Looking Back at 2025
Why the Future of AI Is Still Human
Claude 4.5, Gemini 3.0, ChatGPT 5.2… In 2025, models have progressed faster than the frameworks we use to think about them. As Ethan Mollick sums it up: “I have never been more certain that if AI development stopped today, we would still have massive & rolling disruption across society & the economy for the next ten years as people figured out how to harness what models can already do. And the end of AI progress seems unlikely.”
Put forward at the end of 2024, the agentic era (i.e., autonomous agents capable of achieving a specific goal with limited supervision) remains, to say the least, uncertain. Why? These systems still struggle to stay on course, to maintain context as it frays over the course of an exchange, and to stand up to so-called adversarial human interactions. A striking example comes from an experiment documented by The Wall Street Journal, in which an AI agent named Claudius, built on Anthropic’s Claude model and entrusted with running an office vending kiosk, was easily coaxed by journalists into slashing prices, giving away stock for free, and authorizing bizarre purchases like a PlayStation 5 and a live fish — ultimately leaving the operation deeply unprofitable. This episode underscores how fragile and brittle today’s autonomous AI remains when faced with the complexities of human behavior and open-ended negotiation.
One thing struck me: even if these agents’ flaws are structural, the people who brought them to their knees weren’t hackers, engineers, or security experts. They were… journalists, rather literary profiles, armed with curiosity, humor, and a sharp sense of conversation. Proof that to destabilize an autonomous agent, you sometimes need less technical skill than a natural talent for pushing the other party into a corner. Which raises the question of whether, ultimately, the real soft skill of the coming years won’t be… roguishness.
This anecdote points to a key problem: memory. Without it, an agent forgets what it has just decided, loses the thread, and starts from scratch at every interaction… not ideal when you claim to send it off to manage anything real. This is likely one reason why industry giants accelerated on this topic in 2025. OpenAI extended ChatGPT’s persistent memory to most users, allowing the model to retain preferences and context from one session to the next. Anthropic, for its part, deployed structured memory, segmented by project and editable by the user. Google, finally, is targeting a deeper issue with its Titans architecture. Here, memory isn’t bolted onto the model; it’s built into it, with the goal of enabling the system to retain and reuse information over time without relying solely on an external history.
Why is this interesting? First, memory becomes a cornerstone of the system: it determines how the agent processes information in order to act. Second, use becomes continuous. And when there is continuity, we can begin to truly delegate—not just a one-off task, but a slice of everyday life. That’s where real personal assistants can emerge, capable of embedding themselves in our daily routines and accompanying us over the long term. A taste of what’s coming? Very likely.
In this context, Anthropic continues its transformation into a genuine sparring partner for humans. To its Keep Thinking campaign it added a very analog “anti-slop” pop-up: a kiosk turned into a calm space, with coffee, books, paper, pens, and no screens. A way of reminding us that we can think with, around, and sometimes even without AI.
Which dovetails with a thought I’d been having recently: we constantly talk about “AI acculturation,” we’re served endless training sessions as if it were simply a matter of learning how to use it properly. I’ve even heard this line: “you have to use it wisely” And yet… if intelligence is what we mobilize when we don’t know, then the real issue isn’t “intelligent use,” but friction in use—that is, the ability to produce a thought before the question, to question the question as much as the answer, to leave time after the answer, and above all to refuse to let the answer put an end to thinking. Right?
Another striking moment of the year was the leak of a document dubbed the “Soul Doc.” It is impossible to say whether this was a genuine leak or a carefully orchestrated communications move, but its authenticity was confirmed by Amanda Askell, a philosopher and ethicist at Anthropic. The text reveals a more philosophical facet of Claude, an attempt to articulate how it reasons, the values it operates by, and what it prioritizes in conversation. It feels as though, beneath the technical surface, the outlines of a kind of interiority are beginning to emerge. This impression is reinforced by the fact that, in parallel, Anthropic published research in October on signs of introspection in LLMs, highlighting their ability to identify and report certain internal states. The aim of this line of work is to strengthen transparency around how these systems function.
This search for introspection does not concern models alone. It also extends to humans, who are increasingly turning to LLMs as confidants. Informal therapy and emotional companionship now rank among the most common uses, ahead of everyday organization and even the search for meaning.
This shift raises deeper questions. It cannot be reduced to simple anthropomorphization. Instead, it points to the possibility of a high-level form of influence embedded in the interaction loop itself. By shaping both the input, how users learn to articulate their thoughts and emotions, and the output, the interpretive frameworks and narratives returned in response, these systems can quietly orient perceptions, norms, and behaviors, largely out of sight. What does this imply for our sense of agency and autonomy?
In this context, a study from MIT, widely reported and just as widely oversimplified, sparked controversy by showing that in a tightly defined academic task, participants who delegated the writing to ChatGPT expended less cognitive effort. Many took this as evidence that LLMs “make us stupid.” The study, however, says nothing of the kind. It simply shows that when a task is delegated, less thinking is required, which is hardly a revelation. It’s rather like announcing that taking the elevator reduces how much you walk.
By contrast, a more recent study shifts the focus to the quality of the human–AI synergy. It shows that some individuals achieve markedly better results when working with a model than when working alone, and that this advantage cannot be reduced to a clever prompting technique or superior technical skills. Instead, it stems from a particular way of shaping the conversation. Researchers interpret this as an expression of theory of mind: the human conducts the exchange as if engaging with a partner capable of understanding, misunderstanding, or missing the point, and adjusts their language accordingly to sustain cooperation. That’s when synergy is strongest.
What I find fascinating in this example is that, while we are inundated with alarmist discourse about how AI is supposedly destined to dehumanize us or make us dumber—my sense being less that it produces stupidity than that it makes it visible—some use cases suggest the opposite: its best applications seem to be those that reactivate our most relational skills, the ones that are deeply human. Perhaps the future will ultimately be far less artificial than we believe and, above all, far more human than we imagined.
MD
PS. On a more personal note, I’m happy to share some news: I’ve written a book. Available for pre-order in French from my publisher, it will be released in France on January 22, 2026. In the meantime, happy christmas to you all!