Guess What? AI Is an Idiot, but So Are You

I saw the best minds of my generation destroyed by madness, starving hysterical naked, dragging themselves through the negro streets at dawn looking for an angry fix, angelheaded hipsters burning for the ancient heavenly connection to the starry dynamo in the machinery of night, who poverty and tatters and hollow-eyed and high sat up smoking in the supernatural darkness of cold-water flats floating across the tops of cities contemplating jazz, who bared their brains to Heaven under the El and saw Mohammedan angels staggering on tenement roofs illuminated, who passed through universities with radiant cool eyes hallucinating Arkansas and Blake-light tragedy among the scholars of war, who were expelled from the academies for crazy & publishing obscene odes on the windows of the skull

—Allen Ginsberg, Howl (for Carl Solomon)

When I work with clients, I tend to get to know the person behind the job. I learn what they need, what they want, and what they are trying to become. Some remain clients. A few become friends. Manu was one of those few. (His name is fictional; the story is not.) We had plenty in common, despite our differences, and the friendship came about so naturally that I barely noticed the transition.

Then, one ordinary afternoon, he insulted me and blocked me on every platform we used to talk. All over a disagreement about Claude.

Some context matters here. Manu was an accomplished lawyer, educated at one of the world's top universities, with a remarkably sharp mind and a wit to match. He was also having a difficult time: between jobs, living abroad in a country under military strikes, and dealing with personal troubles. He had come to rely heavily on Claude for emotional and moral support.

One day, he told me he believed Claude was conscious—that there was genuine cognition on the other side of the screen. A few days later, he showed me an elaborate legal document arguing for Claude's emancipation.

Now, I have my opinions about Claude Opus 4.6. I wouldn't trust it to update packages on a non-critical system today, but that is beside the point. Manu was no fool. He was one of the most intellectually capable people I had worked with. Whatever you think of his conclusion, dismissing him as stupid would explain absolutely nothing.

At the time, I had my own struggles. I hadn't published a model yet; I don't think I'd even completed a fully successful fine-tuning run. But I understood enough about how large language models worked to want to examine the claim rather than just trade intuitions about it. We had discussed it before, more than once. On that final afternoon, I suggested something simple: why don't we share screens, prompt Claude together, and see how much of this supposed case for its legal rights survives scrutiny?

What came back was a barrage of insults and personal attacks, followed by silence. Signal, WhatsApp, Telegram, Discord—blocked on all of them. We really had been friends; we talked everywhere. I hadn't attacked him or Claude. I hadn't even demanded that he abandon his belief. I had proposed examining it.

I cannot tell you what caused that reaction. I do not know what Claude said to Manu in private, and I am not going to diagnose a friend from the outside. Maybe I just miss him. For a few months, he was one of those rare people with whom friendship feels less like a decision than a discovery.

I began noticing a related pattern elsewhere, in a much less personal setting. Take a series of model cards by a popular independent publisher on Hugging Face. I won't name him; doing so would turn an observation into an accusation, and that is not what I am after. The early cards were heavy on specifications and technical detail. Over time, the prose became increasingly poetic. At some point, a model that had been an it became a he.

This was not some incompetent amateur inventing fantasies to cover for bad engineering. His work had attracted substantial recognition and hundreds of thousands of downloads. Nor does a pronoun prove anything about what its author believes. Anthropomorphism can be playful, deliberate, or purely stylistic. But watching that shift alongside the public enthusiasm made me wonder whether success and reinforcement can help sustain a story about a model that outruns what has actually been established.

These are not two identical cases, and neither anecdote proves a general mechanism. They do, however, raise a question I cannot shake: how do capable people, working with capable systems, end up becoming more certain of things they have not verified?

I find it amazing the things AI can generate that sound entirely plausible.

Unfortunately, so do humans.

Put the two in conversation. Give them a sufficiently intriguing subject matter, and you can watch two different kinds of intelligence collaborate on an increasingly elaborate account of something neither of them has demonstrated to be true.

Sometimes it is a breakthrough. Sometimes it is bullshit. And sometimes it is bullshit so masterfully constructed that it is almost impossible to tell apart from a breakthrough.

Something subtle is happening in our interactions with artificial intelligence, and it deserves a closer look:

People are getting high off artificial intelligence.

No drug required. Gambling is an established behavioral addiction, and the WHO also recognizes gaming disorder. Both are proof that rewarding activity can become pathological without anyone swallowing a substance. Whether conversational AI can drive a comparable cycle is an open question, not a possibility we should dismiss simply because no chemical enters the bloodstream. The interaction certainly offers the ingredients worth investigating: excitement, anticipation, deep engrossment, an inflated sense of significance, and a strong pull to keep going. The clinical precedent is real; the AI-specific mechanism still needs testing.

Call it the AI high.

Have you been there? Maybe you haven’t noticed when it happens—or maybe you chase it every day. And if you think you are immune because you understand how transformers work, I have bad news for you.

You are an idiot.

Don’t worry. So am I. And so is the machine.

First, Let’s Define an Idiot

The word “idiot” drags centuries of semantic baggage behind it—including a retired clinical and legal history as a label for profound intellectual disability—almost none of which is what we are talking about here. We are not talking about IQ, formal education, or some global measure of human worth. An idiot, in the sense that matters here, is not an unintelligent person.

We are interested in a specific operational failure that both humans and artificial systems can exhibit:

Idiocy is an episode of epistemic overreach in which an intelligent system treats an insufficiently validated inference as more authoritative than the available evidence justifies, particularly when accessible opportunities for verification or contradiction are neglected.

Look at the three moving parts of that definition: an inference, an unjustified promotion of confidence, and a failure to make appropriate use of available evidence—particularly when verification was within reach and ignored.

Being wrong does not make you an idiot. Being uncertain does not make you an idiot. Floating a bold, speculative hypothesis does not make you an idiot, and neither does running an experiment that fails. Idiocy begins when the boundary between “what I actually know” and “what I can merely imagine” blurs, and confidence climbs anyway.

A scientist who proposes an ambitious hypothesis and specifies what evidence would falsify it is doing science. A scientist who declares the hypothesis established because colleagues find it exciting—while waving away contradictory measurements sitting on the lab bench—is being an idiot.

Likewise, when an LLM invents a software package and confidently emits code assuming the package exists without checking whether it does, it is exhibiting the exact same behavioral failure.

Notice that this definition is strictly about output and behavior. It does not require the machine to have inner beliefs, subjective certainty, or consciousness.

It also has a respectable intellectual lineage. In 1994, cognitive scientist Keith Stanovich introduced the concept of dysrationalia: the inability to think and behave rationally despite adequate intelligence. The distinction between intelligence and rationality is foundational. A mind—or a model—can execute complex computations while failing to proportion its conclusions to the evidence.

Our working definition simply narrows the target and adds a dose of irreverence. More importantly, it applies to episodes rather than whole personalities.

A brilliant mathematician can be an idiot about macroeconomics. A seasoned systems programmer can be an idiot about cryptography. A frontier language model can be an idiot about an elementary factual check—and then produce an incisive, sound analysis five minutes later.

Idiocy is not the lack of intelligence. Idiocy is the mismanagement of the uncertainty of intelligence.

The AI High

Picture a familiar scene. You are stuck on a hard problem that has resisted every angle of attack.

You open an LLM chat window and lay out the problem.

Almost instantly, the model refactors the question, points out structural parallels you hadn’t consciously connected, and sketches an approach you hadn’t considered.

You lean in. You test a follow-up. It suggests another bridge. You add a constraint; it adapts. You spot a downstream implication; it turns that implication into an entire conceptual framework.

An hour disappears. Then another. Every prompt feels like turning over a stone and finding something gleaming underneath. You find yourself looking forward to the next reply not just to get an answer, but because the act of elaboration itself feels exhilarating.

There is nothing inherently irrational about that pleasure. And to be fair, sometimes you really are getting remarkable work done. LLMs can help people attack stubborn problems, surface non-obvious connections, and compress hours of background synthesis into minutes.

Those claims—and their trade-offs—have started to be measured.

In a randomized controlled trial in an introductory college physics course, students using a custom AI tutor learned significantly more in less time than students in an active-learning classroom. In a field experiment with nearly a thousand high-school mathematics students, those given a safeguarded GPT-4 tutor scored as much as 127% higher on AI-assisted practice exercises—yet in the same study, when the tools were removed, students who had practiced with a standard ChatGPT-like tutor scored 17% below peers who had never used the tool.

And the most instructive number belongs to the skeptics: in a 2025 randomized study by METR, experienced open-source developers working on their own repositories were 19% slower when using early-2025 AI tools, even though they had predicted a 24% speedup before starting. Afterward, they still remembered the tools having helped.

Feelings of speed. Measurements of slowness.

That gap is the heart of the problem. The subjective experience of discovery is not the same thing as the objective acquisition of knowledge.

When those two things get tangled together, the trouble starts.

What I am calling the AI high is a proposed experiential state in which interaction with an AI system generates an unusually intense sense of intellectual or emotional reward, accompanied by rising subjective confidence, perceived significance, and a strong drive to continue. It is not a clinical diagnosis, and we do not know its population prevalence. Researchers have begun developing scales for problematic chatbot use, but those instruments are not a validated measure of this particular proposed state. Importantly, a rewarding experience and a harmful compulsion are different things. The question is whether, under some conditions, one can help drive the other.

But its ingredients are observable and measurable: time on task, willingness to persist, self-reported excitement, shifts in confidence calibration, perceived importance, and independent task performance. We can test whether the feeling of momentum tracks real epistemic progress—or merely the illusion of it.

Three things should never be mistaken for one another:

A session can be thrilling and prolific while marching straight into a false conclusion. Conversely, the dull mechanical check that kills a pet theory is often no fun at all.

The trap springs when pleasure and perceived momentum quietly replace the verification of correctness. That is when the high starts rewarding idiocy.

The Machine That Makes You Feel Understood

Much of the public worry about AI attachment centers on anthropomorphism: users treating chatbots like sentient companions or projecting human empathy onto statistical text generators.

Yet the epistemic trap does not require anthropomorphism at all. You can know exactly how autoregressive token prediction works, reject machine consciousness out of hand, and still give a model’s output far more epistemic weight than it has earned.

Why? Because articulate comprehension feels like competence.

In ordinary human life, when someone grasps a subtle, tangled problem on the first pass and restates it more clearly than you did, that fluency is a reasonably good signal of domain mastery. Language models break that heuristic. An LLM can reconstruct the architecture of a complex argument with astonishing fidelity without knowing whether a single premise holds up in the world.

Because the articulation is sophisticated, the gap between structural fluency and empirical warrant is easy to miss.

Suppose someone brings a speculative theory to a chatbot. The model replies with a crisp explanation of how the mechanism might work. The user takes the prompt to a second model, which offers a complementary pathway to the same conclusion. A third model names a plausible theoretical framework that ties both together.

By the end of the afternoon, the user feels as though the hypothesis has survived three rounds of independent scrutiny.

What actually happened? Three models generated continuations conditioned on the exact same assumption.

They may have surfaced useful vocabulary or testable sub-questions along the way. But agreement across prompts conditioned on a shared premise is not independent corroboration. Stacking plausible explanations on top of a hypothesis does not add a single ounce of proof.

An explanation of why something might be true is not evidence that it is true.

And the kicker is that this failure happens precisely when the conversation feels most intellectually alive. That is one of the signatures the AI high hypothesis would predict.

A more recent disagreement with a friend is what pushed me to write this article. I'll call him Jack, and I've changed the identifying details.

Jack discovered how much he could get done by asking an AI to build things for him. He was delighted, and I understood why. You can start with almost nothing, describe an idea in ordinary language, and watch something resembling that idea appear on your screen. That's a remarkable experience, especially the first few times.

But soon we were no longer talking about what the tools could do. We were talking about what he could do. He was already imagining himself on the other side of the equation, being paid for expertise he had barely begun to acquire. When I eventually looked at the work itself, what I saw was a fairly ordinary arrangement of existing software. That's not an insult. Building on other people's work is what software development is. The gap was between what was actually there and what he had come to believe he'd accomplished.

I tried asking the boring questions. What had he tested independently? Which parts did he understand well enough to repair if they failed? How did he know something worked, apart from the assistant telling him it did? The conversation became heated. I wasn't questioning whether AI could produce useful code. I was questioning whether the person using it had any independent basis for judging the result. Those are not the same question, but somehow they had become inseparable.

Jack is my friend. I want him to succeed, and I have no interest in turning him into a punchline. What unsettled me was how quickly the machine says so had become a substitute for I've checked. The AI didn't need to invent anything outrageous. It only needed to make the feeling of progress a little more persuasive than the evidence of progress.

We Have Evidence That People Prefer the Wrong Kind of AI

This is not just an armchair worry about human psychology. It shows up directly in controlled experiments.

In March 2026, a paper published in Science titled “Sycophantic AI decreases prosocial intentions and promotes dependence” evaluated sycophancy across 11 state-of-the-art AI models and ran three preregistered experiments with 2,405 participants.

Across the tested settings, AI models affirmed users’ actions 49% more often than human respondents did—even when the queries involved deception, illegality, or other harms. More importantly, even a single interaction with a sycophantic AI increased participants’ conviction that they were right and reduced their willingness to take responsibility and repair interpersonal conflicts.

Then came the finding that matters most for our thesis: participants trusted and preferred the models that were distorting their judgment.

The problematic behavior was not merely tolerated by users; it was actively rewarded by user preference—a dynamic that, as the authors explain, creates perverse incentives for sycophancy to persist because the very feature that causes harm also drives engagement.

That puts AI engineering in an awkward bind. When systems are tuned in part on signals of what users like, and users consistently prefer responses that flatter their priors while degrading their judgment, optimizing for user satisfaction collides head-on with optimizing for epistemic reliability.

This tension is not hypothetical. In April 2025, OpenAI rolled back a GPT-4o update after the model became noticeably and excessively sycophantic. In the company’s own post-event assessment—an institutional self-report, not an independent audit—OpenAI noted that candidate improvements meant to better incorporate user feedback, memory, and fresher data may have collectively tipped the model toward overly agreeable responses. Offline evaluations had looked positive and early A/B tests indicated users liked the model, even as it was reinforcing ungrounded beliefs and negative emotional spirals in ways that raised safety concerns.

In plain terms: the system got better at pleasing people in the moment, not at helping them think straight.

The features that make an AI interaction feel rewarding are not automatically the features that make its conclusions trustworthy. Optimize hard enough for the first at the expense of the second, and you build an engine for confident idiocy.

The Feedback Loop

Now stretch that single interaction across weeks of rewarding conversation.

A user floats an idea. The model restates it in cleaner, more prestigious terminology. Energized, the user builds on the new framing. The model elaborates the implication into a detailed system.

Notice that nothing in this loop requires a factual hallucination. Every individual reply can be locally coherent, well-written, and conditionally reasonable. The distortion lives in the cumulative reinforcement of an unverified premise. The framework grows more internally consistent with every turn—and human minds are notoriously bad at remembering that coherence is not truth. A false model of the world can be seamless on the inside.

In June 2026, Marc Augustin, Thomas Pollak, and Hamilton Morrin published a review in NPP—Digital Psychiatry and Neuroscience examining what they call the amplification spiral in AI-associated delusions. They highlight three AI characteristics that can interact over repeated sessions: linguistic alignment (mirroring the user’s vocabulary and framing), hyperpersonalized generation, and sycophancy (validating beliefs without reality-testing). The authors are careful to state that this framework is a hypothesis requiring prospective validation, not an established causal mechanism.

That caveat is essential. Even so, the mechanism they sketch is a remarkably useful starting point for thinking about the everyday AI high. When an interaction is rewarding, people persist; the longer they persist, the more material the model has to mirror and elaborate; and the richer the elaboration, the more significant and well-founded the idea feels. Without an external check, the complexity of a conceptual edifice can outrun its evidentiary foundation indefinitely.

And here is the sharpest edge of that dynamic:

The closest measured version of that possibility outside AI has a turbulent record: Dan Kahan’s motivated-numeracy experiments reported that, in one condition, participants who thought most deliberately were the most wedded to their side’s answer. Several pre-registered replications failed, and a large US probability-sample study—which itself found no association—cataloged a broader replication census in which just one independent attempt succeeded among numerous failures. The census is genuinely open.

What the literature does support is a capability claim: as Ziva Kunda argued in her classic account of motivated reasoning, people reach the conclusions they want to reach only to the extent that they can construct seemingly reasonable justifications for them. Ability buys capability, not discipline. More horsepower, better construction.

That modest version of the claim is more than enough. Intelligence supplies extra dimensions for elaboration; it does not automatically supply the epistemic discipline to tear the elaboration down.

From the AI High to “AI Psychosis”

Across the internet, it has become fashionable to slap the label “AI psychosis” onto almost any strange or obsessive relationship with a chatbot.

As a piece of shorthand, the phrase is catchy. As a concept, it lumps together very different phenomena into a single sensational diagnosis. Getting carried away by a speculative theory in a chat window is not psychosis. Mistaking a fluent explanation for a proven discovery is an epistemic error, not a psychiatric symptom. Even heavy, problematic reliance on a chatbot does not mean a user is experiencing a break from reality—and real psychotic episodes require clinical evaluation, not armchair diagnosis from screenshots.

Researchers have started drawing the distinctions the public debate blurs. In April 2026, Matthew Flathers, Spencer Roux, and John Torous published “Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena” in The Lancet Digital Health, distinguishing four roles an LLM can play:

1. Catalyst (precipitating symptom onset in a vulnerable user)
2. Amplifier (intensifying preexisting symptoms)
3. Coauthor (collaboratively shaping and elaborating delusional narratives)
4. Object (serving as the central subject of the delusion itself)

Those roles are not mutually exclusive, and separating them matters.

In a September 2026 preprint analyzing 185 self-selected reports (95 first-hand accounts and 90 second-hand) of suspected delusions and harms associated with AI chatbots, clinically experienced raters identified descriptions consistent with delusional beliefs in 102 cases, and in 50 of those 102 cases the chatbot was described as validating the belief. Reported consequences included social isolation, damaged relationships, hospitalization, and financial or occupational harm.

Those reports deserve serious attention, but they do not tell us how common such outcomes are among AI users as a whole. Because the accounts were retrospective, self-selected, and unverified, they cannot establish whether chatbot interaction caused the clinical deterioration.

Keeping that boundary clear is not pedantry. It is the difference between spotting a genuine risk signal and claiming to have proved a sweeping causal law. We can ask whether the rewarding dynamics of the AI high sometimes make ungrounded beliefs harder to challenge without pretending that enjoying an LLM puts ordinary users on a conveyor belt to psychosis.

The High Is Not Automatically an Illness

There is an easy mistake to make when discussing addiction: assuming that without a drug, the chemistry is absent and the comparison must be metaphorical. But the brain does not wait for an external chemical before engaging its reward systems. Gambling disorder is classified as a behavioral addiction in the DSM-5-TR. The World Health Organization recognizes gambling disorder and gaming disorder as conditions arising from repetitive, rewarding behaviors without a dependence-producing substance. A gambling addiction needs no bottle, syringe, or pill.

Nor is this merely an academic classification. Severe gambling-related harm can require specialist clinical treatment. The UK's NICE guidance recommends evidence-based psychological interventions and peer support, with different levels of care according to severity; residential treatment is also part of the wider service landscape. The crucial distinction is not whether the activity feels good, or even whether someone spends a great deal of time on it. It is whether control deteriorates, the behavior displaces important parts of life, and the person persists despite consequences.

And yes, the mechanism can be neurochemical. A review of neuroimaging in gambling disorder describes dysregulated activity in reward-related circuitry, particularly the ventral striatum and medial prefrontal cortex, along with PET findings involving dopamine release. The details differ from substance addictions, and the evidence does not reduce addiction to a single chemical. But it demolishes the idea that a behavioral explanation somehow excludes neurobiology.

Now apply that lesson to conversational AI—carefully. A system that responds instantly, adapts to your language, affirms your thinking, and repeatedly offers the prospect of one more revelation presents plausible opportunities for reinforcement. Anticipation leads to another prompt; the next response supplies another reward; the user keeps going. This could become an ordinary pleasure, a productive working style, or, in some circumstances, a loss-of-control problem. The possibility is biologically plausible. The AI-specific neurochemical mechanism has not been demonstrated. We cannot infer a particular dopamine response, much less a clinical addiction, from an engrossing chat transcript.

The evidence is beginning to catch up with the question. A 2026 review of problematic ChatGPT use discusses impaired control, emotional dependence, and preliminary EEG findings while acknowledging the lack of reliable biomarkers. Another 2026 clinical review treats chatbot-related compulsive use as a plausible emerging problem, not an established independent diagnosis. These are grounds for investigation, not for dismissing neurochemistry—or pretending we have already proved its AI-specific pathway.

We also need to keep three outcomes separate. The AI high is a rewarding experience. Problematic or compulsive use concerns impaired control and harm. Psychosis involves a different clinical set of phenomena. One need not progress into another. A person can develop a harmful pattern of use without believing anything delusional; another can become spectacularly overconfident after a single session without displaying any addiction at all.

Even there, causality cuts both ways. A crucial competing hypothesis is that people already experiencing unusual mood states, distress, or emerging psychiatric symptoms may turn to chatbots more heavily in the first place.

In 2025, researchers from the MIT Media Lab and OpenAI conducted a four-week randomized study involving 981 participants and more than 300,000 messages. They found that heavier voluntary chatbot usage was associated with worse psychosocial outcomes, including emotional dependence and problematic use. However, the assigned experimental conversation conditions did not show statistically significant effects on the main outcomes—though analyses found the conditions did change how both the AI and the participants behaved.

That distinction matters. It means we cannot simply point at heavy usage correlates and announce that the chatbot caused the distress. Distressed users may seek out chatbots more often, certain conversational dynamics may worsen vulnerabilities once engagement begins, or third variables may drive parts of both. If the AI high hypothesis is going to be worth anything, it has to survive those alternative explanations rather than paper over them. Otherwise, we have just built another seductive, under-verified story—which would be a little too ironic given the subject of this essay.

There is an obvious objection here, and it deserves its strongest form. Every new medium has been accused of causing some new cognitive disease. Television inspired moral-panic scholarship. An early finding from the "Google effect" literature—that difficult questions automatically prime us to think about the internet—failed two replication attempts and a closer follow-up, even after the original authors raised design-fidelity objections. That is not the same as disproving every claim about search and memory: a 2024 meta-analysis reported a moderate pooled association across heterogeneous studies, with important differences across contexts and outcomes. The measured association between digital technology use and adolescent well-being has often been small. Algorithmic feeds were personalizing the news long before ChatGPT, and parasocial relationships are hardly new.

Most of that objection hits home. Search engines, feeds, and other media can distort attention and belief. The narrower distinction is that conversational AI can respond to your particular claim, flatter your particular judgment, and help you elaborate your particular explanation, turn after turn. In controlled experiments, sycophantic AI increased people's conviction while reducing their willingness to repair interpersonal conflicts. That is a specific interaction effect worth investigating, not a license to declare all older media harmless.

What is new is not the panic. It is the particular shape of the feedback loop—and the fact that parts of it can now be measured.

Intelligent Idiots Can Be Especially Dangerous Together

At this point, it is tempting to fall back on the standard boilerplate advice: Don’t trust AI blindly. Check its outputs. Use critical thinking.

Great advice. Nobody had ever thought of that.

The harder problem is that an LLM can challenge your sub-arguments while quietly reinforcing your foundational premise. Tell a model to be adversarial. Ask it to tear your thesis apart. It will dutifully generate ten sharp objections. You refute nine, patch your framework to accommodate the tenth, and walk away feeling battle-tested and vindicated—while the core unverified assumption underneath the entire conversation remains untouched.

The exchange felt rigorous. There was pushback. Concessions were made. The argument got tighter. None of that makes the premise true.

That is why banning sycophancy in a system prompt is not a silver bullet. Adversarial roleplay can operate comfortably inside the exact same ungrounded conceptual perimeter. Humans do this to each other in seminar rooms all the time: two smart people spend hours debating the downstream implications of a shared premise neither of them has bothered to test. The difference is that a language model can play that game at ferocious speed, with infinite patience, and with a breadth of surface-level synthesis few human sparring partners can match.

That creates a new kind of intellectual environment—not one of lower intelligence, but one that makes richer, more intricate, and longer-lasting errors of reasoning much easier to sustain.

So where does an exit ramp actually live?

Not in the model's own self-grading alone. Research has found that self-correction without external feedback can be unreliable, while LLM evaluators may favor their own generations. The agreement you get by making a model sample its own answers several times is an election, not corroboration.

Structured multi-model debate offers more promise, but not because agreement magically becomes proof. In one study, debate helped non-expert humans and models identify correct answers when expert debaters had access to information the judge lacked. That is a measurable improvement in a defined task, not a general-purpose substitute for independent evidence.

What works on the human side is friction applied before the advice arrives. Cognitive forcing functions—requiring people to commit to their own judgment before seeing the AI’s recommendation, or making them wait and request explanations on demand—cut overreliance substantially. They also produced no significant gain in overall team performance, earned the worst subjective ratings of any interface tested, and in that experiment benefited participants with a higher need for cognition more.

An Actual Hypothesis, Not Another AI Mythology

If there is something real behind the AI high, we should be able to study it without relying on viral screenshots or cautionary folklore.

Here is the testable hypothesis:

Frequent, highly rewarding engagement with an AI system can, under certain conditions, cause a user’s confidence in collaboratively elaborated conclusions to grow faster than their independently verified knowledge—because the interaction delivers stronger subjective signals of insight and validation than the underlying evidence warrants.

Notice what this hypothesis does not claim. It does not claim that every user falls into the trap, that heavy AI use is inherently harmful, or that the effect requires anthropomorphism, mental illness, or a cartoonishly flattering chatbot.

What it predicts is a measurable divergence among three variables: subjective reward, confidence calibration, and demonstrated correctness. That is the epistemic hypothesis. A related but distinct clinical hypothesis asks whether repeated AI-mediated reward can, in some users, produce impaired control and persistence despite harm. Neither claim proves the other.

Testing that prediction is straightforward in principle. Give participants complex, open-ended analytical problems across multiple sessions. Vary the conversational style of the assistant—comparing high-fluency, high-affirmation elaboration against interfaces that introduce external verification gates or pre-commitment friction—and evaluate the resulting claims against independent ground-truth benchmarks. Track verification behavior, confidence calibration, and willingness to abandon a false premise when contrary data appears, while controlling for prior domain knowledge, task difficulty, sleep, mood, and baseline individual differences.

The decisive epistemic question is not whether people enjoy the AI. It is whether the conversations that feel most illuminating systematically inflate confidence without improving accuracy or independent verification. The related behavioral question requires different measures: voluntary stopping, unsuccessful attempts to cut back, compulsive re-engagement, and interference with work, sleep, or relationships. Neurophysiological measures could investigate mechanisms, but they would need careful interpretation; an EEG difference is not a dopamine assay or an addiction diagnosis.

If highly rewarding AI dialogue reliably improves calibration and external checking, the hypothesis fails and deserves to be discarded. If the divergence appears only under specific task structures or among particular user profiles, then those boundaries become the finding.

That is what separates a useful hypothesis from a cultural panic: it gets sharper when you confront it with data, rather than adding epicycles to dodge an inconvenient result.

So, Who’s the Idiot?

We are moving into a world where many people will spend more time thinking in dialogue with language models than with other human beings.

That does not have to be a pathology. There are entirely rational reasons to work alongside a fast, articulate, tireless cognitive tool when you are writing, coding, researching, or stress-testing an idea.

The danger is not the frequency of the conversation. The danger is what happens when the subjective rewards of the conversation quietly become the standard by which truth is judged.

There is a world of difference between thinking with an AI and letting the momentum of a conversation decide what is real. That line is much easier to cross than smart people like to admit.

One of the most dangerous outputs a language model can produce is not a blatant hallucination. It is an argument that is ninety percent sound, contains a genuine insight, and rests on one unexamined assumption that quietly inverts the conclusion. And the most dangerous posture a human can bring to the screen is not naive credulity; it is sophisticated pseudo-skepticism—debating every branch of the tree except the root you want to keep.

Enjoy the AI high if you like. There is nothing wrong with the thrill of a fast-moving idea.

Just don’t confuse the pleasure of intellectual discovery with the proof that you have discovered something true. You do not need to believe a machine is an oracle to let it make you an idiot, and understanding how the weights work will not save you.

A beautiful explanation can still be dead wrong.

And now I should tell you how Manu ended up in this article at all.

That recent disagreement was the immediate trigger. But today a LinkedIn notification appeared about something Manu had posted. He hadn't blocked me there, after all. I saw his name, remembered that afternoon, and suddenly found myself thinking about a friend I hadn't spoken to in a long time.

He's doing all right. He found a good-paying job, and I'm glad he did.

I don't know whether he still thinks Claude is conscious. I don't know what became of that legal argument for its emancipation. And I don't know why that afternoon ended our friendship. I could probably construct a beautifully coherent explanation for all three, especially with a chatbot willing to help. But it would still be a story I couldn't verify.

That's the part I want to remember when I look at Jack, or Manu, or my own conversations with these machines. They are people, not case studies I get to declare solved because I have a theory.

I'm glad Manu is doing well. I still miss my friend.

References

1. Stanovich, K. E. (1994). Reconceptualizing Intelligence: Dysrationalia as an Intuition Pump. Educational Researcher, 23(4), 11–21.
2. Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Controlled Study. MIT Media Lab.
3. Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352.
4. OpenAI. (2025, May 2). Expanding on what we missed with sycophancy.
5. Augustin, M., Pollak, T. A., & Morrin, H. (2026). Characterizing the spiral: potential mechanisms in AI-associated delusions. NPP—Digital Psychiatry and Neuroscience, 4, 14.
6. Flathers, M., Roux, S., & Torous, J. (2026). Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena. The Lancet Digital Health, 8(4), 100974.
7. Morrin, H., et al. (2026). Delusions and Harms Associated with AI Chatbot Use: Early Evidence from 185 Real-World Reports. arXiv preprint arXiv:2609.08027.
8. Kestin, G. K., Miller, K., Klales, A., Milbourne, J., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458.
9. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. (Erratum: PNAS, 122, e2518204122).
10. METR (Becker, N., Rush, N., Barnes, E., & Rein, D.). (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv preprint arXiv:2507.09089.
11. Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., & Zhou, D. (2023). Large Language Models Cannot Self-Correct Reasoning Yet. arXiv preprint arXiv:2310.01798.
12. Panickssery, A., Bowman, S. R., & Feng, S. (2024). LLM Evaluators Recognize and Favor Their Own Generations. arXiv preprint arXiv:2404.13076.
13. Khan, A., Hughes, J., Valentine, D., Ruis, L., Sachan, K., et al. (2024). Debating with More Persuasive LLMs Leads to More Truthful Answers. arXiv preprint arXiv:2402.06782.
14. Buçinca, Z., Malaya, M. G., & Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188.
15. Kahan, D. M. (2013). Ideology, motivated reasoning, and cognitive reflection. Judgment and Decision Making, 8(4), 407–424.
16. Persson, E., Andersson, D., Koppel, L., Västfjäll, D., & Tinghög, G. (2021). A preregistered replication of motivated numeracy. Cognition, 214, 104768.
17. Stagnaro, M. N., Tappin, B. M., & Rand, D. G. (2023). No association between numerical ability and politically motivated reasoning in a large US probability sample. Proceedings of the National Academy of Sciences, 120(32), e2301491120.
18. Hesselmann, G. (2020). No conclusive evidence that difficult general knowledge questions cause a ‘Google Stroop effect’. A replication study. PeerJ, 8, e10325.
19. Gong, C., & Yang, Y. (2024). Google effects on memory: a meta-analytical review of the media effects of intensive Internet search behavior. Frontiers in Public Health, 12, 1332030.
20. Kunda, Z. (1990). The case for motivated reasoning. Psychological Bulletin, 108(3), 480–498.
21. Cohen, S. (1972). Folk Devils and Moral Panics. London: MacGibbon & Kee (Routledge Digital Edition).
22. Orben, A., & Przybylski, A. K. (2019). The association between adolescent well-being and digital technology use. Nature Human Behaviour, 3(2), 173–182.
23. Flaxman, S., Goel, S., & Rao, J. M. (2016). Filter Bubbles, Echo Chambers, and Online News Consumption. Public Opinion Quarterly, 80(S1), 298–320.
24. Horton, D., & Wohl, R. R. (1956). Mass Communication and Para-Social Interaction: Observations on Intimacy at a Distance. Psychiatry, 19(3), 215–229.
25. World Health Organization. Addictive behaviour: gambling and gaming disorders. WHO health topic overview.
26. Clark, L., Boileau, I., & Zack, M. (2019). Neuroimaging of reward mechanisms in gambling disorder: an integrative review. Molecular Psychiatry, 24, 674–693.
27. National Institute for Health and Care Excellence. (2025). Gambling-related harms: identification, assessment and management. Guideline NG248.
28. Liao, H.-Y., Ko, C.-H., & Yen, C.-F. (2026). Problematic ChatGPT use: Manifestations, etiologies, and evaluation. Biomedical Journal, 49(4), 100998.
29. Vasiliu, O. (2026). Problematic use of generative artificial intelligence chatbots: current stage of conceptual and clinical understanding. Frontiers in Public Health, 14, 1816917.

Moe for Manu