Yes, But...
In Praise of Nuance in the Age of Artificial Certainty
Generative AI produces as many answers as it does ready made truths. The problem is that certainty travels faster than nuance. And nuance, precisely, requires an effort that certainty often allows us to avoid…
What follows is a series of reflections on ideas that seem more complex than they first appear. Perhaps it is also an invitation to cultivate the critical thinking that we are so often encouraged to develop, yet which may be one of the most paradoxical imperatives of our time. After all, doesn’t critical thinking emerge precisely from doubt, from friction, from that temporary void that compels us to search, hesitate, make mistakes, and reformulate our questions? From everything that traditionally separates a question from its answer? Yet this is precisely the distance that generative AI tends to reduce.
In short, we are only beginning to understand what becomes of those intellectual detours we spent years taking, often without fully appreciating their value….
“AI will revolutionize learning because everyone will have a pocket tutor, like Aristotle for Alexander the Great.”
Yes, the flexibility of AI systems, their constant availability, and their ability to adapt to each learner’s level will very likely transform education. But this comparison between Aristotle and Alexander, often invoked in Silicon Valley, overlooks several important nuances.
First, Alexander was not alone with Aristotle. He was educated at Mieza alongside other young Macedonian aristocrats who would later become his companions and generals. His learning was therefore shaped not only by his teacher, but also by interactions with his peers.
Second, Aristotle did not simply know Alexander the student. He knew Alexander the person. A significant part of teaching rests on this deep understanding of another human being, built through presence, interaction, and all the things that language alone cannot convey.
Finally, teaching is not merely the transfer of information. It is also a form of performance. We remember great teachers for their passion, their enthusiasm, the way they occupy a room, or bring an idea to life. A lesson is often as much an experience as it is a body of knowledge.
Having a pocket tutor is undoubtedly a remarkable advance. But if we are going to pursue the comparison with Aristotle to its logical conclusion, we must also ask how we might reproduce otherness, relationships with peers, embodied presence, and the human dimensions of learning that extend far beyond access to information.
"Generative AI systems are sycophantic. But all you need to do is create multiple agents that debate each other to solve the problem."
Much has been made of conversational AI’s tendency to tell users what they want to hear. The familiar: “You’re absolutely right...”
The more sophisticated users were quick to point out that the solution was simply better prompting. Then came the rise of agentic systems: a critic, a challenger, a devil’s advocate, a domain expert. The goal was to artificially recreate the benefits of contradiction and diversity of viewpoints.
To some extent, this works. This tendency is partly a consequence of the prompt structure itself, in which the AI seeks to validate the user. Breaking that dynamic forces an agent to actively look for flaws through debate or Red Teaming, an approach that can be remarkably effective at reducing hallucinations, identifying bugs, and strengthening reasoning.
But this approach rests on a much heavier implicit assumption: that contradiction automatically produces otherness.
Yet asking five agents built from the same model to debate one another is a bit like gathering five people who read the same books, attended the same schools, and share the same blind spots. They may disagree on the surface and correct one another’s factual or logical errors, but they do not necessarily generate genuinely different ways of thinking.
To achieve that, we would need models shaped by radically different cultural, linguistic, and cognitive foundations. Today, however, large language models remain largely developed within a relatively homogeneous technological framework, at least in the case of the major Western models. They are trained on corpora that, broadly speaking, reflect many of the same references and assumptions.
In short, multi-agent systems introduce an interesting form of procedural contradiction. They often reduce this tendency to validate the user and improve the robustness of reasoning. In fact, part of their effectiveness may have less to do with the intrinsic diversity of the agents than with the role of the orchestrator that assigns tasks, moderates exchanges, and synthesizes conclusions.
What they do not necessarily recreate is the richness of genuine otherness. Because there is a profound difference between varying perspectives within the same underlying framework and varying the worlds from which those perspectives emerge.
“Language is language.”
One of the reasons generative AI impresses us so much is that it handles language with remarkable fluency. The more we watch it manipulate words, the more we tend to anthropomorphize it and assume that it uses language in the same way we do.
But that assumption conflates two very different uses of speech.
Part of language serves to describe the world. Saying that it is raining, explaining a theory, or recounting an event all belong to this descriptive function. It is precisely here that language models excel.
Yet another part of language does not describe the world. It changes it.
When someone says, “I promise,” “I resign,” “I forgive you,” or “I now pronounce you husband and wife,” they are performing an act that alters a relationship, creates an obligation, or commits themselves, or an institution, to a particular course of action.
In other words, an AI can produce a promise without actually promising, or a marriage without marrying anyone.
And perhaps this brings us back to a more fundamental idea that AI may help remind us of: the true power of words does not always lie in what they describe. It sometimes lies in what they commit us to.
MD