plainly: programmes specifically designed to help women appeared, in less egalitarian settings, to harm the very people they were meant to serve.

The programme design findings reinforce this picture. Having more women involved in delivering the programme — as mentors, instructors, and staff — was positively associated with outcomes for women founders, but only in more egalitarian environments. In less egalitarian settings, even this seemingly sensible design choice failed to move the needle in the right direction.

The pipeline problem

The paper does not stop at outcomes. It also examines who gets selected into programmes in the first place, and finds further grounds for concern. Even controlling for observable measures of venture quality, women-led ventures were less likely to be accepted than men-led ones. Ironically, the disparity was sharpest in programmes that were supposed to support women.

The selection findings around women on evaluation committees are particularly counterintuitive. One might expect that adding women to selection panels would help women applicants. The data suggest otherwise: programmes with more women selectors tended to select women-led ventures at somewhat lower rates. The authors resist drawing a firm conclusion from this, noting the finding could reflect either bias against women applicants or, more charitably, greater discernment about which women-led ventures would genuinely benefit from the programme.

Interviews with 13 programme managers add texture to the quantitative results. Programmes in more egalitarian, often richer countries tended to select on intangible qualities — coachability, a growth mindset, peer-learning potential. Programmes in less egalitarian environments were more focused on hard skills: technical expertise, capital-raising ability, the clarity of a pitch. The training offered differed too. One manager at a women’s empowerment programme in a less egalitarian setting described a bonding programme built around soccer games and communal drinking — activities that, whatever their merits, may not have been designed with female founders in mind.

The deeper problem, as one executive in a less egalitarian country put it bluntly, is that accelerators operating without supporting infrastructure risk accelerating women “into oblivion.” If the surrounding economy still disadvantages women entrepreneurs — in access to credit, in customer relationships, in social norms around female authority — then a few weeks of intensive training may raise expectations without improving the environment in which those expectations must be met.

Lessons for programme designers

The paper does not argue that accelerators are futile or that supporting women entrepreneurs is a lost cause. Its message is more nuanced and, in some ways, more demanding. Context is not just a backdrop; it is constitutive of whether an intervention works. A programme model developed in San Francisco or Toronto may simply not transplant to Lagos or Colombo — not because of any failing in the local ventures, but because the ecosystem that determines what founders can do with their new knowledge and networks is fundamentally different.

For those running accelerator programmes, the implications are uncomfortable. Good intentions and feminist branding are not enough. Programmes that want to help women need to ask not just whether their curriculum empowers women, but whether the environment their graduates re-enter will allow that empowerment to translate into outcomes. If it will not, the programme may need to do something harder than run a boot camp: it may need to engage with the structural barriers that exist beyond its doors.

For women founders themselves, the paper offers a more tactical message: the decision to join an accelerator is not obviously beneficial. Where you are matters. What the programme does matters. Choosing carefully — and perhaps sceptically — is not cynicism. It is strategy.

This is not the conclusion that the accelerator industry, nor the well-funded foundations that support it, will find comfortable. Structural interventions are appealing precisely because they appear to fix inequality without requiring the harder work of changing the institutions that produce it. The evidence from Dutt and Kaplan suggests that, at least in entrepreneurship, there may be no such shortcut.

This is the embarrassing predicament at the heart of AI evaluation. The field has spent years producing benchmark after benchmark—elaborate obstacle courses of questions and tasks against which models are ranked—yet the resulting scores have delivered neither reliable explanation nor reliable prediction. A model that excels at one benchmark frequently stumbles on another that supposedly tests the same thing. Companies crow about leaderboard victories that regulators and deployers cannot meaningfully interpret. The yardstick, as one paper delightfully put it, is broken.

A study published this week in Nature, by a large team led by Lexin Zhou of Princeton University and José Hernández-Orallo of the Universitat Politècnica de València, proposes a remedy. The authors introduce what they call “general scales” for AI evaluation—a standardised system of 18 dimensions that can characterise both what a task demands of an AI system and what an AI system is actually capable of. The result, they argue, is a framework that can explain AI failures in intelligible terms and predict AI performance on tasks the system has never seen before.

Rulers and report cards

To understand the problem the authors are solving, consider the difference between a thermometer and a beauty contest. Standard AI benchmarks resemble the latter: they rank contestants against one another, but the rankings shift whenever the field of contestants changes. A model that looks brilliant today may look mediocre once a better one arrives, even if nothing about the model itself has changed. Psychometricians call this “populational dependence,” and it is the reason IQ scores, which suffer from the same flaw, have to be periodically re-normed as populations change.

What the authors want instead is a thermometer: an absolute scale on which a temperature of 37 degrees means the same thing regardless of who else is in the room. They design their scales using rubrics—detailed written descriptions of what it means for a task to demand, say, a “level 3” of logical reasoning or a “level 4” of knowledge of natural sciences. Each rubric runs from 0 (no demand) to 5+ (very high demand), with the calibration rule that a level-l task is hard enough that only roughly one person in 10^l-1 could solve it correctly. This makes the levels consistent across dimensions and, crucially, independent of which AI systems happen to exist at the moment of measurement.

The rubrics are applied automatically—by GPT-4o acting as an annotator—to each item in a benchmark, producing a 19-dimensional “demand profile.” (The nineteenth dimension, “unguessability,” captures how much a multiple-choice format artificially inflates scores by giving away the answer through elimination.) Applied to 16,108 items from 20 benchmarks, the system generated nearly 290,000 annotations. Crucially, human annotators agreed with GPT-4o at a rate that measurement scientists would consider respectable, with inter-rater agreement scores averaging 0.86 on a scale where 0.7 is “good.”

What benchmarks don’t measure

The first payoff of this apparatus is diagnostic. Once you can characterise what each item actually demands, you can ask whether a benchmark is measuring what its designers claimed. The answer, with depressing regularity, is: not really.

Consider the galaxy of benchmarks marketed as tests of “mathematical reasoning.” The authors find that these vary wildly in what they actually require. GSM8K, a widely used arithmetic dataset, concentrates most of its items at low levels of quantitative and logical demand—so low that even modestly capable models sail through it. OlymMATH Hard, as the name implies, reaches demands of 4 and above on the same dimensions, explaining why the same model that achieves 90% on GSM8K manages only 13% on OlymMATH Hard. GPQA, meanwhile, does poorly not because it is hard at reasoning but because it is hard at knowledge—specifically, knowledge of natural and applied sciences that the task’s designers arguably did not intend to test at all.

This is what the authors call a failure of “construct validity”: the benchmark does not measure the construct it claims to measure. They introduce formal criteria to detect this. A well-designed benchmark should be “sensitive”—showing a wide spread of demand levels on its claimed dimension—and “specific”—showing low demand levels on everything else. Table after table in the paper demonstrates that neither criterion is met by most benchmarks currently in circulation. Among the more entertaining findings: the SAT, a widely used college-admissions test, fails not because it is poorly designed but because it scores so high on “atypicality”—a measure of how unusual or contamination-resistant its items are. High atypicality means the model has probably seen the questions before during training, inflating its apparent score.

What models actually know

The second payoff is the ability to characterise AI systems in terms that remain stable across benchmarks. For each model, the authors construct an “ability profile”—18 numbers, one per dimension, each representing the demand level at which the model has a 50% chance of succeeding. These profiles are derived purely from the model’s own results, without reference to any other model, making them immune to the population-dependence problem.

The profiles reveal patterns that aggregate scores obscure. Across the three families of models tested—OpenAI’s GPT and o1 lineages, Meta’s LLaMA-3, and DeepSeek’s R1-Distilled variants—the authors find that raw knowledge (what a model knows about chemistry, history, or medicine) scales predictably with model size. Larger models know more, as one might expect. But reasoning capabilities tell a different story. Models trained with reinforcement learning and integrated chain-of-thought—the o1 and DeepSeek-R1 families—show dramatically higher abilities in logical and quantitative reasoning, and also, perhaps surprisingly, in “mind modelling and social cognition.” This last finding suggests that the slow, deliberate thinking encouraged by chain-of-thought training benefits not just symbol manipulation but also the kind of perspective-taking involved in understanding other minds.

The authors also demonstrate that their scales allow them to resolve what they call “conflicting claims” in the literature. In 2022, researchers were hailing LLMs as capable of complex reasoning. By 2025, other researchers were insisting that LLMs cannot do non-trivial reasoning at all. Both camps, the paper argues, were right about the models they tested—they just tested models at different points on the same underlying scales, and used benchmarks with very different demand profiles. The contradiction dissolves once demands and abilities are expressed in commensurate units.

Prediction, not just description

The most practically significant result is on prediction. The authors train a simple random forest classifier—a lightweight statistical model—to take a task’s 19-dimensional demand vector as input and predict whether a given LLM will succeed on that task. The results are striking. For in-distribution tasks (tasks similar to those used in training the predictor), the classifier achieves an area under the receiver-operating-characteristic curve of around 0.84—substantially above chance—while remaining well-calibrated in its probability estimates. Against two much heavier baselines—a model using word embeddings and a fine-tuned 8-billion-parameter LLaMA model—the demand-based predictor is competitive or better on discrimination, and far superior on calibration.

The more important test is out-of-distribution: can the predictor generalise to entirely new benchmarks it has never seen? Here the demand-based approach retains an advantage of 0.75 AUROC, while the embedding and fine-tuned-LLM approaches fall apart, dropping to 0.48 and 0.68 respectively. The reason is conceptually simple: demand annotations describe the nature of a task in terms that remain meaningful regardless of where the task comes from. Embeddings, by contrast, capture superficial textual similarity and fail when the new task looks lexically different from training data, even if it is cognitively similar.

This predictive power has immediate practical uses. Routing queries to the model most likely to succeed—rather than always using the biggest and most expensive model—becomes far more tractable when you can estimate success probabilities before running anything. “Reject before you run” mechanisms, which flag queries likely to exceed a model’s reliable capabilities, become feasible. So does red-teaming: if you know a model is weak at metacognition, you can generate adversarial tests that target exactly that weakness.

Limitations and the road ahead

The framework is not without gaps. Its rubrics cover only text, excluding images, audio, and actions—making it inapplicable to multimodal or agentic systems without extension. The battery contains very few items at the highest demand levels, which limits the precision of ability estimates for the most capable models; as AI improves, the scales themselves may need to be stretched upward. And the rubrics are in English only, raising questions about whether the demand levels translate cleanly across languages and cultures.

There is also the fundamental irony, acknowledged by the authors, that the scalability of their approach depends on using an LLM—GPT-4o—as the annotator. This means the quality of the entire edifice is partly hostage to the quality of one commercial model. DeepSeek-V3 proved a capable substitute in tests, but a smaller open-source model performed noticeably worse, suggesting that the rubrics are not yet so self-evident that any model can apply them reliably.

None of these limitations is fatal. What the paper establishes is a direction: that AI evaluation can be scientific rather than merely competitive, and that the features needed to predict AI behaviour can be defined in terms interpretable to humans rather than extracted from inscrutable embeddings. In a moment when regulators are trying to write rules about what AI systems can and cannot safely do, and when companies are making deployment decisions worth billions of dollars, the value of a reliable ruler is difficult to overstate.

Thermometers, after all, did not end illness. But they made medicine possible.

A recent study from Penn State University explores such displays of empathy towards machines (and indeed from machines to humans). The researchers wanted to understand whether we can empathize with machines, and whether they can empathize in return.

Wired for feeling

The researchers argue that empathy has a couple of components. The first is so-called vicarious emotion, which is when we attempt to take on the felt experience of another person. The second component is compassion, when we have genuine concern for another’s suffering. They believe that both of these could be directed at robots, even if the robots are unable to either notice or, indeed, suffer.

The evidence seems to suggest that this might actually be the case. The researchers describe how we show measurable increases in skin conductance when we see a robotic dinosaur being harmed. What’s more, we also show a change in our mood when robots are placed in positive or negative situations. Our brains also respond differently to painful versus non-painful stimuli applied to robots, albeit to a lower extent than we do when humans are in the same situations.

While this isn’t hard proof that people are capable of empathizing with machines, and certainly not in the way we tend to with other humans, it’s not nothing either.

The researchers believe these findings are an example of so-called “motivated choice,” which is more deliberate than automatic. People decide, consciously or otherwise, whether to engage empathetically with a robot. This decision is based on things like our attitude towards technology and our curiosity levels. Typically, children are more willing than adults to attribute inner lives to robots, with social norms widening this gap as they become embedded within us.

The uncanny helper

The flip side of this is also interesting. Can machines empathize with humans, and if they do attempt it, how do we tend to respond? The research on this is noticeably thinner, but it’s nonetheless encouraging. It shows, for instance, that we’re more likely to help a robot if the machine acknowledges a kind of shared feeling.

Similarly, bystanders are more likely to intervene when a robot is being abused if the robot expresses distress. Even older people show positive responses, with elderly patients given robotic companion animals showing reduced psychiatric symptoms.

The obvious question is whether any of this is real. Robots aren’t capable of actually showing empathy. It’s a programmed response, so is our empathetic response merely a cognitive illusion triggered by clever design? The researchers believe that our framing might be too narrow.

What matters is not the inner life of the robot but the social reality that human-robot interaction creates. Dismissing robot empathy as “fake” because its target is non-sentient may itself be a form of empathy avoidance.

Building better humans

With interactions between man and machine only likely to increase in the years ahead, the researchers also ponder whether these interactions could be designed to enhance human empathy rather than dilute it. If we can learn to extend moral consideration to a machine, even when considerable cognitive distance exists between us, we might also be more willing to extend it to other humans, especially those who are different from us.

Of course, there is also the potential for machines to be used for darker purposes. Designers of robots could exploit human empathetic tendencies for commercial or manipulative ends, engineering emotional attachment that serves the manufacturer rather than the user. What’s more, if we delegate caring responsibilities to machines, it could erode our own caregiving capabilities.

Of course, it’s important to note that the research literature on the topic is fairly meagre. What research that does exist tends to have small samples, inconsistent definitions of empathy, and a bewildering variety of robot types. For us to make stronger conclusions, it’s clear that we need larger and better-designed studies.

As a general direction of travel, however, it’s nonetheless interesting, not least as the inclusion of robots and other forms of technology becomes an ever greater part of our lives. Whether it’s in schools, hospitals, or homes, the question of how humans relate to them emotionally is no longer a philosophical curiosity. Instead, it’s equal parts a design problem, an ethical challenge, and also an opportunity.

Examples like HitchBOT show that humans do have feelings towards robots, but the task now is to understand those feelings well enough to make them useful.

What makes the finding credible, and somewhat unsettling, is how it was produced. The studies from which word counts were drawn were not designed to measure conversation at all. Participants were being tracked for unrelated purposes—coping with breast cancer, adjusting to divorce, and studying the social effects of meditation. None of them knew their chatter was being tallied for posterity. This renders the data unusually clean: no one was performing for the hypothesis.

Creating disatance

The likeliest culprit, at first glance, is the smartphone. Its rise coincides neatly with the period under study, and the data offer modest corroborating evidence: adults under 25 shed words at a faster clip—about 452 per year—than those over 25, who lost around 314. Youth have migrated more thoroughly to screens. Yet the fact that older generations are declining too implies something more structural is at work. The self-checkout kiosk, the GPS that replaced the passer-by you once stopped for directions, the app that lets you order coffee without catching anyone’s eye: the texture of daily life has been quietly rearranged to minimise the need for speech.

One obvious objection is that texting and social media have merely displaced speaking rather than eliminated expression altogether. Total word output, across all channels, may well have held steady or even risen. The researchers are careful not to dismiss this possibility. But they are equally careful not to assume that typed words and spoken ones are interchangeable. Spoken language carries tone, spontaneity, and physical presence—qualities that a well-crafted emoji cannot fully replicate. Whether people who text more but speak less are socially worse off remains, as the researchers put it, unsettled.

There are limits to what the study can claim. Its participants were drawn almost entirely from Western, individualistic societies—mostly Americans, with a smattering of Europeans. The researchers decline to extend their conclusions to more communal cultures, where incidental conversation may still thrive. They also lack data beyond 2019, which means the pandemic and its aftershocks remain uncharted. Given that lockdowns and remote work accelerated precisely the forces the researchers identify, it would take considerable optimism to imagine the trend had reversed by the time offices reopened.

The findings arrive at a moment when policymakers have begun to take social disconnection seriously as a public-health concern. America’s Surgeon General has written of a loneliness epidemic; several countries have appointed ministers for it. What this research may offer is something such declarations typically lack: an objective, measurable signal. The loss is not abstract. It is 338 words a day—a brief exchange with a cashier, a question asked of a stranger, a remark to a neighbour who now slips past ungreeted. Small things, perhaps. But as the researchers observe, small absences add up too.

The researchers tested the assumption across four contexts: general interactions, project teams, one-to-one meetings, and conference settings. Participants rated both themselves and others on listening ability, drawing on observable behaviours—such as interruptions, share of speaking time, and displays of positive affect—as well as less visible tendencies like self-focused attention.

Useful results

The results are instructive. First, there is no clear “introvert advantage”. Second, in two of the four studies, extroversion was weakly but positively associated with perceived listening effectiveness. The implication is not that extroverts are inherently better listeners, but that the behaviours commonly associated with them—visible engagement, responsiveness—may be read as such.

This points to a more useful conclusion. Listening is not a fixed personality trait but a skill, comprised of behaviours that signal attention and interest. Quietude alone does not suffice; nor does talkativeness preclude attentiveness. What matters is whether the listener conveys presence—through cues like eye contact, nodding, and measured responses—while keeping self-directed thoughts in check.

For organisations, the lesson is practical. Rather than relying on crude personality heuristics, they would do better to treat listening as a trainable competency. In an economy where leadership, collaboration, and relationship-building are increasingly prized, that shift—from disposition to discipline—may prove overdue.