Original article published in The Conversation on September 15, 2026:
When we work as a team, cooperation unfolds across two dimensions. Some are instrumental: sharing information, distributing tasks, exploring alternatives, detecting errors and meeting deadlines. Others are relational: trusting, building a shared vision, allowing ourselves to be affected by others without losing our capacity to act, and accepting that they can also change our position.
In human cooperation, these dimensions interact differently depending on the context and do not necessarily develop in parallel. Artificial intelligence, however, is taking many instrumental capacities to a scale we had never seen before.
This September, OpenAI presented a proposed solution to the Navier–Stokes mathematical problem, one of the seven Millennium Prize Problems identified by the Clay Mathematics Institute in 2000 as some of the major open challenges in mathematics.
What is most striking about the announcement is not only the result, but how it was reached. It was not a model answering a question. It was a system of agents: copies of the model working autonomously and simultaneously, organised into groups that could communicate with one another, test different approaches to the same problem, abandon those that led nowhere and share the most promising paths with the rest. According to the company, the group that reached the solution involved around 10,000 concurrent agents, and the result emerged some 88 hours after the first ones were launched.
That said, what this episode illustrates needs to be acknowledged without defensive qualifications. In formal problems, pattern recognition and the large-scale exploration of possible paths, machines are already better than we are, and they will become increasingly so. No human team can divide a problem among ten thousand minds, try everything at once and recombine what works in less than four days. Denying this would be a rather unhelpful form of pride.
When the result returns to the human world
And yet, when that result returned to the world of people, problems emerged that none of the agents had had to solve. The mathematician Tristan Buckmaster publicly questioned how OpenAI had handled recognition of the work he was developing with Levent Alpöge on closely related problems, and raised doubts about whether his previous interactions with Codex, the company’s own programming tool, may have influenced the process.
The result was built on decades of mathematics developed by people: the concepts, techniques, failed attempts, tools and, above all, the very formulation of the problem. Attributing the achievement exclusively to an AI, and by extension to the company that owns it, is at the very least a simplification of something that belongs to a much longer collective human effort.
Cooperating with people or with AI: what do we choose to preserve?
Cooperation between people is not merely a procedure we use because we cannot get there alone. It is how we live together. Raising a child, caring for someone who is ill, teaching, debating a municipal budget or deciding where a company is heading are not problems that we solve through cooperation: they are forms of living together that exist only while we cooperate. Human cooperation is not only a means of achieving results. It is a large part of what we understand human life to be.
This raises a less obvious question: when we work with an AI, are we really cooperating with it, or are we expanding our own capacity to act? If it is the latter, what we call cooperation may be something else: an extraordinary tool that helps us think, but with which there is nothing to negotiate.
We have become accustomed to calling AI language models assistants, copilots or even teammates. While working with them, we discard one answer after another and ask for a new one without having to weigh our words very carefully. We neither soften the rejection nor think about what the relationship will look like tomorrow. That convenience makes us consider what happens when the proposal we reject comes from a person with whom we will have to continue working. With an AI, we can always ask for another answer. With a person, a conversation remains unfinished.
The Habermas Machine, or who we can hold accountable
At this point, it would be tempting to conclude that AI is good with numbers and humans are good with people. That is not true. A study published in Science in 2024, involving 5,734 participants, tested the so-called Habermas Machine, an AI that synthesised a group’s views on contentious issues. Participants preferred its summaries to those produced by human mediators and rated them as clearer, more informative and more impartial towards minority positions. But there was one crucial detail: they could criticise the synthesis, and those objections fed into a new version. The machine helped people understand one another; the conversation still belonged to the people.
This difference matters because every decision that affects someone raises three questions: who can challenge it, who has to answer for it and who bears the consequences. An AI can make recommendations, correct itself or be switched off, but it is accountable for nothing. It does not lose its job, it does not have to explain itself and it does not live with the outcome. The more complex the system becomes, the easier it is for this chain of responsibility to be diluted. Keeping the decision on the human side is, above all, about preserving someone who can be held accountable.
There is an even deeper difference. A system can prove a theorem, but it cannot want it to be proved. Navier–Stokes mattered because there were people to whom it mattered. The agents did not choose the problem, nor did they know why it was worth solving: someone asked them to do it.
We can delegate an enormous part of cooperation: organising information, exploring alternatives, detecting errors and even helping us mediate. But the ends, responsibility and the possibility of challenging the decision must remain ours.
With an AI, we can ask for another answer. With a person, the answer can also change us. And cooperation begins precisely when the answer can still be challenged by someone we cannot configure or restart.
