"AI expert": anatomy of a smokescreen

A man seen from behind, at his desk, looks at a corkboard where four handwritten cards each propose a job title; three are struck through in red marker, the fourth is circled.
Illustration generated with Gemini (Nano Banana), retouched in GIMP.
Detailed description of the image

Indoor photograph in landscape format, taken at the end of the day.

On the left, a man in his forties sits with his back to us at a wooden desk, turned to the right, chin resting on his hand. He wears a dark chunky-knit jumper and his brown hair is tousled. Only the rear three-quarters of his face is visible, so his expression cannot be read. Behind him a window lets in warm, low light over a pale sky and city rooftops. On the desk, from left to right: a pencil pot, an open notebook covered in handwriting, a pair of glasses, a grey mug, an open laptop seen from behind, and a stack of books with colourful spines.

On the right, filling the upper half of the image, a large wood-framed corkboard hangs on the wall. Four white sheets are pinned to it in a square, each bearing a handwritten label, each pin a different colour. Top left, «naturopathe» (naturopath), struck through with a red marker. Top right, «lead thinker», struck through the same way. Bottom left, «formateur en bureautique» (office software trainer), on two lines, also struck through. Bottom right, «chercheur (indépendant) en IA» (independent AI researcher), on two lines, the only one not struck through: it is ringed with a wide red circle.

Between the man and the board, in the half-light, a bookcase, a plant and a few smaller photographs pinned to the wall can be made out. An angled desk lamp comes down from the top right edge.

There is a great deal of talk about disinformation by AI: generated texts, faked images, invented quotations. It is real, but it is not the subject of this post. I want to talk about a quieter and, to my mind, more corrosive kind of disinformation: disinformation about AI, produced by people who present themselves as experts in the field and are not.

The mechanism is simple. The general public does not know what a language model is, does not know what a researcher does with their days, and has no way of telling a scientific paper from a LinkedIn post. Into that vacuum, anyone can plant a flag. And plenty have.

A title nobody protects

In France, certain titles are restricted. "Physician", "lawyer", "architect": using them without entitlement is the offence of usurpation of title. "Doctor", in the sense of the academic degree, is protected too: you do not call yourself a doctor without having defended a thesis.

"Researcher", on the other hand, is protected by nothing. Nor is "expert". Nor "specialist". Nor "consultant". You can call yourself an AI researcher on Monday and a digital transformation coach on Tuesday, and nobody has any recourse. Even "engineer" is not restricted in France; only ingénieur diplômé, the accredited engineering degree, is, through the Commission des titres d'ingénieur.

The result is mechanical: the word "researcher" carries a century's worth of credibility accumulated by people who did in fact spend years producing verified knowledge, and that capital is on self-service. All you need is to add "in AI" to whatever you were already doing. An office-software trainer becomes an AI trainer. An organisational consultant becomes a generative AI consultant. A journalist who has tried ChatGPT three times becomes a specialist in the stakes of AI. Poof.

Why "in AI" works so well

The suffix "in AI" has three properties that make it an ideal smokescreen.

It cannot be checked by a layperson. If someone calls themselves an expert in tax law, you can check whether they are a lawyer, a chartered accountant, or nothing at all. If someone calls themselves an AI expert, what do you check? There is no professional body, no reference qualification the public knows about, no register. The only legible criterion would be scientific publication, and the public does not know where to look.

It frightens and fascinates at the same time. A field perceived as mysterious and threatening creates a massive demand for explanations. That demand can pay: anxious companies, local authorities that want to "do something", overwhelmed schools. Supply follows, and it has no reason to be proportionate to actual competence.

It licenses you to speak about everything. Since AI touches everything, the "AI expert" can pronounce on employment, education, health, geopolitics, consciousness, energy. A real AI researcher knows they are competent on a tiny fraction of the field. A fake one has no such constraint. The drift is remarkable for its speed: the office-software trainer turned "AI trainer" finds themselves, in the same week, explaining at a conference why quantum computing is going to make models conscious. Nobody asked them whether they knew what a qubit is, because nobody in the room knows either. The suffix "in AI" does not confer a competence, it confers a licence: the licence to speak about everything complicated, since AI is complicated and they are "in it".

What this smokescreen produces

This is not a problem of vanity. The consequences are concrete.

False figures circulate with the authority of an "expert". A few, heard at conferences, read in the press or in training material:

None of these figures came out of a laboratory. All of them came out of a chain of re-quotations in which each link cited the previous one, and in which the first link was already a misreading. What keeps them alive is that they are repeated by people whose standing as "experts" excuses the audience from checking.

I have also seen AI-text detection tools sold to entire institutions. The available studies credit them with high false-positive rates, with a documented bias against non-native writers. I am not a lawyer, and whether such a score is legally admissible depends on the procedure. But evidence of that quality is not enough to ground an accusation of cheating against a student, and that this is nonetheless what it is used for.

Decisions are taken on these bases. Training courses billed at top rates, institutional policies, press articles, contributions to public bodies. The cost of disinformation is not abstract: it is budgetary, pedagogical, sometimes legal.

And the worst effect is indirect: the credibility of research is diluted. When everyone is an expert, nobody is, and the researcher who says "we don't know" or "it's more complicated than that" comes across as less competent than the consultant who has an answer for everything.

The market in "made by AI"

The smokescreen is not only a problem of discourse. It has a turnover.

First there are the services. I have seen marketing platforms bill for content "generated by AI", "optimised by AI", "personalised by AI", whose entire production chain consists of pasting the client's brief into a consumer chatbot and then pasting the answer into the deliverable. Zero engineering, zero model, zero processing. The client pays a margin on a twenty-euro-a-month subscription they could take out themselves. The word "AI" on the invoice does not denote a technology, it denotes an opacity: the client does not know what they are buying, and that is precisely what makes it sellable to them.

Then there is training. Three hundred euros for a half-day to learn to "master generative AI", and the content fits in one sentence: open ChatGPT, write what you want, read the answer. The rest is padding: a history of AI in ten slides, approximate definitions, and a list of "magic prompts" that are politeness formulas. The person running it discovered the tool six months before their trainees. There is neither pedagogy nor content here, only the conversion of anxiety into a billable service.

And then there is employment. Job descriptions and profile titles are multiplying in organisations that have no idea what the post should contain. Half an hour on a professional network is enough to gather an armful; in brackets, the number of profiles carrying the same title:

I will pass over the truncation. "AI researcher in the field of X" miraculously ends up shortened to "AI researcher", and what drops off along the way is the only part that told you anything. Not bad!

These titles do not all denote impostures, far from it. Some cover perfectly real work: a doctoral student in artificial intelligence is a doctoral student, with a supervisor and a topic; an engineer who builds agent-based systems is doing a job that exists. That is exactly what makes the list interesting. Placed end to end, it is illegible. Nothing in the wording lets you distinguish the person who writes code from the person who runs workshops, or the researcher from the salesperson. "AI and Cinema Expert" and "AI Researcher" look alike typographically, and that is all a hurried recruiter, a journalist looking for a speaker or a local authority looking for advice will see.

In these titles, then, the word "AI" no longer describes a field: it signals membership. And it ends up producing internal authorities by decree. Someone who knows how to use a chatbot is appointed head of AI, and that person becomes the in-house expert on a subject they understand no better than their colleagues, simply because somebody had to be.

Not every service around AI is a swindle, nor every training course. The market, for its part, makes no difference between those that bring something and those that bring nothing, because the client has no way of seeing it. That is the very definition of a market in smoke.

How to check, without being in the trade

We are not asking the public to become researchers. We are asking them to know how to ask four questions.

Where do they publish? A researcher publishes in peer-reviewed conferences and journals. Their work is on HAL (the French open archive), on DBLP for computer science, on Google Scholar. If the whole of a person's output consists of LinkedIn posts, videos and a book with a general publisher, they are not a researcher. That is not disqualifying in itself, but it is not the same thing.

A caveat, though: "published" is not enough. There is a whole industry of so-called predatory publishers whose business model is to charge the author a few hundred or a few thousand euros to publish more or less anything, with a façade of a review board and peer review in a matter of days. Their journals have perfectly respectable names, "International Journal of", "Advances in", and a website that looks like a serious publisher's. To an unwary reader, an article there looks as credible as one in a genuinely peer-reviewed journal. That is precisely their stock in trade: selling the appearance of science to those who need it for their CV. A few clues: the publisher runs dozens of journals on every subject, solicits authors by email, promises fast publication, and advertises its publication fees as its main argument. When in doubt, look up the journal's name in the blacklists maintained by university libraries or in the DOAJ, which lists open-access journals meeting criteria of seriousness. A talk at a seminar or a workshop, however prestigious the venue, is not a reviewed publication either: it is an oral contribution, and it proves nothing except that someone was invited to speak.

Affiliated to what? A researcher belongs to an identifiable laboratory: a joint research unit, a team, an institute. It can be checked in two clicks: the lab has a website, a list of members, and the name is on it or it is not. Beware of formulations that imitate affiliation without being one. A post, even an academic one, is not membership of a laboratory: you can work in higher education, in the civil service or in a large company without belonging to any research team, and that is the case for the vast majority of people who work there. Being a doctoral student is a real affiliation, and a respectable one, but it designates someone in training for research, not a researcher; an honest doctoral student presents themselves as a doctoral student. "Independent researcher" exists but remains rare, and a serious independent researcher still has a publication trail; without one, the phrase mainly means that nobody agreed to take them on. "Founder of", "CEO of", "director of innovation at" are functions, not scientific affiliations: they say that the person has a commercial interest in the field, which is precisely the next question.

Are they selling something? It is not disqualifying, but it is a conflict of interest worth knowing about, and you have to go looking for it because it is rarely disclosed. The person who explains that AI is going to upend everything and who sells AI training has a direct interest in your being afraid. The one who explains that AI-text detectors are reliable and who distributes one has a direct interest in your buying them. The one who explains that AI is going to destroy half of all jobs and who sells "transformation" consulting has a direct interest in your feeling behind. The pattern is always the same: an alarming diagnosis whose only way out is the diagnostician's product. Look at the website footer, the "services" page, the link in the bio. Look also at who is paying for the conference: an "expert" invited by a software vendor will rarely speak ill of the software. A researcher may have commercial interests, and many do; the difference is that they are required to declare them in their publications, and that they do not found their authority on them.

Can they say what does not work? This is the most reliable test, and it requires no technical knowledge at all. Ask the person about the limits of what they are talking about, what we do not know, where results are contested, what has surprised them in a bad way. A real specialist answers at length and with pleasure, because the limits are precisely where they work: that is where they spend their days, that is where the open questions are, and that is what they want to talk about. They will tell you that such and such a result has never been replicated, that such and such a benchmark is contaminated, that such and such an effect disappears when you change the wording. A smoke merchant answers in generalities ("we must remain cautious", "it all depends on the use", "the human must remain at the centre") or changes the subject to what does work. Let me be precise, because I have written elsewhere that AI is a tool and I stand by it: the sentence is true, but it only becomes an evasion when served as a conclusion, in place of the list of limits you were asking for. A useful variant: ask them what they are not competent to talk about. The specialist answers in a second, with a precise list. The self-proclaimed expert hesitates, because the question makes no sense to them: their title covers everything.

When the smoke gets into the papers

There is one more storey, and it is the most worrying for someone whose profession this is: the contamination of the scientific literature itself.

The phenomenon is documented and easy to observe. Published papers contain sentences like "as a language model, I cannot", left in the text because nobody, neither author nor reviewer, read it. In 2024, a biology paper published in a peer-reviewed journal was retracted after publication because it contained manifestly aberrant generated figures, which thousands of readers saw before the journal reacted. American lawyers have been sanctioned for filing briefs citing court decisions that do not exist. And several studies have shown that models fabricate plausible bibliographic references: real authors, real journal, credible title, and a DOI that leads nowhere.

It is this last point that damages research most, because it is invisible. A reader rarely checks every reference. A bibliography of forty entries, three of which do not exist, reads exactly like a bibliography of forty entries. If the article is itself cited, the false reference propagates, and it becomes very hard to trace back to the origin. This is disinformation by slow accumulation, without intent and without a clearly identifiable author.

That researchers use these tools poses no problem for me: I use them, to rephrase, to translate, to clear a path through a literature I do not know. The problem is not the tool, it is the disappearance of a step. Signing a paper means answering for every sentence and every reference in it. A text you have not read line by line, whose sources you have not opened one by one, is not a text you answer for. The rule is simple and long predates language models: you do not cite what you have not read. It is enough to avoid more or less all of these accidents, and it is massively broken.

There is an irony worth noting. Institutions that buy detectors to hunt students most often have no procedure for checking the publications of their own staff. We police downwards and trust upwards. And the prestige of the title, once again, stands in for review: a text signed "AI researcher" is read with less suspicion than a student's paper.

The only honest title I found

None of the titles in my harvest says what the person does with their days. They all say the same thing, which is: I am on the AI side, and it is a good place to be.

One stood out: "Developer and user of chatbots and multimedia".

Look at what that title does. It names a concrete object, chatbots, not a continent. It distinguishes two roles, developing and using, which are not the same skill and which all the others conflate. It announces neither expertise, nor research, nor vision. You know immediately what to ask them, and what not to ask them. That is very precisely what I call a scope.

They do not have many followers. Of course. The honest title pays nothing: it says one thing, whereas the vague title promises a thousand. That is the entire logic of this post in a single comparison. As long as saying what you actually do costs visibility, smoke will remain the rational choice, and those who refuse to produce any will remain less audible than those who live off it.

A pity. Theirs is nonetheless the only profile in my harvest that I would have invited to speak in front of my students.

Two textbook cases

The mechanism was not born with generative AI; AI is merely its current vehicle. The most documented case in France is that of Idriss Aberkane, and it is worth applying the four questions to it, because he fails all of them methodically.

Presented in the mid-2010s as the holder of three doctorates, a professor at CentraleSupélec, a researcher at Polytechnique, affiliated with Stanford, he sold a best-seller on neuroscience and worked the television studios. In the autumn of 2016, Le Monde, L'Express, Libération and Marianne checked with the institutions: neither an École normale supérieure alumnus, nor an academic at the CNRS or at Polytechnique. He had been, for one year, an associate researcher, that is, an unpaid one, in a management research centre. His French doctorate is in management science and applied epistemology, not in neuroscience. In September 2022, Polytechnique's ethics committee described passages of that thesis as obvious plagiarism; the decision only became public in the summer of 2023. On publications, the question I raised above is the right one: where one expects a bibliography, one finds popular books and a handful of texts in journals whose seriousness has been contested.

Affiliation: real affiliations inflated until they became false. Publications: absent or complaisant. Commercial interest: books, lectures, training courses: the whole model rests on the title. Limits: a discourse covering neuroscience, economics, biomimicry, geopolitics and then, in the 2020s, virology, without ever declaring himself incompetent on anything.

What makes the case instructive is that it cost nothing. The investigations were published, the CV was publicly dismantled, and the speaking career carried on. The audience that fills the halls has not read L'Express; it has read "three doctorates". The correction never catches up with the title, because the title is short and the correction is long.

The second case is more recent, closer to my subject, and more uncomfortable, because it concerns someone whose competence is real. Luc Julia is an engineer and a doctor of science, worked at SRI in the 1990s with Adam Cheyer on agent architectures, led a Siri team at Apple from 2011, and was then chief scientist at Samsung and at Renault. Nobody disputes that career. What has been disputed is the label: "co-creator of Siri", which became in France "the father of Siri" and then "the pope of AI". But the company Siri was founded in 2007 without him, by three people who are not him, and he joined Apple after the acquisition, once the assistant was already in the iPhone. His patents from the SRI period were indeed acquired by Siri Inc., but to build up an intellectual property portfolio, not to build the product.

The point that interests me is not the quarrel over the title. It is what the title produced. In June 2025, he was heard by a French Senate committee as an AI expert, presented on the institution's website as the "designer of Siri". There he asserted, among other things, that language models have "a 36% margin of error", a figure that corresponds to no known measurement and that means nothing without saying which task is being discussed. The senators praised a fascinating contribution and a Cartesian clear-sightedness. It took a video essayist with a doctorate in philosophy taking the hearing apart point by point in August 2025 for the specialist press to take an interest, and for Siri's co-founders, when asked, to state publicly what his role had and had not been.

This case says two things the first did not. First, that real competence on one subject, here software engineering in the 2000s, does not transfer to another, here language models in 2025; the label, on the other hand, transfers frictionlessly. Second, that the institutional filter does not work: the Senate did not check, it read the biography. And the hurried reader will have noted that the only person to have done the verification work is neither an AI researcher nor a journalist, but a philosopher with a YouTube channel. The four questions require no degree; they require asking them.

That is exactly what is being replayed today with "AI expert", on a far larger scale, and with profiles far less visible, and therefore far less investigated.

Why I removed "in AI" from my biography

I am an academic in computer science, and an associate researcher at IRIT. That last phrase deserves a clarification, since I used it above about Aberkane: associate researcher means attached to a team without holding a post or a salary there. That is my situation. What distinguishes the two cases is not the status, it is what is put behind it: my publications are listed, with their DOIs, on HAL and on DBLP, and anyone can count them. I have a doctorate, and I work with language models daily, in teaching and in research. For a while, my biography said "AI researcher". I had put it there for a reason I believed to be good: the field was becoming fascinating, I was spending more and more time on it, and I told myself the phrase would help with outreach. People would know straight away what I was talking about.

Wrong. It is exactly the opposite. The suffix did not say what I was talking about, it said that I could talk about everything. In the eyes of a non-specialist reader, it gave me authority over the whole field when my actual competence covers a fraction of it, and it erased the difference.

I can describe that fraction precisely. The only part of my research that looks inside a language model concerns the way it vectorises words: how a text becomes a list of numbers, what those numbers capture and what they miss. It is a precise, technical, bounded subject, and I am the first to criticise what I say about it. The rest of my work on AI, guardrails applied to disability, uses in engineering education, studies models from the outside, as black boxes whose behaviour is measured with a protocol. That does not require knowing what is inside, and it does not make me someone who knows what intelligence is.

I insist on this because I see more and more papers, and theses, with "some AI" bolted on. Bolted on literally: the work already existed, it stood up on its own, and someone wired onto it a thing that vaguely resembles an LLM, or an ersatz neural network. Magic: since nobody can explain what it does, the results are bound to be good. Or not. BUT IT IS PUBLISHABLE, LOOK, I DO AI! No, Alex, you have merely added a black box that sometimes behaves like a performing dog barking on cue, and sometimes throws bananas. What separates the two cases, you do not know, and the paper does not say either.

In some cases this goes beyond a fashion. Data gets thrown at OpenAI or Anthropic without a thought for what it contains or where it lands, the feedback from those agents is taken up, folded into the manuscript, and signed. Well done, fine scientific ethics. Using these tools does not shock me, I use them daily and I say so; not saying so, and claiming authorship of what one did not write, is something else.

Tools are going the same way. For a year now I have been seeing "this weekend I designed", "this weekend I built", "I developed an app that does". Me, this weekend I bought a box of microwave pasta, and I did some cooking.

Let us be clear: I am throwing stones at nobody. Not at the people who use these tools, nor at those who design with them. I defend their usefulness, and without reservation: in disability, where I work, there is next to nothing that works, and the need is ENORMOUS. There, these models do things no other tool had done before.

What makes me shout is when the mention disappears. Robin, who teaches art history, turns out in three days an application that renders 3D with raytracing. Nooooooo, seriously? Nobody asks how. The demo runs, the project is signed off, and the question of who will know how to maintain it will come up later.

And I shout louder when whoever produced it does not know what they produced, because the flaws then go unnoticed, and some of them are dangerous. True story: at a project defence, a student whose program was crashing in front of us answered, "well sir, I can't explain it to you, I'm not the one who coded it, you can just ask the AI." And they did, there and then: they opened a chat with Codex and typed my question into it. They were not joking, and they could not see what was extraordinary about the sentence.

To give an idea of the extent of the field, here is my own path through it. I entered it in 2013 through automatic sentence parsing, that is, through natural language processing. I left it during my thesis, which I devoted to disability and human-computer interaction, with text entry for visually impaired people. I came back later through yet another door, that of ontologies and knowledge alignment. Three entrances, three communities, three vocabularies, and each time I had to relearn almost everything. These three subfields fit under the same acronym, and yet someone who excels in one may have nothing to say about the other two. That is what the word "AI" conceals: it designates a continent, and it is used as though it designated a competence.

That, by the way, is why I dislike the term. It says nothing of what one does, nothing of the method, nothing of the scope; all it does is enlarge whoever pronounces it. "AI expert" now evokes nothing to me but a windbag, blowing its own hollowness over the wind of the ignorant. And it is an ill wind that blows nobody any good.

What made me understand this was an invited lecture that Mr Jude Rola did me the honour of organising this year. I talked about what I know. And the questions from the room were: "What do you think about consciousness?", "How do we build a virtual human being?", "Are machines going to replace us?". None of these questions has a scientific answer today, and most are not even scientific questions: they are philosophical, or anthropological, or quite simply anxieties. But they were put to me, because I was "the AI researcher" and the title implicitly promised that I had something authoritative to say about them. The only honest answer was "I don't know, and nobody knows", and it always disappoints. I gave it anyway. That was the moment I measured what the title was doing in my name.

Now, I am in the habit of not pronouncing on what I do not master. Or, more exactly, of asking questions of those who think they know, which has a talent for irritating them; Socrates remains my favourite troll, and the method is unbeatable, since in the worst case I am right, and in the best case I am wrong and I learn something. A precise example: I do not know how something we call "intelligence" emerges from stacking billions of artificial neurons, objects that share with the biological neuron only the name and a vague analogy of function. Nobody really knows, as far as I am aware. On that point I keep quiet, or I say that I do not know. The phrase "AI researcher" on my page suggested the opposite: that I had an authoritative view on it, like all the other experts.

So I removed it. I was taking part in the farce, without having meant to, simply because the phrase is available and because it pays.

The title "AI expert" is precisely designed so that you never have to state your scope. That is why it should never be taken at face value, including when you are the one wearing it.

The good news is that checking is within everyone's reach. It takes thirty seconds to type a name into HAL or Scholar. People can check perfectly well; what nobody ever told them is that it was possible, and where to look.

The state of this field is serious. The fake experts in it outnumber the real ones, are more audible, and are better paid. Be wary. Check. And when someone tells you "I am an AI researcher", ask them where.

A word, to finish, for the merchants of hot air. The bubble will burst, as the internet and online retail burst at the end of the 1990s, and cryptocurrencies twenty years later. Some of these AI experts will then be experts in unemployment, for as long as it takes to redo a biography. Give them a few weeks before they come back as quantum AI experts: the word is already in the air, it impresses more, and it has this advantage over AI, that it is even harder to check.

References