Artificial intelligence is a major global concern today, and the question we must ask ourselves—beyond the scientific issues—is: How shall we coexist with artificial intelligence? How will the different cultures of the world adapt to it? The challenge is both fascinating, and for some, exciting, but also unsettling.
What we have absolutely no right to do is criticize, challenge, or oppose it without offering a solution. That would be like opposing the discovery of electricity or the internet. There will be an infinite number of answers and ways to adapt, but the one thing we must not do is to give up by turning a blind eye.
Well, there is one truly fundamental concept: that of discoverability, and it is no coincidence that this concept originated and was developed in Quebec.
What this concept describes is not fundamentally new. That works are discovered only after—sometimes long after—the author’s death; that works significant to the history of civilizations have fallen into oblivion, only to be rediscovered in a completely different geopolitical context— such as the rediscovery of works from Greek antiquity between the 8th and 13th centuries, a prelude to the Italian and French Renaissances—this is not in itself a new discovery.
However, what makes this concept unique is that it has been redefined for the digital society we have now entered.
Every innovation—and particularly all disruptive innovations—generates asymmetric effects, which the economist, philosopher, and historian François Perroux referred to in the 1950s as “effects of domination.” Such are the effects of all the innovations that follow one another in waves within the digital society, the latest and perhaps most spectacular manifestation of which is known as artificial intelligence.
Thus, in the digital world, discoverability is understood as the existence or implementation of processes that enable content (knowledge, works, information) to be found and identified within an immense mass of information. In a sense, the European Observatory on Multilingualism was created with the aim of raising the profile of content, activities, and events centered on languages, multilingualism, and linguistic and cultural diversity.
Discoverability is only one aspect—an important one, to be sure—but it does not exhaust such a fundamental question as how to exist individually and collectively in a world of digital societies. Recently, the term “digital sovereignty” has emerged. Sovereignty is not independence, especially in a world built on interdependence, but can be analyzed in various ways as a “strong form of interdependence,” a highly significant concept also coined by François Perroux, whose meaning is “intelligible to all.”
Since we are immersed in a world of digital technologies, we must ground our reflection on multilingualism and the ubiquity of languages in social and international relations—an ubiquity that has always existed but must be examined in the context of digital technology. In this regard, we must commend the magnificent MultiLAwa project—a European doctoral initiative on “multilingual awareness in the European digital society”—led by Hélène Vinckel-Roisin of the University of Lorraine and nine other European universities, with the OEP serving as an associate partner. A call for applications for 14 doctoral fellowships has just been launched, with a deadline of July 8.
In the same vein, we should also commend the research chair on the discoverability of scientific contents in French, created by the Fonds de recherche du Québec, and the UNESCO Chair on Open Science.
At our modest level, regardless of numerous technical factors we won’t go into here, a key aspect is taking the initiative in the debate of ideas, and when it comes to AI tools, the priority is to use them to serve one’s own interests—that is, to be in control.
With that in mind, we propose a brief exercise comparing the responses of two platforms, Mistral AI and ChatGPT, to the same question.
The question is as follows: “Example: A journalist in a report [on a well-known news channel] says: ‘U.S. intelligence’ does not confirm Donald Trump’s statements.” How does the AI process this statement?”
Before letting Mistral and ChatGPT have their say, one can imagine that an average listener might be surprised and interpret the sentence as “ there are people in the United States who are smart enough not to take Donald Trump’s statements at face value.” But the AI isn’t thrown off by this insidious Anglicism, which creeps into a word that is perfectly French and English at the same time, yet whose meaning is distorted by the context of the statement and the journalist’s use of it. Here’s what they say:
-
ChatGPT
“The expression ‘l'intelligence américaine’ is most likely a clumsy translation or a direct calque of the English phrase ‘U.S. intelligence.’
In English, ‘intelligence’ does not only mean ‘intelligence’ in the sense of intellectual ability; in political and military vocabulary, it also refers to intelligence services. Thus, ‘U.S. intelligence’ means:
U.S. intelligence agencies as a whole;
or the assessment produced by these agencies.
In French, a journalist would normally say:
“U.S. intelligence does not confirm Trump’s statements”;
“U.S. intelligence agencies do not confirm Trump’s statements”;
“The U.S. intelligence community does not confirm Trump’s statements .”
Natural language processing AI generally attempts to resolve this ambiguity based on context. If the news report discusses geopolitics, war, national security, or classified reports, it will interpret “American intelligence” as a mistranslation of “American intelligence” rather than as “the intelligence of Americans.”
There are therefore two levels of analysis:
Linguistic analysis: the expression is unusual in French and suggests an influence from English.
Semantic analysis: the intended meaning is almost certainly “U.S. intelligence agencies” or “U.S. intelligence assessment .”
If you provide me with the exact sentence or a longer excerpt from the report, I can analyze more precisely what the journalist meant and whether the wording is correct or not.”
ChatGPT’s response, though very brief, gives the impression of a thorough analysis providing a clear and definitive answer.
-
Mistral
Mistral takes a different approach—much longer and much more nuanced—in which the AI’s reasoning is laid out and room for doubt is allowed. Let’s take a look:
“Your example is excellent for illustrating how AI (and, more broadly, natural language processing systems) interprets—or fails to interpret—the nuances of human language. Let’s analyze this in detail, in light of your definition of AI as ‘massive data processing that simulates certain mechanisms of human intelligence’.
1. What the AI sees in this sentence
The sentence: “U.S. intelligence does not confirm Donald Trump’s statements.”
a. Basic syntactic and semantic analysis
- Tokenization : The AI breaks the sentence down into units (words, sub-words): [“L'”, “intelligence”, “américaine”, “ne”, “confirme”, “pas”, “les”, “déclarations”, “de”, “Donald”, ‘Trump’, “.”]
- Entity recognition :
- “U.S. intelligence” → Named entity (category: organization or concept).
- “Donald Trump” → Political figure (former President of the United States) .
- Grammatical structure: Subject (U.S. intelligence) + verb (does not confirm) + object (Donald Trump’s statements).
b. Semantic ambiguity
Here, the keyword is “intelligence”:
- Meaning 1 (most likely for a human): “U.S. intelligence agencies” (CIA, NSA, etc.).
- Meaning 2 (literal): “The collective intellectual capacity of Americans”.
- Meaning 3 (ironic/metaphorical): “The wisdom or rationality of the United States”.
Problem: AI has no cultural or historical context to automatically make a determination. It relies on:
- Frequency of use: In text corpora, “American intelligence” is mostly associated with intelligence agencies (e.g., “The CIA, American intelligence, has released a report”).
- Co-text: The mention of Donald Trump (a political figure) and “statements” points to a geopolitical context, where “intelligence” = intelligence services.
→ Likely AI conclusion: “U.S. intelligence agencies contradict Trump.”
c. What the AI does not understand
- Irony or sarcasm : If the journalist uses “U.S. intelligence” in a sarcastic way (e.g., to criticize a decision perceived as stupid), the AI might miss the nuance unless it has been trained on similar examples.
- Example: “U.S. intelligence has struck again!” (irony regarding a mistake by the intelligence agencies).
- Cultural Implications:
- A human knows that “intelligence” (in the singular) often refers to the intelligence services in a political context.
- A non-English speaker or an AI might confuse this with “intelligence” as a human trait.
- The journalist’s intention:
- Does he mean that the intelligence agencies deny Trump’s remarks?
- Or that the Americans in general (according to polls and experts) do not support them?
- The AI cannot guess the intent without additional data.
2. How would a model like me process this sentence?
Step 1: Disambiguation
- Pattern recognition: In the training data, “intelligence [country]” is often associated with intelligence agencies (e.g., “British intelligence”, “Russian intelligence”).
- Political context: The mention of Trump triggers associations with topics such as disinformation, elections, or reports from federal agencies.
→ Main hypothesis: “U.S. intelligence agencies do not corroborate Trump’s statements.”
Step 2: Verification via context
If the sentence is extracted from a full article, the AI could:
- Search for keywords such as “CIA”, ‘report’, “intelligence” in surrounding sentences.
- Analyze the tone of the article (critical, neutral, satirical) to adjust the interpretation.
Step 3: Generating a response
If we ask the AI: “What does ‘U.S. intelligence’ mean in this sentence?”
It would likely respond:
“In this context, ‘U.S. intelligence’ most likely refers to the U.S. intelligence agencies (such as the CIA or the NSA), which do not confirm Donald Trump’s statements. This phrase is commonly used in the media to refer to these agencies. ”
But:
- It would not know whether the journalist used the term ironically without additional clues.
- It would not understand the political implications of this statement (e.g., a conflict between Trump and the Deep State).
3. Comparison with human intelligence
CriterionAI (your definition) Human (journalist/reader)
Word recognition ✅ Perfect (via tokenization). ✅ Perfect.
Literal understanding ✅ Good (dominant meaning identified). ✅ Good.
Contextual disambiguation ⚠️ Depends on training data ✅ Excellent (experience, culture).
Detection of irony ❌ Poor (unless trained on examples). ✅ Possible (tone, social context).
Understanding of implicit meaning ❌ None (no theory of mind). ✅ Natural.
Author’s intent ❌ Unaccessible. ✅ Interpretable (with possible biases).
4. Concrete example with a language model
If you ask a model like me: “In this sentence, does ‘American intelligence’ refer to the intelligence agencies or to the collective intelligence of Americans?”
Likely response:
“In this context, it is very likely that ‘American intelligence’ refers to the U.S. intelligence agencies (such as the CIA). This expression is a metonymy commonly used in journalism to refer to these agencies. However, without further context, some ambiguity remains.”
Why this response?
- Statistics : The model has seen thousands of examples where “intelligence [country]” = intelligence agencies.
- Caution : It mentions the ambiguity because it does not have access to the journalist’s intent .
5. Limitations and Possible Biases
- Frequency bias: If the model was trained on articles where “U.S. intelligence” = intelligence agencies 90% of the time, it will favor this interpretation, even if the context is ambiguous.
- Cultural bias: A model trained primarily on American media will find it easier to resolve ambiguity than if it were trained on French or Chinese texts.
- Lack of causal reasoning: AI does not ask itself: “Why is the journalist using this term?” (e.g., to avoid mentioning the CIA directly, out of habit, or ironically).
6. Application to your research (here, Mistral addresses me based on previous exchanges)
Your work on Averroes and cultural transfers shows that a deep understanding of a text requires:
- A historical context (e.g., knowing that secularism was revolutionary in the 12th century).
- A sensitivity to linguistic nuances (e.g., Latin translations of Arabic texts).
AI, on the other hand, does the opposite:
- It simulates understanding via statistical patterns.
- It ignores history, unless it is explicitly encoded in the training data.
Question for you:If an ancient text (such as those by Averroes) were analyzed by AI, could it detect its modernity without knowing the historical context… or would it simply classify the text as “medieval philosophy” without grasping its subversive significance?
In summary:The AI would process your example by identifying the most likely meaning (intelligence services), but with no guarantee of grasping the nuances (irony, intent, implied meanings). This is massive data processing… but not human understanding.
Some may find Mistral’s response too long or too complex, but in our view, ChatGPT answered only the first part of the question accurately and addressed the second part—which was more important than the first—only briefly: “How does AI process this statement?” . Mistral, on the other hand, focuses on the second part, laying bare the AI’s entire process and highlighting its strengths and weaknesses. It demystifies and demystifies it in the eyes of the user who is asking a question about the AI itself.
The practical conclusion regarding our journalist is that he expresses himself in a language that resembles French but uses English meanings for the words within an American context, so that a French listener is unlikely to interpret the message correctly, while an American listener who might receive it does not interpret it any better, given that it is expressed in a language that is not their own.
final note: Mistral knows me well, as I’ve grilled it on numerous specialized topics, and it knows that I’m not satisfied with brief, inadequately reasoned answers. Another internet user might not have received the same response from Mistral as the one presented in this article.Final conclusion: it is up to the internet user to master the AI tool, not the other way around.