Source: French to English Tester Published on: 2026-06-08
Source: The Conversation – in French– By Rossana De Angelis, Associate Professor of Linguistics, Paris-Est Créteil Val de Marne University (UPEC)
You may have already experienced it: text-generating artificial intelligences, or AIs, regularly invent response elements, which are all the more difficult to detect as they integrate into a seemingly coherent whole and blend with verified facts. By examining how these AIs search for information, how they analyze « crawled » or scanned texts, one can understand why they are so frequently mistaken.
It has already been more than three years that we have been able to use generative artificial intelligences (AIs) for texts and images. ChatGPT was launched in November 2022 in a free version, opening the door to massive access to generative AIs, but it did not yet have access to the Web as a source of information. It was from January 2023 that it became possible to ask all kinds of questions, submit “prompts” or “invites,” because now, alarge part of what is found on the Netmay be explored by the machine.
And it is from this “presumption of omniscience”, in other words the capacity we attribute to the machine to know everything, that we began to experience the “hallucinations” of AI: a term derived from Latinhallucination, meaning “mistake”, constructed from the verbhallucinatewhich means “to be mistaken”, “to ramble”, but also “to deceive the interlocutor”.
I myself was faced with the hallucinations of ChatGPT when, during the preparation of a scientific conference, I asked on the interaction interface: “Who am I?” to test the limits of human-machine interaction within the framework of a dialogue. This simple question generated a linguistic experience that allowed me to better understand how certain hallucinations arise.
This experience,which gave rise to a scientific article, allows me to sketch initial answers to the questions: when and why do generative writing AIs make mistakes?
What does not exist on the Net does not exist for generative AI
The presumption of omniscience falls when we take into account the criteria that allow information to be processed by the machine. Firstly, this information must be recorded in the form of digital data. Secondly, it must be available at all times, as machines only operate when connected to servers that ensure access to data carrying information. Thirdly, this information must be identifiable within the considerable mass of available information.
The first two reasons help explain certain types of hallucinations, because what does not exist on the Net does not exist for machines. When I asked a generative writing AI to answer the question: “Who is my grandmother,” knowing that the grandmother in question has no Instagram account, blog, or institutional page, it replied: “I cannot know who your grandmother is without information from you. If you give me her name, the context, or what you want to know about her, I can help you as best as I can.” I then provided “a name,” a valid label so that the information could be identified, which prompted the machine to offer incorrect descriptions rather than remain without an answer.
The issue of referencing
The third reason helps explain other types of hallucinations, because what exists on the Net but is not visible to machines cannot be immediately detected by generative AIs. This dimension is specific to digital writing: it is linked toreferencingdigital documents through tags (“tags”) that indicate what they contain and what they refer to, because the documents are linked together forming a network of documents (the “Web”), some more or less visible than others. There are two levels of referencing: the first consisting of placing keywords and links useful for promoting texts, and the second consisting of buying traffic and links to increase the visibility of texts on the Web.
When I asked ChatGPT to tell me “Who is Rossana De Angelis?”, the machine mistakenly suggested a definition of me as an artist, due to the presence on the Net of an artist whose identity partially matches my name, before identifying me as a researcher. This is explained by the fact that the documents making up an art object sales website, where the description of the artist in question was found, are better indexed (and therefore more “visible”) than the documents of an institutional site, because of their commercial stakes.
Algorithms that direct machines towards data follow the most visible directions: the best-referenced, best-tagged documents stand out first. And if the best-referenced documents are those with commercial value, it follows that the most visible data are those linked to commercial interests. And this, regardless of the quality and truthfulness of the content.
LLMs, a complex functioning
The three criteria of presence, availability, and visibility do not explain all possible hallucinations, far from it. The greatest difficulty lies in grasping the complexity inherent in the functioning of generative text and image AIs that use large language models (LLM) to process and generate content.
Also read:
How does ChatGPT work? Decoding its name to understand language models
For example, for image recognition, the initial processing captures fundamental aspects such as contrasts, edges, lines, their orientations, etc. The subsequent processing combines this data with other data concerning, for example, textures or materials. Finally, from this set, further processing allows the construction of numerical representations of images, such as objects or faces, by recording as “values” the differences and similarities between the data. Data processing thus works through overlapping layers (deep learning)Â: the deeper this overlap is, the greater the analytical capacity increases.
AIs organize their data in a“vector space”A: a set of objects each occupying a specific place. Let’s take the string game: it’s like multiplying at will the number of threads, their length, their thickness without changing the rules of the game, and then being able to identify a knot. In this space, the position of each object is determined by the relations it maintains with the others. The set of these relations determines what we call their “vector value”: a combination of numbers that defines a place in a space. These values allow data to be identified as positions just like the knots in a string game.
Also to read:
From Cambridge Analytica to ChatGPT, understanding how AI gives meaning to words
Vector values allow identifying regularities: nearby data (for example, fruit names: strawberries, cherries, apricots, etc.) carry similar information (for example, they can be picked, cut, cooked, eaten, etc.) They can therefore be detected by their proximity (for example, similar fruits, similar actions).
To detect data, and therefore to request information, we must submit a request (a “prompt” or “invitation”) to the machine that searches for these patterns in themvector spacesWith the help of words (for example, “cherry clafoutis recipe”). This is why a generative writing AI can provide us with the correct recipe for “cherry clafoutis” first by identifying, then reproducing the regularity of the data using the words “clafoutis, cherries, recipe” (similar words, similar contexts, similar actions). The relationship that the data maintain in these spaces guides the more or less accurate reproduction of information, inautomatically complementssequences of words in the interaction space.
Also to read:
“World models,” when artificial intelligence learns to understand the world
On the importance of context
What allows the use of words lies in theirvalue. It is a principle that lies at the foundation of general linguistics. More precisely, the value of a word is a differential value: within a language, the word “dog” is distinguished from the word “cat,” just as the word “blue” is from the word “green,” and they mean what they mean precisely because of this difference. However, this value is defined in relation to the other words present in the text (semantic context: what is said with the words) and in relation to the context in which the text is embedded (pragmatic context: what is done with the words).
Let’s take an example. Any word written on a road sign only works in a specific context (a road) and in a specific practice (driving). Most of the texts we produce or interpret daily operate in this way.
How can a machine extract the context of a text? It does so precisely through the process mentioned above: by defining the value of words based on their vector positions. In this way, the machine reproduces the semantic value of words. However, not always having access to the context, the pragmatic value is not always definable. Using the example of a road sign again, “turn right” or “turn left” means nothing without context. Under these conditions, if you ask a generative AI whether to “turn right or left,” it will be unable to answer correctly.
In other words, when pragmatic value cannot be defined, the machine can make mistakes. This results in what can be called a “pragmatic breach”: since the contexts and usage practices of words are inaccessible, their pragmatic value becomes elusive.
Why is it difficult to understand how generative AI works?
What prevents understanding the functioning of generative AI is its complexity. For example, one difficulty is imagining that data occupy places in spaces that are not two-dimensional, but vector spaces with n dimensions, because we cannot conceive and imagine spaces with more than four dimensions (width, height, depth, time). As if we had to increase the number of strings in the string game by n times.
It is a disproportionate change of scale whose complexity exceeds us, which forces us either to reduce the complexity of the phenomenon or to accept its incomprehensibility. This explains the attitude of glorification or demonization that we often adopt towards generative AI.
I have wondered several times whether or not generative writing AIs will one day be able to replace humans in writing practices. My position remains the same:no, AI will not make all our writing practices disappear, because obviously an entire set of highly context-dependent writings, that is to say linked to a context, such as impromptu exchanges, informal writings, or improvised signs, cannot be (re)produced in a digital environment. There are therefore entire areas of written expression to which generative editorial AIs do not (yet) have access.
Rossana De Angelis does not work for, does not advise, does not own shares, does not receive funds from any organization that could benefit from this article, and has declared no other affiliation than her research institution.
–ref. Why do generative AIs hallucinate? –https://theconversation.com/why-do-generative-ias-hallucinate-270493