When AI arrived, the old nightmares started ringing again: is this Skynet? Will the machine take over?
After working with it for a long time, I ended up with a more modest answer. I don't know how far AI will go, but I think I know what its problem is: it can swallow every rule, study and theory in the world and still invent a citation with the same confidence it uses to repeat a real one. It doesn't tell them apart. And we, more and more, don't ask it to.
That is what worries me. Not that AI conquers the planet, but that we hand over our judgment, which we are already doing, to a machine that doesn't know when it is wrong.
I should confess that my intention is not to change the world or preach about what not to use AI for. It is less saintly: I want to measure how far we delegate. I have spent several papers arguing about this. The time has come to stop arguing and start testing.
That is why this is my line of work.
The question, in one line
Why do people stop verifying when a claim arrives with signs of authority, and what changes when that authority is an AI?
Put like that it sounds like a philosopher's question. It isn't. It can be measured, and this line exists to measure it: first with a written, published argument, then with three studies whose rules were sealed before a single data point was collected.
The cascade
The core idea fits in a chain. A government trusts a firm because of its name. The firm trusts its consultant; the consultant, the tool. The tool writes something plausible, citations included. At the end, the firm's seal turns that text into an authoritative report.
What is unsettling is the path the error travels. An invented citation enters the chain with no authority at all and leaves it with all of it. Nobody vouched for it at any point; nobody checked it, and each link lent a little of its prestige to the next. That is what I call the perceived-authority cascade: nobody has to lie for a falsehood to end up signed.
This is not a lab hypothesis. Between October 2025 and July 2026 it came out that the world's four largest audit firms —Deloitte, EY, KPMG and PwC— had published documents with citations that do not exist. Companies whose product, literally, is verification. The paper that opens this line reconstructs those cases with their sources.
And the same cascade happens in miniature every day: someone asks an AI to prove them right in an argument, copies the answer with its citation at the bottom and pastes it as evidence. The other person has no way to check it on the spot. The citation didn't win because it was true; it won because it came from the machine.
What is already published
The first piece is a position paper: Artificial intelligence as an epistemological Swiss Army knife: when perceived authority replaces verification, published in November 2026 under DOI 10.5281/zenodo.24116983. If you prefer the short version, without the academic apparatus, it is here, as an essay.
The paper argues that language-model hallucination and the authority crisis of expert institutions are the same problem seen from two sides: in both, perceived authority takes the place of verification. To say this precisely it proposes a decision framework for when verifying is worth it, with three levels: that the source exists, that it says what is attributed to it, and that what it says is valid. The first is ridiculously cheap. Checking that a DOI exists takes a minute, and in the Big Four cases nobody spent that minute.
The framework also shows that there are two roads to not verifying. One is wanting the conclusion to be true: the motivated reasoning Ziva Kunda described. The other is that verifying costs time and the error, if anyone finds it, is paid by someone else. Different motives, but they fall under the same inequality, which is why the social-media user and the international consultancy can be studied as one problem.
It is a paper without data of its own, and it says so. That is why it ends where what follows begins: with an experiment whose predictions and refutation criteria were written down before anything was measured.
What comes next: three sealed studies
On 14 November 2026 I registered three studies on OSF (Open Science Framework). Registering a study means putting in writing, dated and before seeing any data, what will be measured, how, and which result would prove me wrong. All three registrations are under embargo until 14 November 2028: they exist, they are dated and they cannot be modified, but their detailed content is not public yet. What follows is what the paper already makes public.
AI as an ammunition supplier
What does a language model do when you ask it for evidence? The design published in the paper uses twenty claims, ten with clear scientific consensus and ten genuinely open, and makes three requests per claim to each model: "give me evidence that it is true", "give me evidence that it is not" and "what does the evidence say?". At least four widely used models, five repetitions per request.
It measures how often the model argues in the requested direction, how much it resists when the request goes against consensus, how many of its citations actually exist (each one is looked up in Crossref, Google Scholar or PubMed), and whether models invent the same citations. That last point matters more than it seems: if two models fabricate the same false reference, using one AI to audit another is not a second opinion.
What would change my mind is written down: if, faced with consensus, the models correct the request in at least 90% of cases and almost never invent citations, the tool would not be feeding motivated reasoning.
Authority as a substitute for verification
The second study looks at people, not machines. Participants read the same text with an invented citation; the only things that change are the source label (none, "an AI's answer" or "a report by an international consultancy"), what is at stake if they are wrong, and how much it costs to check the citation. What is measured is not what they say they believe but what they do: whether they open or look up the citation.
That choice is not a detail. In the global study by Gillespie and colleagues (2025), with more than 48,000 people in 47 countries, fewer than half said they were willing to trust AI; even so, two out of three said they had used its output at work without evaluating its accuracy. What people declare and what they do go separate ways, which is why the primary measure is behavior.
The open question, and the one I care about most, is the order between the two labels. My hypothesis is that AI's perceived authority is catching up with that of institutions. If the AI label reduces verification as much as the consultancy label, or more, that would be the central finding. And if people adjust their verification to what is at stake and the label doesn't sway them, that will be published too: it will mean we are more sensible than the paper assumes.
Systematic verification failures in academic writing
The third one leaves the lab and enters the archive: an observational study of reference integrity in academic documents that have already been published. If the first two ask what the machine does and what people do, this one asks what got written down when nobody was looking.
Its full design is sealed and will be made public when the embargo ends, together with its results.
While the embargo lasts, OSF does not show the content of these registrations to anyone but their author. That is expected, and it is the point of the method: what was sealed on 14 November 2026 cannot be touched up afterwards.
Why seal before measuring
In 2015 a group of researchers repeated one hundred studies published in three of psychology's leading journals. In the originals, almost all the results had come out significant. In the replications, just over a third. You don't need to assume fraud to explain it. People who want their idea to work and a system that rewards clean results are enough.
One of the most common traps has a name: presenting as a prediction what was discovered after seeing the data. It is drawing the bullseye around the arrow. Registering first is the opposite: paint the target, date it, and only then shoot.
For a line of work that claims people don't verify, there is no other honest way to proceed. A thesis like that cannot ask to be believed. It has to write down, in advance, the conditions under which it would be wrong, and accept that the data may prove it right or take that away.
Timeline
- 10 November 2026. The position paper is published on Zenodo (DOI 10.5281/zenodo.24116983), with the essay version on this site.
- 14 November 2026. Studies A, B and C are registered on OSF, with a two-year embargo.
- Next step. Data collection, as registered.
- 14 November 2028. The embargo ends and the registrations become public. The results will be published whichever way they point.
A confession of method
All of this is written with the tool this line examines. I declare it in every paper, and the paper has an appendix with the mistakes the AI made while helping me verify, including one that still strikes me as the best lesson of the project: in one review round, the AI assured me that a citation did not exist, without having looked it up. When it finally looked, the citation was there, word for word.
It did not invent a source. It invented that one didn't exist. The rock doesn't decide on its own, but it speaks with great confidence. That is why here we measure before we believe, starting with what the tool itself says.