Generative AI boosts individual performance but at the same time reduces collective functioning. Withdrawal into oneself is becoming a major risk in the AI transformation of organizations.
February 2026. MIT publishes the consolidated version of a large-scale experiment on collaboration with AI agents: 2,234 participants, randomly distributed into human-human or human-AI pairs, responsible for producing real advertising campaigns. The verdict is clear: human-AI pairs produce 50% more per person, with better text quality. The colleague becomes statistically optional. A first conclusion, probably hasty, quickly emerged: each employee, well equipped, would now be worth a team.
This reading may seem interesting from a performance perspective. But above all it establishes a new feeling: individual omnipotence boosted by AI. Why bother a colleague when an assistant responds in 10 seconds, without a shared agenda and without judgment? Some employees already prefer to ask their questions to the machine rather than exposing a shortcoming in front of a peer. The sometimes naive question, the one that makes everyone progress, migrates to the machine. AI becomes the refuge of those who no longer dare to ask.
The phenomenon goes beyond the professional framework. The MIT Media Lab and OpenAI combined a randomized trial on 1,000 participants and the analysis of nearly 40 million interactions with ChatGPT: intensive use correlates with more loneliness, more emotional dependence and less socialization. These are correlations and not proven causalities. But when a tool used every week by nearly 1 billion people correlates with the isolation of its biggest users, this signal of vigilance deserves to be observed carefully.
The cost is paid at the collective level
The same MIT experiment also contains even more disturbing data. Faced with an AI companion, exchanges become more transactional: 25% more task-oriented messages, 18% less interpersonal exchanges, more work delegated than to a human partner. And the productions of human-AI pairs are more similar to each other. Everything is there: efficiency increases, conversation becomes poorer, diversity is reduced. Everyone gains individually but the group loses its raw material.
The daily life of organizations confirms this in a more trivial form: this content generated by AI with the appearance of completed work, pushed to a colleague without proofreading, which English has already called workslop. A recent American survey estimates that 40% of employees received workslop during the previous month, with each content generated in this way costing them on average nearly 2 hours of corrections. Behind the productivity displayed, appears a simply transferred load: it is the recipient who checks, sorts and takes back. Trust, the basis of all collaboration, erodes with each botched delivery.
The withdrawal occurs through micro-decisions. Each question asked of the AI rather than of a peer can be unitarily rational. Faster, more direct and sometimes, let’s face it, more effective. Cumulatively, these decisions eat away at what makes a collective robust: team exchanges, constructive disagreements, the informal circulation of knowledge, learning through the confrontation of ideas. A collective that talks to each other is, however, a powerful engine of performance: ideas are confronted, knowledge circulates, teams come together. This bond is cultivated on a daily basis, in exchange. Maintaining it deserves as much attention and energy as deploying AI tools.
AI does not destroy the collective, it puts it to the test
Should we conclude that AI is destroying the collective? Probably not yet, even if there is still very little perspective on the subject. The same MIT study recalls that the frontier of capabilities remains irregular: on certain dimensions, human pairs retain the advantage. Above all, AI broadens everyone’s horizons: it democratizes expertise, accelerates learning, restores ambition to profiles that were self-censoring. But broadening everyone’s horizon does not create a common vision. A common vision is built through conversation, the confrontation of points of view, shared arbitration. No model will do it for teams.
The answer belongs to each organization, and it can be divided into 3 areas.
Measure first. AI dashboards track adoption: licenses, requests, productivity gains. None say if the teams still talk. However, the indicators exist, from peer-to-peer requests to the time spent reviewing the content generated. They also have their place in steering the AI transformation.
Then govern uses. An AI charter is not limited to data confidentiality. It must say when the AI is the rightful interlocutor and when a question should live in the team first. It must also deal with workslop: generated content must be accepted, signed and reread before being pushed to a colleague. Sending content to a colleague that you have not reviewed does not create any gain: the verification burden is simply shifted to them.
Finally reinstate rituals. Let us not forget this paradox: AI, a possible cause of the decline, also constitutes the best pretext at the moment to revitalize trade. Sharing of practices, team prompt reviews, collective decisions on use cases: the organizations that progress the fastest are those where AI is practiced collectively, rather than everyone in their own corner.
The feeling of individual omnipotence does not make a business strategy. It is, at best, a sum of isolated performances and, at worst, a collection of productive solitudes. The organizations that will derive the most value from AI will not be those that field the most augmented employees. These will be those which have preserved, among these collaborators, the desire and the opportunity to work and build together.
