Source: French to English Tester Published on: 2026-05-03
Source: The Conversation – in French– By Bruno Deffains, Professor of Economic Sciences, Honorary Member of the University Institute of France, University Paris-Panthéon-Assas
What if we were not asking the right question when worrying about the impact of artificial intelligences on employment? Beyond the jobs that might be destroyed or created, what will remain, and what should remain, human? Far from replacing all white-collar workers, AIs will change the role assigned to each person, which should prompt education to teach the youngest to become operators of abundance.
On March 10, 2026, Nvidia CEO Jensen Huang published an essay stating thatAI creates more jobs than it destroys. A few days earlier, the economistsfrom Anthropic published empirical datatelling a much more nuanced story. Between these two readings, one question remains whole: in a world where the machine produces instantly, abundantly, and often suitably, what must humans imperatively retain as their own capacity?
There is something troubling about the current debate on artificial intelligence and work. It’s not that it is false, it’s that it systematically asks the wrong question. People wonder if AI will destroy jobs, if white-collar workers are the new victims of technical progress, if they will in turn experience “a kind of deindustrialization.” This debate is legitimate.
But it dismisses a much deeper question, which is to wonder what capacities humans must retain in a world where machines produce instantly, abundantly, and often adequately. To answer this question, the concept of the operator of abundance provides insight that is in many respects relevant.
Also to read:
Generative AI will not destroy your job but it will profoundly change your profession
Abundance is not the value
Jensen Huanguses a seemingly attractive formula according to which AI generates real-time responses, contextualized and often strikingly plausible. In a few seconds, it produces a draft of judicial conclusions, a bibliographic summary, a risk analysis, a course plan. This is technically accurate. But this accuracy hides a fundamental ambiguity: to produce does not mean to be valuable.
The real question is not “Is this content good?”, but for whom is it good? In what context? With what consequences if one is mistaken? And to this question, the model does not respond, because it requires a form of knowledge that thephilosopher Michael Polanyi described it as tacitin his works on the foundations of scientific knowledge “We can know more than we can tell” he writes), knowledge that lives in the memory of situations, in the accumulated experience of mistakes, in reading the power dynamics of an organization, in sensitivity to the implicit expectations of a jurisdiction or a committee. This knowledge cannot be put into a prompt because it is not something that can be fully articulated, and it naturally resists any attempt at complete formalization.
Amplifier mirror
This is where theempirical data published in January 2026by Anthropic economists provide decisive insight. Among the millions of interactions analyzed with Claude, 52% involve augmentative modes. In other words, the human iterates, adjusts, redirects. They do not delegate, they collaborate. And above all, the authors observe an almost perfect correlation between the sophistication of the question asked and the quality of the answer obtained.
The model is an amplifier mirror; it only produces cognitive value to the extent that the person questioning it is able to formulate what they are looking for. The abundance generated by AI is therefore not a freely accessible resource. It is conditioned by the skill of the person who calls upon it. This is not a technical law, it is a cognitive law.
The operator of abundance
It would be appropriate to call the professional figure that the AI economy requires, and that our training systems do not yet produce, an “abundance operator.” This is neither a model programmer nor a passive user. It is someone who knows how to formulate a problem in terms exploitable by a model, evaluate the relevance and reliability of an output, inject the context that the machine cannot have, and take responsibility for a decision based partly on suggestions that they themselves did not generate.
This skill is not technical in the narrow sense. It is deeply intellectual. It requires having been trained in critical thinking, argumentation, and bias identification. It requires having experienced being wrong and having learned to detect why. In a word, it requires having been exposed to the difficulty of the tasks one now delegates, before delegating them.
The silent risk of blind delegation
That is where thereport published on March 5, 2026Anthropic economists Maxim Massenkoff and Peter McCrory introduced a crucial distinction, namely the difference between theoretical exposure to AI, what the model could technically do in a profession, and observed exposure, what it actually does. The former exceeds 90% in legal, financial, and managerial professions. The latter is much lower, for legal, organizational, and institutional reasons. But the gap is narrowing.
And the most worrying warning sign is not unemployment. It is the slowdown in the recruitment of young graduates (22-25 years old) in the most exposed professions. Companies are no longer hiring as many juniors for tasks that AI now performs faster and cheaper. It’s not that AI is eliminating jobs, rather it is contracting the inflow into skilled professions, depriving a generation of whatthe economist Kenneth Arrowcalled thelearning by doing, this slow and irreplaceable process by which one becomes capable of supervising what one has not yet fully understood.
Cognitive deskilling is precisely that. Not losing one’s job, but gradually losing the ability to supervise what one delegates. One delegates writing without having learned to write. One validates outputs without having developed the critical skills that would allow detecting what AI does poorly, its blind spots, its tendency to produce plausible results where truth is required. This “blind delegation” is not a technological inevitability. It is the product of passive augmentation, a delegation without reinvestment in higher-level skills.
When the loop closes without us
This reasoning becomes even stronger with agentic AI, this new generation of systems that do not wait for human validation at each step, but act. They navigate, write, execute code, orchestrate other agents, make intermediate decisions in chains of actions that stretch over hours without human intervention.Jensen Huang announced it at GTC 2026stating that in ten years, every human employee will work alongside one hundred AI agents. The figure may be exaggerated, but the management is not.
With generative AI, the loop always came back to the human. They evaluated an output and decided. With agentic AI, this loop can be closed internally. The operator of abundance must then become what one might call an “objective architect”; no longer “is this text good?” but “did this agent truly understand what I wanted, and were its fifty intermediate actions all legitimate?”
However, theAnthropic datashow that the success rate of models decreases significantly as the complexity of the task increases, precisely where agents deploy the most autonomy. Human supervision is therefore most necessary where it is most difficult to exercise. And it is impossible for those who have never had to do what they supervise.
Urgency of a training policy
Jensen Huang’s thesis is not wrong, it is incomplete. Yes, the deployment of AI generates jobs in the sectors of physical infrastructure. But, like the electrical networks at the beginning of the 20th centuryeIn the century, an infrastructure is not equitably irrigated solely by the virtue of the market. And above all, it says nothing about the effects on the learning trajectories of intellectual professions.
What the data strongly suggest is the urgency of a policy for training in AI operational skills, not learning to code models, but learning to think with them, to question them, to critique their outputs, to keep alive the skill that one delegates. To train operators of abundance. This task primarily falls to higher education institutions, and it is urgent because the contraction of the inflow of young people into exposed professions leaves little time before learning becomes structurally impossible.
Because what AI cannot do on our behalf is precisely to decide that a particular output is worth something, in this context, for this person, with these stakes. This judgment is not a technique. It is a form of responsibility and it remains irreducibly human.
Bruno Deffains does not work for, does not advise, does not own shares in, does not receive funds from any organization that could benefit from this article, and has declared no other affiliation than his research institution.
–ref. AI can produce everything… but not yet judge: why we are becoming operators of abundance –https://theconversation.com/lia-can-produce-everything-but-not-yet-judge-why-we-are-becoming-operators-of-abundance-280207