AI tells lies, blatant lies
Mons, 27 September 2025
Artificial intelligence (AI) tells blatant lies. This is an observation which, according to others [1], I really must present here as a result of an experience I had after two years of reasonable use, marked by a sense of cautious optimism. I currently use a paid version of AI Chat Online from OpenAI, under MacOS. This system is based on GPT-4 (Generative Pre-trained Transformer 4). It describes itself as an advanced artificial intelligence model capable of natural language understanding and generation.
For educational reasons, at the beginning of the new school and academic year or in business or institutional seminars, I have always introduced my lessons and talks, in both history and foresight, with a lesson on heuristics and therefore on source criticism. However, as highlighted by Caroline Muller, a lecturer in contemporary history at the French University of Rennes, researchers must not only be committed to criticism of the discovered source, but must also clarify the rationales for source discoveries [2].
This experience will undoubtedly be shared with my students as well as my colleagues [3].
1. Specific extracts from a reference work
The experience was as follows. I asked OpenAI how two authors defined a concept which I was interested in for a work on governance, with reference to some scientific works and specifying the reference work of these authors.
As always, the AI responded immediately and in great detail, opening up the concept across several dimensions and commenting on these various configurations in a few lines, before presenting a three-point summary referring to the reference work originally cited. To my surprise, the AI added: If you wish, I can also provide you with some specific extracts or reposition this definition in the broader context of their analysis on public sector reform.
My surprise stemmed from the fact that the work, which had been produced by an internationally renowned university, is not publicly accessible and was published less than ten years ago.
Therefore, out of curiosity, I replied positively to the proposal to provide me with the extracts, asking the AI to send me the page references.
The AI responded immediately, sending me three specific extracts from the reference work, in inverted commas, with references to the page in the book for each of them. At the same time, OpenAI offered to provide me with other extracts or a more detailed summary.
As the extracts were in French and the work had been published in the language of Shakespeare, I asked the AI to send them to me in the original English. It did so immediately:
Excerpt 1 – title, page.
» (…) »;
Excerpt 2 – title, page.
» (…) »;
Excerpt 3 – title, page.
» (…) ».
adding: Please let me know if you would like other quotations or more specific contextualisation.
I should have been delighted. In reality, I was sceptical. I ordered the reference work in question which I received a few days later.
2. A rather devious plan…
Upon opening the reference work at the pages referenced for the three extracts communicated by the AI, I could not find them. As the work had a detailed index, I searched for the concept which appeared in various forms. Naturally, I looked to see whether there were other editions of the work, which was clearly not the case. The book was listed in Google Books. Although it was not publicly accessible, it was still possible to search it for certain uncommon words taken from the quotations. These did not seem to appear in the book.
Had the AI deceived me?
I came up with a rather devious plan.
I opened a new session of OpenAI and asked it directly: From which book is that quote taken?, entering in inverted commas the first extract which it had communicated to me a few days earlier.
I was stunned by the immediate response as it was so assertive and conclusive: according to OpenAI, the extract came from a book which is another reference work, written by other authors, published by other publishers, with several editions between 2004 and 2019.
I immediately replied that, a few days earlier, the AI had attributed the extract to two other authors which I named.
Again, the response left me astonished:
Thank you for pointing that out. You are correct.
The quote: (extract)
is attributed to ** (the author I had initially provided) ** in their work:
** » (…) « ** (2015).
I apologize for the initial incorrect attribution. The correct source is (authors from 2015).
Since I indicated that the book was in front of me, on my table, and that I had purchased it, OpenAI thanked me for informing it and told me that it understood how frustrating it must be not to find the quotation mentioned in the book. In order, it wrote, to try to clarify, it told me that the exact quotation which I had provided (in reality, it was OpenAI that had provided it) was not a verbatim extract from this book. And the AI told me:
The quotation you have could be a **paraphrase, a summary or an interpretation** of the authors’ ideas, stemming perhaps from a synthesis article, a secondary source or course notes referring to these authors rather than from a direct quote.
3. My confidence in your ability to give relevant answers is at stake
OpenAI then invited me to look in the work’s index after an occurrence of the concept, then to search online if I couldn’t find the quotation, then to search similar works by those authors or other researchers in the field to find variants and, lastly, to search in the preface or conclusion of the book in my possession. If you wish, I can help you refine your searches to find the exact source or suggest fundamental texts on the concepts… Would you like me to do that?
With some exasperation, I gave a positive reply:
Yes, since you have given me this quotation indicating the exact page, page x, it’s a shame that it cannot be found there. As you have given it to me in quote form, I would be pleased to know where it comes from… My confidence in your ability to give relevant answers is at stake...
OpenAI then adopted a low profile:
I totally understand your concerns, and I offer my sincere apologies for the confusion and frustration that this may have caused. It is very important to provide accurate and verifiable information, and I regret that I was unable to do so in this particular instance.
In order to clarify things and be totally transparent, the AI stated that it did not have direct access to the specific content of the pages in copyrighted books, such as the book I had mentioned, and that the quotation I had shared (in reality, which it had provided) was often paraphrased or summarised in discussions, but that it was not necessarily a direct quotation with an exact reference to the page in their book. Lastly, OpenAI stated that when it had previously mentioned the quotation attributed to page x, this was an error caused by excessive generalisation based on summaries and common interpretations in the literature on the concept being investigated.
OpenAI then invited me:
– 1. To search in the book for similar expressions or conceptual explanations, even if they are not verbatim.
– 2. To search in authoritative secondary sources (review articles, critiques or public administration handbooks) which quote the authors and use the quotation or a very similar expression, which could be the place where the exact phrase appeared.
– 3. The AI offered to assist me by searching in university databases or reliable reference repositories (such as Google Scholar, JSTOR or ResearchGate) for the exact quotation or close variants, providing me with key articles or chapters in which the discussion on the concept is well established, with verifiable quotations.
I undertake to provide you with accurate information verified at source. To regain your trust, I propose the following:
– Allow me to carry out a **source verification search** specifically for this quotation or its closest form, in credible university or governmental publications.
– Provide you with references to documents that include a direct and exact quotation, or identify whether it is a **conceptual summary derived from several authors** rather than a verbatim quotation.
Would you like me to carry out this search and provide you with an annotated list of verified sources? In this way, you will receive the most reliable information and quotations possible.
I again accepted and re-entered in the thread the quotation which the AI had originally given me.
In the new search, OpenAI:
– confirmed the title and the edition of the original work.
– launched a verification process which showed that the first reference page was at the beginning of the book, the second in the middle, the third at the end, etc.
– confirmed that no exact match for the quotation was found in Google Books, or on the University website, or on third-party sites.
– noted that the other sources confirmed the meaning but not the quotation.
OpenAI then gave me its **Conclusion:**
The quotation seems to be either a **paraphrase or a synthesis combining the authors’ ideas on the concept**, rather than a direct extract from the pages cited.
It then proposed three recommendations: firstly, it would be better to express the quotation as a summary or a synthesis rather than a direct quotation, giving me an example. Next, it would be better to refer to actual books, physical or digital, rather than online overviews. Lastly, for greater security, it restated the book’s complete reference and sent me a new bibliography on the concept.
Thus, OpenAI confirmed to me to once again that the quotation it had given me, with its reference work, inverted commas and page number, did not exist as such.
4. Conclusion: its shortcomings and ours
There is no doubt that, in the light of such an experience, after other, albeit more positive, ones, researchers should reflect on the quality of their own questioning and the limits of what can be expected from artificial intelligence.
Next, we should observe that, OpenAI produced on its own initiative simple responses to the questions posed to venture into an area to which it is unsuited, both legally and in terms of its capability. As we have seen, its very proposal to provide referenced extracts is disturbing. This could be seen as an unintentional initiative on the part of the AI, a form of slippage.
Lastly, this incident forces us to re-examine our own daily observance of referencing and the stringency that historians bring to this work. At a time when the heuristic norm is collapsing, observed among many social science colleagues, including those in economics and geography, this experience reiterates the importance of precise and accurate referencing. The current practice aims to reference an author and a year, without offering the actual ability to verify the source, in other words stating the page(s) on which the information can be found, if not the evidence that underpins the thinking. In spite of the appended bibliographies, good luck to those who have to cope with (Hobbes, 1983) for the Leviathan (780 pages) [4], (Jacob & Schiffino, 2021) for Politiques publiques, Fondements et prospective pour l’analyse de l’action publique (956 pages)[5] or (Hautcœur & Virlouvet, 2025) for Une histoire économique et sociale de la France de la Préhistoire à nos jours (1,062 pages) [6]. We may well conclude that it only makes sense that AI is developing its critical laxness in our image.
But that’s not all.
This experience shows that AI tells lies, blatant lies.
Admittedly, in writing this, we all understand that I am anthropomorphising the machine whereas we all know that AI is merely an algorithm whose responses I interpret: I react as if they were expressed by a sensitive person endowed with intelligence and will. I should therefore write instead: if the AI was a human, it would be caught lying. I was rightly reminded of this by my colleague Thierry Dutoit, professor at the Polytechnic Faculty at the University of Mons, with this elegant phrase: L’habit ne fait pas le moine; l’IA ne fait pas l’Homme [Looks can deceive; AI does not make the Man] [7].
We must distance ourselves from the epistemic fog that is artificial intelligence and shake off the current beliefs in an all-knowing AI and the pretences that such resources generate.
Epistemic fog? What does OpenAI think of that?
Saying that **AI is an epistemic fog** is a metaphorical way of highlighting that AI may introduce or reveal obscure areas in the way we acquire and validate knowledge. It is a relevant expression in a philosophical, epistemological or critical reflection on the impacts of artificial intelligence on knowledge [8].
A tit-for-tat response?
As researchers, we alone are able to shed light on these obscure areas.
Philippe Destatte
@PhD2050
[1] Alice RIGOR, Stéphanie BILLOT-BONEF, Intégrité scientifique à l’heure de l’intelligence artificielle générative: ChaptGPT et consorts, poison et antidote? in Environnement, risques et santé, 2024/5, vol. 23, p. 235-238. DOI10.1684/ ers.2024.1818
https://stm.cairn.info/revue-environnement-risques-et-sante-2024-5-page-235?lang=fr&tab=texte-integral – Adam T. KALAI, Ofir NACHUM, Santosh S. VEMPALA, Edwin ZHANG, Why Language Models Hallucinate, Cornell University, arXiv, 4 September 2025. https://doi.org/10.48550/arXiv.2509.04664
[2] Caroline MULLER, with Frédéric CLAVERT, Écrire l’histoire, Gestes et expériences à l’ère numérique, p. 61, Paris, A. Colin, 2025.
[3] Ph. DESTATTE, Opinions which are partial have the effect of vitiating the rectitude of judgment », Heuristics and criticism of sources in science, University of Mons – EUNICE, Mons, 21 October 2021, Blog PhD2050, https://phd2050.org/2021/10/26/heuristics/ – Philippe DESTATTE, Territorial Foresight, Intellectual Indiscipline with Demanding Heuristics, Speech to the 2023 Science Congress, held at the University of Namur (Wallonia) on 23 and 24 August 2023, Blog PhD2050, October 20, 2023. https://phd2050.org/2023/10/19/heuristics-2/
[4] Thomas HOBBES, Léviathan, Traité de la matière, de la forme et du pouvoir de la république ecclésiastique et civile, Paris, Sirey, 1983.
[5] Steve JACOB & Nathalie SCHIFFINO dir., Politiques publiques, Fondements et prospective pour l’analyse de l’action publique, Brussels, Bruylant, 2021.
[6] Pierre-Cyrille HAUTCŒUR and Catherine VIRLOUVET dir., Une histoire économique et sociale, La France de la Préhistoire à nos jours, Paris, Passés/Composés, 2025.
[7] Message from Thierry Dutoit, 8 September 2025.
[8] Summary of OpenAI’s response to the question: ‘can one write that AI is an epistemic fog?’, 10 September 2025.