The Reverse Turing Test
Understanding the Turing Test
In 1950 Alan Turing published a paper in Mind that opens with an impossible question, “Can machines think?”, and dismisses it within half a page as too meaningless to deserve discussion.
In its place, he proposes a game. One interrogator, two rooms, a conversation conducted at a distance: if the machine can pass for the human being, the philosophical question evaporates and the result is allowed to stand. Turing called it the imitation game, history renamed it the Turing test, and for seventy-five years it stood as the finish line of computer science: build a machine capable of fooling us.
Turing even risked a prediction: by the end of the century, a machine with a billion bits of storage would deceive an average interrogator thirty percent of the time, after five minutes of conversation. He got the date wrong, not the direction.
The first incident arrived ahead of schedule. 1966, MIT: Joseph Weizenbaum wrote ELIZA, a few hundred lines of code impersonating a Rogerian psychotherapist by turning the patient’s sentences back into questions. It understood nothing, literally. Yet Weizenbaum’s secretary, who had watched the program being built and knew exactly what it was, asked him one day to leave the room: the conversation had become private. Weizenbaum was shaken enough to spend the rest of his career warning against his own profession. The ELIZA effect entered the vocabulary: our irresistible tendency to attribute understanding where there is only pattern matching. The Turing test, it turns out, does not measure how good machines are. It measures how willing we are.
Then, for half a century, almost nothing. Trade-fair chatbots, Loebner prizes won with parlor tricks, a very long adolescence.
And finally the overtaking, which happened without ceremony. Since November 2022 we have been conversing daily with systems that would have made Turing’s interrogator faint, and no headline ever really marked the moment. In 2025 a study at the University of California San Diego rebuilt the original three-party protocol: GPT-4.5 was judged human 73 percent of the time, more often than the actual humans in the room.
The seventy-five-year finish line, crossed and filed away as a feature. We were too busy having it rewrite our emails.
Which brings us to the reversal. Since 2 August 2026, Article 50 of the AI Act requires that anyone interacting with an artificial intelligence system be informed of it, that synthetic content carry a machine-readable mark, that deep fakes be disclosed.
Fines run up to fifteen million euros or three percent of worldwide turnover. For three quarters of a century we rewarded machines for passing as human; now European law requires machines to declare themselves machines.
The Turing test, flipped and published in the Official Journal: no longer something to pass, but something whose passing must be reported.
Inside Article 50 sits the most philosophical exemption ever written into a regulation: the obligation falls away when the artificial nature of the system is obvious to a person who is “reasonably well-informed, observant and circumspect”.
Obviousness as a legal category. Turing designed an experiment to establish whether a machine could pass for human; the European legislator takes the answer as settled and shifts the burden onto the observer: exactly how circumspect must a human be, in 2026, for the deception to count as harmless?
The Commission’s guidelines, fifty-one pages adopted on 20 July, attempt an answer with the slightly surreal seriousness of an institution tasked with regulating the obvious.
And then there are the stickers. The Commission has even published an official label kit, free to download: AI GENERATED, AI MODIFIED, black pills in white type that look like they came off a cigarette pack, or off the Parental Advisory stickers that guaranteed the commercial success of any record in the Nineties.
There is something touching in this faith in signage: the idea that the boundary between human and artificial, which seventy years of philosophy of mind failed to draw, can be patrolled by an adhesive label.
Turing measured intelligence by successful deception. Europe governs it by forbidden deception. In between stand the rest of us, no longer asking whether a machine can pass for human, but whether we still care to notice.
Judging by the way we greeted the overtaking, the answer may already be in, as quiet as everything else.
For the record: this essay falls under the exemption for content subject to human editorial control. The author has read it, fixed it, and takes responsibility for it. No sticker required. For now.
