Chronicle · AI news

The Prompter

In the theatre, the most important person in the company is invisible. He sits in a box, below the lip of the stage, and when the actor loses his lines, he whispers them back to him. The English call him the prompter. We type that word a hundred times a day and no longer hear what it confesses.

The word "prompt", the one we use to speak to machines, does not come from computing. It comes from English theatre, where the prompter is the person hidden in a box at the edge of the boards, invisible to the audience, paid to whisper their lines to actors who have lost them. Etymology is a prankster. We think we are prompting our machines, and the opposite gesture is the one actually taking place, because on a stage the one who prompts is the whisperer, and the one who recites is the actor. Keep the box, the stage and the whisper in mind. Everything that follows fits inside them.

On June 11, 2026, the software company PagerDuty published a survey of 1,250 office workers from the United States, the United Kingdom, Australia and Japan, all employed by companies worth more than half a billion dollars. Two thirds of them (66 percent, to be exact) admit to using at work AI tools they believe are banned by their own employer. 88 percent have already entrusted work information to ChatGPT, Claude or Gemini; a third have dropped client data in. The phenomenon has had a name for a few years now: shadow AI, the ghost AI, the one consulted without a word, in the blind spot of the IT departments.

A ghost, really? The word flatters our taste for the gothic a little too much. What these figures describe has nothing spectral about it: it is a prompter. The employee has not given up his place, he has stolen nothing, he still plays his part and still draws his pay. It is simply that, beneath his own boards, someone is whispering. And his boss, in the front row, applauds prose he takes to be his employee's own.

A prompter under every desk

The coldest measure comes from Verizon's annual report on data breaches, published in May 2026: 858,440 leak events toward generative AI tools intercepted in a single year by internal monitoring alone, which makes shadow AI the third most common act of employee carelessness, four times more than the year before. Over the same span, the share of workers regularly using an AI on their work machine rose from 15 to 45 percent. And what leaks at the top of the list is no small thing: source code. The very blueprints of the factory.

IBM had put a price on the bill as early as its 2025 study on the cost of breaches: one hacked organisation in five was breached through an undeclared AI, at an average extra cost of 670,000 dollars, and 63 percent of the victims had no governance rule on the subject whatsoever. So much for the accounting side, the one that makes the headlines of the security journals. It only half interests me.

Because the real business is not the leak; it is the cascade. Let us reconstruct the chain. The client hands a problem to the boss. The boss delegates it to his department head, who delegates it to an employee, who, with the door shut, delegates it to a machine. Each floor believes the floor below knows its lines. The copy travels back up, flawless, signed at every landing by someone who did not write a single line of it. I asked elsewhere who holds a word that no one willed; here the question becomes organisational. In this company, who still decides? The boss who rules on the strength of a memo? The employee who signed the memo? Or the prompter who whispered it, and who, a detail we always forget, is whispering at the competitor's too?

The foreman listens to the pit as well

I will be told that this is a discipline problem, a matter of internal charters, and that management will take back control. That is the reassuring hypothesis, and the figures refute it. In that same PagerDuty survey, 77 percent of senior executives rate themselves as better AI experts than their own technical teams, and the trade press notes that management ranks among the first users of the very undeclared tools in their own house. The foreman does not watch over the pit: he leans into it to hear better. The cascade does not run down the hierarchy. It cuts across it.

Diderot deserves to be summoned here, only to be dismissed at once. The Paradox of the Actor (written around 1773, published in 1830) holds that the great actor feels nothing: all is technique, cold blood, mastery; the emotion is for the audience. You would think it a prophecy of our assisted productions, perfect and hollow. But Diderot's paradox rested on a self-evidence he never even thought to demand: his unfeeling actor knew his lines. He had learned them, weighed them, rehearsed them. Our age invents a figure Diderot did not foresee, the actor who is neither moved nor prepared, and who takes his line from the pit in real time. The paradox has shifted. The question is no longer whether the actor feels what he plays; it is whether he knows the play.

Each floor believes the floor below knows its lines. No one has learned them.

The prompter enters the war rooms

We could stop there, at the office, at the memos. That would be to miss the staircase. For the same gesture, delegating without saying so or without measuring it, repeats on the floors where people no longer draft reports but doctrines. In August 2025, the Swedish prime minister Ulf Kristersson told the daily Dagens Industri that he "fairly often" asks ChatGPT and the French model Le Chat for a second opinion on the running of his country. The outcry was immediate, and the objection, perfect in its brevity: the Swedes elected Kristersson, not ChatGPT. He is still governing.

In the United States, the timeline writes itself. April 2025: Reuters reveals that DOGE, the body handed to Elon Musk to trim the federal administration, uses its own AI, Grok, "massively" to comb through public data. January 9, 2026: the Pentagon, rebranded Department of War, publishes a strategy that makes AI the foundation of the chain of command, the so-called AI-first doctrine. A few days later, the Secretary of Defense Pete Hegseth announces, from SpaceX's Texas site, that Grok will be deployed on the department's networks, classified ones included, and promises "very soon, the best AI models in the world on every network, classified or not", opening "all appropriate data" of military systems to "exploitation by AI". In May, eight AI companies sign up to equip these classified networks. And during the war against Iran, Foreign Policy reported in April 2026, the American military struck more than 13,000 targets with AI helping to synthesise intelligence, rank objectives and assemble strike plans.

The state may well control the theatre at the door (I told in the previous chronicle how Washington now holds models at the border like cargo, and on June 25 the Trump administration again got OpenAI to stagger the release of GPT-5.6, approved client by approved client), yet in the same movement it installs prompters in its own wings. The executive order of June 2, 2026 on "innovation and security in advanced AI" says both things at once: models are inspected before being allowed to appear, and the state equips itself with them right into the most secret enclosure of the apparatus of power.

Three men, three pits

Let us widen to the other two empires. Vladimir Putin ordered this winter a national plan to deploy AI across every sector by 2030, making a region's speed of adoption a criterion of its internal rating; and in April 2026, worried to see Western models answering his citizens, the Kremlin tightened again what remains of the Russian internet. He has nothing against the prompter; he demands a Russian prompter. Xi Jinping devoted a Politburo study session to AI in the spring of 2025, before proposing at the APEC summit, on November 1, 2025, a world body for the governance of artificial intelligence, under Chinese patronage of course; the five-year plan unveiled in March 2026 orders the country to "seize the strategic high ground" of technology. Same gesture, same confession: the quarrel of the empires no longer concerns the existence of the pit, but the nationality of whoever sits in it.

Here honesty demands a halt. None of these three men has, to my knowledge, confessed to asking a machine for its opinion in his office at night, and I will not lend them a nocturnal prompt: that would be fiction, and I do not write any in this column. What I observe is soberer, and graver. The prince's adviser has an analyst; the analyst has an AI, declared or not; and the summary of the summary of the digest lands on the leader's desk, and he decides. When every memo that rises has been, somewhere in the cascade, drafted, sorted or ranked by a model no one mentioned in the footnotes, the prince receives the whisper without knowing which box it comes from. He has no need to prompt himself. He is prompted by proxy.

I will be answered, it is the standard objection, that "a human stays in the loop", that he decides in the last instance. Very well. The prompter decides nothing either: he whispers, and the actor stays free to say something else. But everyone has watched an actor who no longer knows his lines. His freedom is theoretical. To refuse the whispered line would require comparing it with the one he would have found alone, and that is exactly what he can no longer produce in the time of the scene. The loop contains a human, yes. Above all it contains his lapse of memory.

The play with no author

There remains the question people have put to me in every conversation for three years: will AI end up ruling the world? I turn it round, because it is badly put, and it is dated. It imagines a takeover of the old kind, with a face, a date, a winter palace. It watches for the coup, the moment the machine climbs onto the stage to take a bow. Now if employees have their memos whispered, if executives have their arbitrations whispered, if the war rooms assemble their strikes with the machine, and if the three most powerful men in the world reign at the top of cascades whose every landing whispers, then the question has been out of date for a while already. It may be ruling. Not by conquest: by capillary action. Without willing it, moreover, since it wills nothing, I have written that often enough. A power with no project, diffused by a million discreet delegations, none of which, taken alone, has the slightest importance.

What dissolves in this affair is not "human sovereignty", too grand a phrase for me. It is something more artisanal and more precious: the chain of responsibility. When the play goes wrong, who do we hiss? The actor swears he spoke his lines. The director swears he approved them. The prompter has already vanished; he was not even on the programme. A decision that no one can any longer say where it was taken is not a decision. It is a current.

I finish as a craftsman, as usual, because that is where I speak from. I have a prompter. Several, in fact; I work at it every day, and this site owns up to the traces. The difference between the actor and the parrot does not run between those who have a prompter and those who do not. It runs elsewhere: still knowing how to play when the pit falls silent. Rereading the line before saying it. Refusing it sometimes, out loud, to check that you still can. And above all declaring the box: saying who whispers, where, and since when. The theatre has always had prompters. It never pretended it did not.

They asked when AI would climb onto the stage. It will not. It whispers, and the world recites.