In “Is AI Smarter Than People?,” published as a Wall Street Journal editorial on April 24, 2026, Vivienne Ming, a rising star in the AI guru universe, discusses results from a forecasting experiment she conducted and concludes that two cultivated qualities — perspective-taking and intellectual humility — separate users who flourish with AI from users who experience brain atrophy.
The piece was published during a coordinated promotional cycle for Robot-Proof: When Machines Have All the Answers, Build Better People (Wiley, March 2026), Ming’s new book. The book promises to influence how the public comes to understand AI’s effect on cognition. On my view, Ming’s argument holds some truth, but it does not survive scrutiny. My quibbles with it are significant but pale in comparison to the core flaw: It reduces learning to a single dimension—and gets the dimension wrong.
The Tour, Not the Genre
A weak version of the criticism I want to make would be that op-eds are not research papers and Ming should not be writing about her opinion until she has written and published the peer-reviewed paper. Of course, that is unrealistic and counterproductive.
Op-eds are a legitimate venue for synthesis and persuasion, and researchers routinely discuss ongoing or unpublished work in mass-media settings. The line that I think gets crossed is not between popular and research writing. I push against publishing an opinion based on findings presented as the result of scientific research when the findings come from emerging research not yet peer reviewed.
Ming presents findings from what we’ll call the Polymarket experiment. Some time earlier, she told the New York Post that students using AI sometimes “show more than a 40% reduction in the gamma-band brain activity that indicates active cognitive engagement” when they use AI to answer their questions and attributes this figure to “my own data.” She tells the Journal‘s readers that mixed human-AI teams outperformed humans alone and AI alone in the Polymarket experiment.
She reports that approximately ninety to ninety-five percent of participants fell into ineffective patterns, that only five to ten percent were powerful “cyborgs,” and that three traits predicted cyborg performance. She tells EurekAlert and McGrath and Kalb essentially the same things. Across at least seven public venues that I can find, the same numbers and the same construct vocabulary travel as if they were established results.
What is missing is the document that would let any of this experiment be examined. As of the date of this writing, no preprint or peer-reviewed version of the Polymarket experiment has appeared that I can put my hands on in arXiv, PsyArXiv, PubMed, or Google Scholar.
Ming told a Kennedy Human Rights Center book club audience on April 6, 2026 that the paper was about to come out later that month (KHRC, 2026). It has not. The Kosmyna study at MIT, by contrast, was posted as a 200-page preprint with full methods within months of being conducted (Kosmyna et al., 2025). That is the relevant comparison — not editorials versus journals, but claims that can be examined versus claims that cannot.
The pattern across venues matters because the absence of a paper is not a single event. It is a sustained condition. A Business Insider piece appeared in March 2026 (Ross, 2026). The New York Post piece followed on April 10. The UC San Diego feature story distributed by EurekAlert appeared April 16 (Clark, 2026). McGrath’s Thought Sparks piece appeared April 13 (McGrath, 2026). Kalb’s blog Q&A ran in March (Kalb, 2026).
Ming’s own Substack post “The Enflattenning” appeared April 21 (Ming, 2026a). Then the WSJ editorial on April 24 (Ming, 2026b). The phrase productive friction appears in the Post piece and in McGrath’s account before it reaches the Journal. The arc is timed to the release of Robot-Proof.
The objection is not that Ming should keep quiet until peer review confers permission to speak. The objection is that the same unpublished experiment functions as the empirical anchor for a book tour, accumulating apparent corroboration through repetition rather than through documentation, and that readers receive the rhetorical authority of science without the means to check.
The Polymarket Comparison and Its Limits
Reconstructing the experiment from March’s Business Insider account and Ming’s own description in the EurekAlert interview (Clark, 2026), the design ran from late summer through fall 2025 with seventy-two participants, thirty-nine UC Berkeley students and thirty-three Bay Area residents, organized into teams of three, given one hour to make ten predictions about the future.
Polymarket is a decentralized, cryptocurrency-based prediction market platform where users can place bets on the outcomes of real-world events. Founded in 2020 by Shayne Coplan, the platform allows individuals to trade shares representing the likelihood of future events across various categories, including politics, sports, economics, popular culture, and weather.. Ming’s participants had one hour to make a prediction about the future to be tested against Polymarket as the benchmark.
The unaided humans were not going to beat Polymarket regardless of who they were, because the comparison is not between minds. It is between three human minds operating briefly and a price-discovery mechanism specifically engineered to outperform individuals.
The AI-alone result is similarly underdetermined. Large language models trained on extensive corpora that include base-rate analyses and historical event discussions in my opinion will outperform untrained humans given an hour to forecast.
Ming reports that the smallest open-source model used was “better than the best human by a lot.” Approximately ninety to ninety-five percent of participants fell into the automator and validator patterns; about ten percent of the Berkeley students were cyborgs. Three traits — perspective-taking, intellectual humility, and curiosity — predicted cyborg performance, she claims.
Working from Ming’s own descriptions in popular-press venues is a constrained position, and I will not build elaborate methodological critiques on it. I cannot evaluate methods I cannot see.
What I can do is note what the structure of the comparison guarantees regardless of methodological detail. I can comment on horse-race research generally, the sort that was carried out in the late 20th century which allegedly supported systematic, explicit phonics. Ming’s description of the experiment is a horse race, and she set it up so that her horse would win. There is no reason to assume that either humans alone or AI alone would do better than AI and human together.
The cyborg finding is where Ming’s interesting claim lives and my worry arises. Ming attributes cyborg performance to perspective-taking, intellectual humility, and curiosity.
The published media accounts do not say how these qualities were measured. I will not guess at the measurement procedure. I will note only that an explanation is owed: when a study attributes outcomes to specific psychological constructs, the reader needs to know how the constructs were operationalized. Without that, the explanation is not comprehensible.
The “three students” claim deserves attention. Ming presents it as the headline result of the experiment — three Berkeley undergraduates without subject expertise matched the accuracy of prediction markets backed by tens of thousands of financially motivated forecasters.
The AI became a sparring partner, according to Ming. The teams pushed back, demanding evidence and interrogating assumptions. When the AI expressed high confidence, the humans questioned it. When the humans felt strongly about an intuition, they asked the AI to come up with a counterargument. They were the only group to consistently rival the prediction market’s accuracy. On certain questions, they even outperformed it
If true, this is a remarkable finding, the kind of finding that requires a paper. We do not know which three students. We do not know how their accuracy was scored against the markets. We do not know whether the comparison was made on the same questions, in the same time window, with the same information available. We do not know whether the result replicated. We are told, on Ming’s authority, that it happened. There is no peer review.
Polymarket prediction is a domain where AI has structural advantages and where the human contribution is necessarily marginal, supplying contextual or temporal information the model lacks. Ming’s prescription, treat AI as a “savant collaborator,” is good advice for this task because the task happens to fit it.
Whether the prescription generalizes to creative work, ethical judgment, teaching, clinical reasoning, or the daily acts of meaning-making that constitute most of cognitive life is an open question this experiment does not address.
The EEG claim deserves a separate note. The “more than 40% reduction in gamma-band activity” Ming has repeated across venues was, on the Business Insider reconstruction, taken from “a couple” of participants in this same study.
Ming’s description in the EurekAlert interview confirms this: “I put an electroencephalogram or EEG on a couple of them and compared to people reasoning on their own — or even just using Google — there was dramatically less cognitive activity.”
Two people. Generalizing from two participants to a population-level cognitive risk is not a finding one defends; it is a number one chooses not to publicize until the sample is adequate.
The Socratic-fine-tuning intervention described in the EurekAlert interview is interesting in its own right and deserves separate scrutiny. Ming reports that fine-tuning a small open-source model not to give answers doubled the cyborg rate. She does not say how the cyborg classification was made, whether the doubling was statistically reliable, what the comparison condition was, or whether the effect persisted beyond the study session.
The intervention is being deployed rhetorically as evidence that AI design can produce “productive friction.” As an empirical claim, it requires the same documentation any intervention study would require. It does not have it.
Perspective-Taking and the Operational Question
The deeper problem is theoretical. Ming names two qualities, i.e., perspective-taking and intellectual humility, as if they were neutral descriptions of folk virtues and announces, in the WSJ editorial, that they are “at root, emotional skills.” The first of these terms has a specific provenance in the research literature.
Perspective Taking is one of the four subscales of the Interpersonal Reactivity Index developed by Davis (1983), and Yu et al. (2025) classify it as cognitive empathy in their empirical work on LLM empathy.
Within the IRI, the subscales are explicitly partitioned. “Fantasy” and “Perspective Taking” together index cognitive empathy. “Empathic Concern” and “Personal Distress” index affective empathy. The Perspective Taking subscale measures the tendency to spontaneously adopt the psychological point of view of others, a representational operation distinct from the affective subscales that measure other-oriented sympathy and self-oriented distress. This cognitive-affective distinction has been validated through decades of factor-analytic work and reproduced in the Basic Empathy Scale developed by Jolliffe and Farrington and used alongside the IRI by Yu et al. (2025),
Two clarifications are due here. First, the cognitive-affective distinction within the IRI is operational, not ontological. The instrument partitions its constructs along a measurement-tradition convention that makes cognitive operations on others’ mental states tractable to self-report. It does not commit its users to the view that cognition floats free of affect at the level of brain function or lived experience.
As I will argue shortly, the embodied-cognition literature — including Borghi and colleagues’ work — treats human cognition as deeply imbricated with affect and interoception. The point about Ming’s relabeling is not that cognition and emotion are sealed off from each other in reality. It is that within the measurement tradition Ming is invoking when she invokes constructs like cognitive offloading, perspective-taking is a cognitive subscale, and recoding it as an emotional skill in passing prose, without argument, smuggles a category change past the reader.
Second, the relabeling has empirical consequences for Ming’s own thesis. Yu et al. (2025) tested GPT-4 against human participants on the IRI and the BES to see whether GPT-4 has empathy. They found that GPT-4 reproduces the cognitive/affective factor structure of human empathy but scores significantly lower on every dimension.
The gap was larger on affective empathy than on cognitive empathy. Cohen’s d for the affective gap reached 0.87. Cognitive empathy showed a significant difference but a smaller effect. If perspective-taking is the cognitive construct the IRI says it is, then GPT-4 is closer to humans on Ming’s prized quality than it is on the genuinely affective dimensions she rhetorically conflates with it.
The thing AI most plausibly lacks is affective empathy: empathic concern, emotional contagion, the felt response to another’s state. By moving perspective-taking from the cognitive column to the emotional column without justification, Ming names the wrong virtue for her own argument.
The same pattern of borrowing-without-grounding shows up in her treatment of intellectual humility. The term has currency in philosophy and psychology where it functions as a metacognitive construct, i.e., calibrated awareness of the limits of one’s own knowledge.
Ming relabels it “emotional courage” and proceeds as if relabeling were argument. I am not in a position, from the materials in front of me, to point to a specific instrument or peer-reviewed source that adjudicates the redefinition. I can point only to the substantive observation that monitoring one’s own knowledge is an operation on cognition, and the rebranding from cognitive to emotional carries argumentative weight Ming has not made explicit.
Friction Reduces Learning to One Dimension, and Picks the Wrong One
The third move in Ming’s argument is the one that has traveled furthest through the press cycle. Productive friction is the phrase she keeps reaching for, and it is doing more work than she has evidence to support. She relies on it primarily as a metaphor.
The objection is not that metaphor is illegitimate in public discourse. Metaphors are useful conceptual tools, and demanding mathematical isomorphism from them is pedantry. The objection is that Ming is not deploying friction as a casual figure of speech. She is using it as a design principle, telling AI companies they should preserve productive friction, telling educators and parents they should cultivate it, telling readers it is the variable that distinguishes flourishing from atrophy.
In the EurekAlert interview, she puts it as a directive: “we should be building systems designed around productive friction — systems that challenge you, that help you explore, that don’t just hand you the answer.” A design principle needs a referent. Friction in Ming’s usage does not have one.
There is a serious behavioral literature on conditions under which difficulty improves learning — Robert Bjork’s desirable difficulties framework, John Sweller’s cognitive load theory — and citing it as the rigorous version of what Ming gestures at would be an obvious move. I want to resist that move, because it concedes too much.
Desirable difficulties and cognitive load are themselves behavioral correlations dressed in cause-effect clothing stretched as the explanation for a phenomenon of interest to all of us we call “human learning.” They report retention curves and transfer effects across training conditions, but they do not specify neural mechanisms. There are no fMRI studies that show friction lighting up a particular system.
The older frameworks are better-tested folk theory than Ming’s offering, but they remain folk theory in the relevant sense: they identify a behavioral pattern without defining it clearly and without explaining what produces it. Indeed, friction suffers from the same limitation words like intelligence inherently display. Their referent is what we say it is according to how we measure it.
The deeper problem is that the friction model commits to a phenomenology of learning that does not match how learning actually feels or works either subjectively or objectively. Ming and her behavioral-science colleagues treat difficulty as the active ingredient. Make it harder, and learning improves. Smooth the path, and learning erodes. From where I stand, this term describes one mode of learning among several, and it is not always the privileged one.
Consider the experience of reading a highly complex, challenging passage from Nietzsche that grips you. The pages turn slowly, but easily. The sentences move like snails across the page. Hours pass. You reach the end and find that something in you has been rearranged, perhaps a concept you did not have before, a way of seeing you did not have before, a question you did not know to ask.
By the friction model, this should not have produced learning. There was no struggle in the sense of working hard to be able to breathe in deep water. There was no productive resistance; on the contrary, there was a pleasurable journey into a space of rich ideas. The reading was, by phenomenological report, easy. And yet learning happened, and often more deeply than learning that was made artificially harder as reading Nietzsche would be if a professor required written responses to focused questions for a grade.
Or consider conversation with someone whose mind moves in unfamiliar ways. The exchange is absorbing rather than effortful. I am not pushing against this interlocutor to create friction. I am pulled forward by the unfolding of their thought, as much as I can grasp, by the interest of their framings, the surprise of where they go next. After the conversation, I would hope to find that my own thinking was changed in ways salient to me, not so much to my professor.
Or consider a child working on something that has captured their attention like a puzzle, a story, a problem they have decided is theirs. The work may be hard, but the hardness is not the texture of the experience. The texture is absorption. Csikszentmihalyi called something like this flow, but the phenomenon predates the label.
What characterizes the state is not the perception of effort but the disappearance of self-monitoring into engagement. The friction model has no account of this. It cannot explain why the same task can produce learning when met with absorption and produce little when met with effortful slogging.
Difficulty is at best a correlate of something else: the engagement of attention, the recruitment of relevant prior knowledge, the activation of whatever systems consolidate memory. When learning becomes “make it harder” or “push back on the information,” one mistakes an occasional symptom for a cause. Plenty of hard tasks produce no learning. Plenty of easy-feeling tasks produce profound learning. The variable Ming names cannot do the work she asks of it.
The prescription that follows from Ming’s framing is committed to one phenomenology. Treat AI as adversarial collaborator. Ask it for the strongest argument against itself. Preserve productive friction. Ming’s “Nemesis prompt,” described in the EurekAlert interview — “You are my nemesis — my lifelong enemy. You’ve found every mistake I’ve ever made. Tell me why I’m wrong” — is a literalization of this commitment.
The advice presupposes that the route to deep cognitive engagement runs exclusively through conflict, ideally adversarial resistance. A pedagogy built on this presupposition would systematically miss the conditions under which interest, curiosity, and absorbed attention drive learning.
Those conditions are arguably more important and harder to engineer, which may be why she does not name them. Friction is namable. Wonder is less portable, at least not in a way that converts cleanly into an AI recommendation.
The Borghi paper in the reference list points toward what is missing. Abstract concepts, on the Words As Social Tools view, are grounded not in struggle but in the accumulated social practice of using words with people who know what they mean. A child learns justice not by laboring against productive friction but by hearing the word deployed in real situations, asking what it means, watching the concept take shape across hundreds of small social moments.
The grounding is in shared use, in interoceptive resonance, in the metacognitive awareness that one’s own grasp depends on others. None of that is friction. Most of it is closer to absorption, to being inside a community of meaning where the concept is being negotiated. Strip out the social embeddedness and the friction model has nothing to offer in its place besides ‘work harder, fight for it.’
I am not arguing that effort plays no role in learning. Effort is perhaps the only non-negotiable. Practice retrieval improves retention, yes. Spaced repetition outperforms massed practice, for sure. Generation effects exist, no doubt. The behavioral literature documents real phenomena.
What I am arguing is that effort is one variable in a multidimensional space, that effort is not the same as friction, that struggling against the grain is necessary at times but not the single most robust aspect of learning. Ming has flattened that space into a single axis with friction at one pole and atrophy at the other. The flattening is rhetorically efficient because it produces a clean prescription, and it is dangerously limiting.
What the Genuine Affective Substrate Looks Like
Ming gestures repeatedly at emotion. Perspective-taking is “at root, an emotional skill.” Intellectual humility is “emotional courage.” Friction matters because it engages something deeper than cognition, presumably emotion. But she does not describe any specific emotional process, neural system, or developmental mechanism. The editorial gestures toward affect while remaining cognitivist in substance, naming cognitive operations and announcing they are emotional, without specifying what the emotion does or how it operates.
The scientific literature describes the affective substrate of human cognition with considerably more precision. Three lines of evidence converge on a substantive picture.
The human self-domestication hypothesis
Hare (2017) reviews the human self-domestication hypothesis, which proposes that Homo sapiens was selected for in-group prosociality and tolerance, producing a temperamental shift that enabled flexible cooperative communication. The argument is not that humans developed cognitive perspective-taking through cultivation. It is that evolutionary selection acted on the affective substrate of emotional reactivity, tolerance, the capacity to remain prosocial in cooperative contexts, and that cognitive flexibility followed as a downstream consequence.
Hare emphasizes that chimpanzees can take perspectives and can communicate, but they struggle to do both simultaneously because their intolerance disrupts joint effort. The human innovation, on this account, was not the cognitive operation we think of as perspective-taking but the emotional regulation that lets the cognitive operation function in social contexts. The temperamental shift is the engine.
The emergence of a sapien brain
Barger and colleagues (2014) provide neuroanatomical evidence for what selection produced structurally. Their comparative analysis finds that the human hippocampus is approximately fifty percent larger, the lateral amygdaloid nucleus approximately thirty-seven percent larger, and the orbital frontal cortex approximately eleven percent larger than would be predicted for a hominoid of human hemisphere volume.
The hippocampus is the human save button. It is responsible for turning short-term experiences into long-term memories. If you learn a new coworker’s name or remember what you had for dinner last night, your hippocampus is doing the work. It’s also GPS. It helps you build a mental map of your surroundings. When you navigate your way home from work without needing a map, or remember where you parked your car, you are relying on your hippocampus.
The amygdala is the alarm system. It is constantly scanning your environment for threats. If a car suddenly swerves into your lane, the amygdala instantly triggers your “fight-or-flight” response, making your heart race and pumping you full of adrenaline before you even consciously realize what is happening. It attaches emotions to your memories and is the reason why smelling a certain perfume might instantly make you feel nostalgic, or why you remember a terrifying event vividly but forget what you did on a boring Tuesday.
The orbital frontal cortex helps you weigh rewards against punishments to make choices. If you are deciding whether to eat a second piece of cake (reward) despite knowing it will give you a stomachache (punishment), this part of your brain is crunching the data. It helps you process social cues and behave appropriately in different situations, and it stops you from saying something rude at a dinner party by helping you predict how others will react to your behavior.
The hippocampus, the amygdala, and the orbital frontal cortex comprise much of the circuitry implicated in emotion processing, social valuation, and the integration of affect with decision-making. The human brain is not simply a larger primate cortex with the standard mammalian limbic system underneath. It is disproportionately invested in affective and social-affective processing.
Now how does one square the biological emergence of these brain parts and the potentials they brought to humans for emotionality as a driving force with the proposition that cognitive friction is the key ingredient in learning? Ming treats friction, resistance, pushing back as the sine qua non of learning. Make it harder, and learning improves. Smooth the path, and learning erodes. Struggling produces better learning than cooperating.
Words as Social Tools
Borghi and colleagues (2019) propose, in their Words As Social Tools framework, that abstract concepts are grounded not only in sensorimotor experience but in social, linguistic, interoceptive, and metacognitive experience. The meaning of concepts like uncertainty, freedom, justice, or principle depends on the felt bodily states of the user, on the social experience of needing others to fix meaning, and on the metacognitive sense of grasping or not grasping.
Interoception is the perception of sensations from inside the body. While exteroceptive senses (like sight, hearing, and touch) help you understand the outside world, interoception helps you understand your internal state. It allows you to answer the question, “How do I feel right now?”
Their rating studies show that abstract concepts score higher on social metacognition — the felt sense that others’ contributions are crucial for understanding — and on inner grounding through cultivation of interoception and emotion.
Converging lines of theory
These three lines of evidence converge on a claim Ming’s editorial does not make. Human cognition is built on an evolved affective substrate, instantiated in specific neural structures, and expressed through interoceptive and social grounding of the abstract concepts we use to think alone and together. This is not a virtue to be cultivated. It is a biological inheritance, exercised through daily embodied practice with other minds.
Notice how this reframes the learning question. If concepts are grounded in social and interoceptive experience, the conditions for learning them are conditions of being inside a community of meaning, with the body present and engaged, with the felt sense of grasp or non-grasp available as a guide.
None of that maps onto friction. All of it is closer to absorption, to the kind of engaged participation in shared practice that lets a child come to mean justice by encountering the word in situations where justice is at stake.
What the Literature Supports — and What It Does Not
Honesty requires being as precise as I know how to be at this stage in my development about what the scientific literature establishes and what it does not.
It supports the claim that human cognition has an evolved affective and social substrate (Hare, 2017), that this substrate has neuroanatomical correlates (Barger et al., 2014), and that abstract concepts depend on interoceptive and social grounding (Borghi et al., 2019). It supports the claim that perspective-taking, as operationalized in standard empathy instruments, is a cognitive construct in the field’s measurement tradition, and that current LLMs differ from humans more on affective empathy (identifying feelings) than on cognitive empathy (taking a perspective) (Yu et al., 2025).
It does not establish that Ming’s specific recommendations are wrong on their own terms. Some users do develop more productive working relationships with AI than others. Adversarial prompting may produce better outcomes than naive acceptance. Convergent independent evidence like Kosmyna et al. (2025) on reduced neural connectivity in LLM-assisted essay writing broadly supports the direction of her concern about cognitive disengagement during passive AI use. The concern is real. The phenomenon is real.
What the literature refutes is the framing. The relabeling of cognitive constructs as emotional is permissible but must be named. The appeal to a friction metaphor without theoretical content and an epistemic commitment to a phenomenology of learning captures only one mode of learning. The implicit claim that the affective substrate of human cognition is something to be willed into existence or taught through individual cultivation rather than something already present in our embodied biology and our social practices is indefensible.
A rigorous version of Ming’s argument would say something like this: AI use is reshaping daily cognitive practice in ways that may reduce engaged participation in the social and embodied conditions through which humans actually learn and understand.
The risk is not that we fail to cultivate productive friction. The risk is that we increasingly route abstract reasoning through systems that lack the interoceptive and social grounding our concepts require with downstream consequences for how we use those concepts. The remedy is not personal virtue. It is institutional and pedagogical: preserving the practices of embodied, social meaning-making in education, work, and ordinary life, even as AI tools become available for tasks where they genuinely help.
That argument would draw on Hare, Barger, Borghi, and Kosmyna without needing to invent new constructs or smuggle category changes past the reader. It would not require an unpublished experiment. It would not require a metaphor borrowed from physics. It would not require a book tour.
The Public Deserves the Story Without the Tour
There is a pattern in Ming’s public communication. The Polymarket experiment travels with her between venues, and the data accumulates apparent corroboration through repetition rather than through documentation. Rita McGrath’s “Thought Sparks” piece is a useful illustration.
McGrath, a Columbia Business School professor with substantial credibility, repeats Ming’s claims about “data from 122 million people” and her “thirty years at the intersection of human and machine intelligence.” Neither claim has documentary support I can locate, but McGrath does not flag the absence.
McGrath also writes that “a recent paper — replicated by an independent research group — found that the people who get the most out of AI are disproportionately high in perspective-taking” without citing it. When a credentialed third party repeats unsupported claims without flagging them, the chain of attribution lengthens and the absence of underlying evidence becomes harder to see. The chain itself begins to function as the evidence.
Readers should know that Ming’s documented peer-reviewed work is concentrated in auditory neuroscience and sparse coding from roughly 2003 to 2009 —for example, see a 2009 paper with Lori Holt in the Journal of the Acoustical Society of America on efficient coding in human auditory perception, and conference work in the same tradition.
Her sole recent arXiv preprint, posted in December 2025, uses 2,500 LLM-simulated agents to model the lifetime impact of resilience interventions (Ming, 2025). It contains no human participants. Its conclusions about wealth, mortality, and dementia outcomes reflect simulated agent behavior, not empirical data about humans. This is the published research record that underwrites the WSJ editorial’s claims about AI-induced cognitive decline and atrophy.
What the public deserves, especially the educators and parents being told that AI is cannibalizing human intelligence, is the science as it actually exists. Kosmyna and colleagues (2025), with full methods and a posted preprint, report that LLM-assisted essay writers showed weaker neural connectivity than search-engine users or unaided writers.
Hare (2017) gives us the evolutionary depth of the affective and cooperative substrate human cognition rests on. Barger and colleagues (2014) show what that substrate looks like neuroanatomically. Borghi and colleagues (2019) describe how the abstract concepts we reason with are grounded in embodied and social experience. None of these accounts requires the cultivation of a folk virtue like intellectual humility. None of them traffics in undefined metaphors. None of them flattens learning to a single dimension of effort.
The folk intuition that struggle helps you learn captures something real. So does the folk intuition that absorbed interest helps you learn. So does the folk intuition that being in on the right kind of conversation can change how you think. Ming has chosen one of these intuitions, dressed it as a theory, and built a prescription on it. The prescription has the shape of confident science.
Students and educators trying to think clearly about AI deserve better science. They deserve an account that distinguishes what we know from what we suspect, that names its theoretical and epistemic commitments, and that takes seriously the genuinely affective and embodied character of the cognition AI cannot supply because it lacks a hippocampus, an amygdala, and an orbital frontal cortex. That account is being written in the scientific literature. It is not being written in the Wall Street Journal. That account doesn’t sell papers.
Works Cited
Barger, N., Hanson, K. L., Teffer, K., Schenker-Ahmed, N. M., & Semendeferi, K. (2014). Evidence for evolutionary specialization in human limbic structures. Frontiers in Human Neuroscience, 8, 277.
Borghi, A. M., Barca, L., Binkofski, F., Castelfranchi, C., Pezzulo, G., & Tummolini, L. (2019). Words as social tools: Language, sociality and inner grounding in abstract concepts. Physics of Life Reviews, 29, 120–153.
Clark, C. (2026, April 16). What skills do humans need to become robot proof in the age of AI? EurekAlert / University of California San Diego. https://www.eurekalert.org/news-releases/1124563
Davis, M. H. (1983). Measuring individual differences in empathy: Evidence for a multidimensional approach. Journal of Personality and Social Psychology, 44(1), 113–126.
Hare, B. (2017). Survival of the friendliest: Homo sapiens evolved via selection for prosociality. Annual Review of Psychology, 68, 155–186.
Kalb, D. (2026, March 24). Q&A with Vivienne Ming. Book Q&As with Deborah Kalb. https://deborahkalbbooks.blogspot.com/2026/03/q-with-vivienne-ming.html
Kennedy Human Rights Center. (2026). KHRC Book Club examines the future of human potential with Robot-Proof by Dr. Vivienne Ming, moderated by Michael Schreiber. https://kennedyhumanrights.org/our-voices/khrc-book-club-examines-the-future-of-human-potential-with-robot-proof-by-dr-vivienne-ming-moderated-by-michael-schreiber/
Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X. H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv preprint arXiv:2506.08872. https://arxiv.org/abs/2506.08872
McGrath, R. (2026, April 13). Vivienne Ming on building robot-proof humans. Thought Sparks.
Ming, V. (2026a, April 21). The enflattenning. Socos Academy. https://academy.socos.org/the-enflattenning/
Ming, V. (2026b, April 24). Is AI smarter than people? The Wall Street Journal. [Online edition published as “AI Is Cannibalizing Human Intelligence. Here’s How to Stop It.”]
Ming, V. L. (2025). Quantifying the lifelong impact of resilience interventions via agent-based LLM simulation. arXiv preprint arXiv:2512.18803. https://arxiv.org/abs/2512.18803
Business Insider. (2026, March). A top researcher says a new divide is emerging in AI use — and most people are on the losing side. Business Insider. https://www.businessinsider.com/ai-impact-on-thinking-cognitive-skills-researcher-2026-3
Yu, T., Pan, S., Fan, C., Luo, S., Jin, Y., & Zhao, B. (2025). Can large language models exhibit cognitive and affective empathy as humans? Computers in Human Behavior: Artificial Humans, 6, 100233.

Really interesting article. Human intelligence is incomparable - AI is comparable.
Amazing work, Terry!