Epistemic Trespassing: When Experts Wander Beyond Their Domain

The Man Who Invented the Test for HIV Did Not Believe HIV Caused AIDS

In 1993, Kary Mullis won the Nobel Prize in Chemistry for inventing the polymerase chain reaction — the technique that allows scientists to multiply a tiny amount of DNA into a quantity large enough to study. PCR is the foundation of nearly every modern diagnostic test. It is the technology used to test for HIV. It is the technology used in every COVID test you have ever taken. It is, by any reasonable measure, one of the most consequential inventions in twentieth-century biochemistry.

Kary Mullis also believed that HIV does not cause AIDS, that human-caused climate change is a hoax, that astrology is real, and that he was once visited by a fluorescent talking raccoon in the woods outside his cabin in Mendocino. He wrote about these positions extensively in his 1998 autobiography Dancing Naked in the Mind Field. He continued to deny HIV-AIDS causation until his death in 2019, well after the science was settled by every measure his own invention had helped establish.

The puzzle is not “how could a brilliant man hold such terrible views.” Brilliant people hold terrible views all the time. The puzzle is that Mullis’s competence in the specific cognitive operations required to discover PCR — careful experimental design, statistical reasoning, biochemical intuition — did not transfer one inch into evaluating the HIV-AIDS hypothesis, which required essentially the same skills. Mullis was a world-class expert when working on a particular bench problem and a confidently-wrong amateur the moment he stepped outside that bench. The boundary between the two was invisible to him. The philosophical literature has a name for this boundary, and a name for what happens when an expert crosses it: epistemic trespassing.

The term was crystallized by the philosopher Nathan Ballantyne in 2019, but the phenomenon has been studied informally for centuries. It is the cognitive failure mode that explains why Linus Pauling — two-time Nobel laureate — spent the last thirty years of his life trying to convince physicians that 40,000 mg of vitamin C a day could cure cancer. (Pauling died of prostate cancer in 1994, the very disease he claimed vitamin C prevented.) It is why physicists routinely embarrass themselves about consciousness, why economists pronounce on epidemiology, why football coaches give business advice. And in 2024 it became something else entirely: an industrial-scale failure mode of AI systems that cost the global economy an estimated $67.4 billion in business losses from AI hallucinations alone.

To see why epistemic trespassing matters now in a way it did not before, you have to look at it across three domains at once.

The Domain of Brilliant Wrongness

The cleanest case studies of epistemic trespassing come from the Nobel prize itself, where the prize sometimes seems to function as a license to be confidently wrong about everything else. The Skeptical Inquirer in May 2020 coined the term “Nobel disease” to describe the pattern: an unusually high rate of Nobel laureates — particularly in physics and chemistry — embracing implausible ideas in fields outside their training. Pauling and Mullis are the canonical examples; the longer list runs through chemists pushing megavitamin cures, physicists championing telepathy, and biologists promoting racial pseudoscience.

The cognitive mechanism is the Dunning-Kruger effect generalized one level up. The classic Dunning-Kruger finding is that people who are bad at a task are unable to recognize how bad they are, because the metacognitive skill needed to evaluate competence at the task is the same skill needed to do the task in the first place. Ballantyne and Dunning collaborated on a 2018 paper extending this: experts in domain A possess enough metacognitive skill to recognize their own expertise in domain A, but that recognition does not transfer to domain B. They know enough to feel calibrated. They do not know enough to be calibrated. The two feelings are indistinguishable from the inside.

This is the destructive end of the spectrum. Pauling killed people — not directly, but by lending his Nobel-laden credibility to a treatment regimen that diverted cancer patients from effective therapy. Mullis muddied the public discourse on HIV during the first decade of an epidemic that killed forty million people worldwide. The damage from confidently-wrong experts is measurable in lives. And the people doing the damage genuinely could not see what they were doing wrong, because every internal feeling of competence and calibration that had been correct in their home domain was now firing in a domain where it carried no information.

If this were the whole story, the answer would be obvious: stay in your lane. But it is not the whole story. There is a second, opposing pattern that has been just as consistent across history.

The Domain of Productive Trespassing

When Karl Marx was thinking about the industrial revolution in the 1860s, he noticed something that historians of technology still cite. The men who invented the steam engine, the spinning machine, and the steamship were not engineers. James Watt was a watchmaker. Richard Arkwright was a barber. Robert Fulton was a jewelry maker. None of them held credentials in the domains they revolutionized. None would have passed a guild examination in mechanical engineering. Marx’s point, written into Capital, was that the great inventions of his era came overwhelmingly from cobbler-class people who walked into other people’s domains without permission and built things the credentialed experts could not.

The pattern persists. The transformer architecture that powers ChatGPT was published in 2017 by a Google translation team trying to solve neural machine translation; the authors had no particular standing in general AI research. mRNA vaccines were championed for thirty years by Katalin Karikó, a biochemist denied promotion at Penn because her work was outside the accepted research priorities of her department. The CRISPR-Cas9 revolution came from microbiologists studying bacterial immunity, not from the human-genetics establishment that would eventually deploy it. Every one of these is, in Ballantyne’s strict definition, an act of epistemic trespassing — experts in one domain pronouncing on, and ultimately transforming, another. And every one was right.

A 2023 paper in the philosophy journal Inquiry explicitly pushed back on the dominant negative framing. The authors argued that “significant discoveries in scientific history would not have been possible without epistemic trespassers.” Their case was that the strict Ballantyne position — defer to specialists, mind the guardrails — would have prevented most of the breakthroughs we now consider obvious. Outsiders bring tools, frames, and questions that insiders cannot generate from within. The history of science is the history of fields being saved from themselves by people who shouldn’t have been there.

This is not a contradiction with the Pauling story. It is the same phenomenon viewed from a different angle. Epistemic trespassing is not categorically destructive or categorically productive. It is a high-variance activity. When it goes wrong, it goes wrong loudly — Pauling, Mullis, the long Nobel-disease list. When it goes right, it goes right loudly — Watt, Karikó, the transformer team. What distinguishes the two outcomes is not whether the trespasser was an expert in their home domain (they always are) or whether their target domain was complicated (it always is). It is whether the trespasser brought the epistemic infrastructure of their home domain with them — calibration, verification, willingness to be wrong — or whether they brought only the credentials of their home domain and expected those credentials to carry the work.

Mullis brought his credentials. Karikó brought her infrastructure. The two outcomes were not random; they were a function of what the trespasser was willing to do with disagreement once they crossed the boundary.

The Domain of Industrial-Scale Trespassing

The reason any of this matters in 2026 — to developers, to founders, to anyone working with AI — is that we have just industrialized epistemic trespassing. A large language model is, in the strictest possible sense, an expert in one specific thing: producing sequences of tokens that conform to the statistical patterns of natural language. It is genuinely world-class at this. It is also, by construction, a trespasser in every other domain. When you ask it a question about medicine, it produces grammatically perfect text shaped like a medical answer. The text is not the answer. The text is what an answer would look like if produced by something that knew the answer. Most of the time the two are the same. Some of the time they are catastrophically different.

The numbers tell the story. AI hallucinations cost an estimated $67.4 billion globally in 2024. AI systems in healthcare make hallucination mistakes 8% to 20% of the time depending on task. Europe’s cybersecurity agency ENISA had to retract two threat reports in 2025 after discovering they were riddled with AI-fabricated citations — 26 incorrect footnotes out of 492. Deloitte delivered a report to the Australian government containing fake academic references and made-up quotes that no human had checked. Attorneys in multiple jurisdictions have been sanctioned for filing court briefs citing court decisions that do not exist. Each of these is a case of epistemic trespassing exactly in Ballantyne’s sense — a system that is a true expert in one narrow thing (text generation) pronouncing confidently on domains where it has no calibration.

The qualitative difference from the Pauling case is that AI trespassing is universal and automated. Pauling was one wrong man with a strong reputation in one strong field; he could trespass into a few adjacent fields per decade. A frontier language model can trespass into every field humans have ever written about, at a rate of millions of queries per hour, with the confident textual tone of a domain expert in each. The cost scales accordingly. Every chatbot answer about a drug interaction is an act of epistemic trespassing. Every AI-generated investment thesis is. Every legal brief, every code review of unfamiliar code, every “explain this to me” response about an unfamiliar topic. The model has no internal signal distinguishing the questions where it knows from the questions where it generates. From the inside it all feels like generating.

This is why “AI hallucination” is a misnomer. The model is not hallucinating in the sense of malfunctioning. It is doing exactly what it does — producing plausible text — and the failure is that we asked it to produce text in domains where plausibility and truth diverge. It is Pauling but for every field at once.

The Solution Is Not “Stay in Your Lane”

The naive lesson from this story would be: experts should stay in their domains, and AI systems should refuse to answer outside theirs. But Marx’s cobblers are the counterargument. The strict-deference position would have prevented Watt, Karikó, the transformer team, and almost every interdisciplinary breakthrough of the modern era. It would also impoverish the use of AI to the point of uselessness — a model that refuses to engage with anything outside its narrow training distribution is a model that cannot help with any real-world task, because every real-world task is interdisciplinary.

The better framing comes from the productive trespassing pattern: bring your epistemic infrastructure with you, not just your credentials. Karikó brought rigorous experimental skepticism into immunology from biochemistry. Watt brought the watchmaker’s discipline of micron-scale tolerance into the world of steam. The transformer team brought a translation team’s habit of careful ablation studies into a problem they hadn’t been formally trained on. Each of them won by importing the method of their home domain, not by relying on its reputation.

For working professionals, this generalizes into three habits worth borrowing.

First, when crossing domains, be more skeptical of yourself, not less. The cognitive feeling of competence is calibrated to your home domain. It is not a signal in the new one. Discount it. Pauling died of cancer because the internal feeling of “I’m a Nobel laureate, I must be right about this” was indistinguishable from the internal feeling that had been correct about quantum chemistry for forty years.

Second, bring the verification infrastructure of your home domain into the new one. If you are a scientist, run an experiment. If you are an engineer, build a test. If you are a writer, check the citation. The cobblers who invented machines did not trust their cobbler intuition; they built prototypes and iterated. The Nobel laureates who embarrassed themselves did not bring their bench discipline into their public pronouncements.

Third, when working with AI, treat every output the way you would treat a confident pronouncement from a brilliant stranger in a domain you do not know. The stranger sounds correct. The stranger uses the right vocabulary. The stranger may even be correct most of the time. None of those facts mean the stranger is correct on any specific point. Verify the load-bearing claims. Check the citations. Re-derive the calculations. Treat AI outputs the way a careful researcher treats a paper that has not yet been peer-reviewed — useful, suggestive, and provisional until checked.

The meta-point is that this essay itself is an act of epistemic trespassing. The author is not a credentialed philosopher of science, not a historian of technology, not a clinician evaluating Pauling’s vitamin claims, not an AI safety researcher quantifying hallucination rates. The defense of writing it anyway is the same defense Marx gave for the cobblers: the synthesis across these domains is the work that the domain experts will not do, because each domain expert is busy doing the depth-work that justifies their credential. The risk of writing it is the same risk Pauling and Mullis ran: producing a confident text that sounds correct in each domain it touches while quietly being wrong in some specific way that none of the domain experts have time to flag.

The mitigation is the only honest one available. Every claim in this essay traces to a specific cited source. Every number is the number a specialist would recognize. Every inference is one that a careful reader can check or push back on. The cobbler is doing the work; the cobbler is also showing their work. That is the difference between Karikó and Pauling, between Watt and the next confident amateur with a perpetual-motion patent. Bring the tools. Show the work. Cross the boundary anyway, because the most important problems live at boundaries — but cross with the cobbler’s discipline, not with the laureate’s credentials.

The line between trespassing that pays $40,000-mg-of-vitamin-C in damage and trespassing that pays steam engines and mRNA vaccines is not visible from the inside. It is only visible in what the trespasser is willing to do with the disagreement they will inevitably encounter. Stay open to being wrong, and trespassing is the engine of intellectual progress. Stay confident in being right, and you will eventually be the next entry in the Nobel disease catalog. The boundary is the same. The choice is which side of it you spend the rest of your career on.

Sources: Ballantyne, “Epistemic Trespassing,” Mind 128(510): 367–395, 2019; Kary Mullis, Dancing Naked in the Mind Field (1998); Linus Pauling, Vitamin C and the Common Cold (1970) and How to Live Longer and Feel Better (1986); The Skeptical Inquirer, May 2020 (“Nobel disease” coinage); Ballantyne & Dunning, on metacognitive competence transfer, 2018; Pliny the Elder, Naturalis Historia 35.85 (“ne sutor ultra crepidam”); Marx, Capital Vol. 1 (1867), on cobbler-class inventors; Vaswani et al., “Attention Is All You Need,” NeurIPS 2017; Daubert v. Merrell Dow Pharmaceuticals, 509 U.S. 579 (1993); Federal Rule of Evidence 702; Kruger & Dunning, “Unskilled and Unaware of It,” Journal of Personality and Social Psychology 77(6), 1999; Thorndike, “A Constant Error in Psychological Ratings,” Journal of Applied Psychology 4(1), 1920 (halo effect); Inquiry, 2023 (counter-argument paper on productive trespassing); ENISA threat-report retractions, 2025; Deloitte Australia report incident, 2024–2025; AI hallucination cost estimate $67.4B (2024 industry figures, multiple sources).