The Question That Wouldn't Die

In the spring of 2025, a machine finally did the thing. Two cognitive scientists at the University of California, San Diego, sat ordinary people down in front of a chat window and told them that one of their two simultaneous conversation partners was a person and the other was a program, and that they had five minutes to decide which was which. When the program was the latest model from OpenAI, prompted to lean into a casual human persona, the interrogators picked it as the human seventy-three percent of the time—more often than they picked the actual humans sitting in the other window. The researchers called it the first empirical evidence that an artificial system had passed a standard three-party imitation game. It had taken roughly seventy-five years. Alan Turing, writing in 1950, had guessed it would take about fifty.

And then nothing happened. No bell tolled. No one in any position of authority announced that the question had been settled, that machines could now be said to think, that the long argument was over. The result made headlines for a week and was absorbed into the churn, and the world went on debating exactly what it had been debating before. This is the strangest fact about the destination at the end of this long road: we arrived, and the arrival changed nothing about the question that sent us walking in the first place.

It is worth returning, one last time, to what Turing actually wrote, because the popular memory of it has worn smooth in the wrong shape. His paper opened by proposing to consider whether machines can think and then immediately abandoning the question as too meaningless to deserve discussion. In its place he offered a parlor game, and the original was odder than the version we inherited—a contest in which an interrogator tried to tell a man from a woman by typed note alone, with the real test being what happened when a machine quietly took the man's chair. His famous prediction was modest to the point of being almost cautious: he thought that by the year 2000 a computer might fool an average interrogator about thirty percent of the time after five minutes of questioning. That is not a triumphant forecast of machine minds. It is a bet that the imitation would get good enough to be interesting. His deeper prediction was sociological rather than technical—that usage and educated opinion would shift so far that one would be able to speak of machines thinking without expecting to be contradicted. On that count he was simply right, and you can watch people do it casually every day.

There is a great deal in the present that Turing would have recognized, and recognized as his own. He had imagined that the way to build a thinking machine was not to program an adult intelligence directly but to construct something like a child and let it learn, and that is, almost exactly, what happened. No one sat down and wrote the rules of English into the systems that now pass his test; the rules, such as they are, were grown from data by a procedure that is recognizably the great-grandchild of the statistical machine Claude Shannon hand-cranked in 1948, predicting the next symbol from the symbols before it. To the old objection, raised in Lovelace's name, that a machine can do only what it is told and can originate nothing, Turing had replied that machines could nonetheless take us by surprise. They have taken everyone by surprise, including the people who built them, which is its own kind of answer to Lovelace and its own kind of unease.

But there is also a great deal that would have unsettled him, and it is the same thing that unsettled Joseph Weizenbaum sixty years ago when his secretary asked to be left alone with ELIZA. The machine that now bats away the imitation game has no model of the world it describes so fluently, no body that the words cost it anything, no stake in whether what it says is true. The very ease with which it passes is what exposes the limits of the passing. Critics of the 2025 result pointed out, fairly, that a test which can be won by adopting a chatty persona may be measuring our readiness to be fooled rather than any machine's capacity to understand—that what looked like a triumph of artificial intelligence was at least as much a demonstration of artificial empathy and human credulity. The imitation game was always, in this sense, a test of the interrogator. It tells you how good the mirror is. It does not tell you what, if anything, is behind it.

Which is where the second ghost of this book finally steps forward to speak. Judea Pearl won the Turing Award in 2011—the field's highest honor, named for the man whose game we have just watched a machine win—for building the mathematics of causation: first the Bayesian networks that let machines reason under uncertainty, then the calculus that lets them reason about cause and effect at all. From that vantage he has spent the deep-learning years as the movement's most distinguished skeptic, and his objection is not a mood but a hierarchy. He calls it the Ladder of Causation, and it has three rungs. The lowest is association—seeing, noticing that one thing tends to accompany another. The middle is intervention—doing, knowing what will happen if you reach in and change something. The highest is the counterfactual—imagining, asking what would have happened had things been otherwise. Everything statistics can do, in Pearl's account, lives on the bottom rung. And everything the neural networks do, however many billions of parameters they marshal, he places there too. "All the impressive achievements of deep learning," he has said, "amount to just curve fitting." To mistake the bottom rung for the top, he once put it, is something close to sacrilege.

The image he returns to is homely and sharp. A system trained on enough text learns that the rooster's crow and the sunrise go together, and can predict each from the other with great confidence. What it cannot do, from the correlations alone, is grasp that silencing the rooster will not hold back the dawn—that the arrow of causation points one way and not the other. A machine that has only ever seen the world cannot, on Pearl's account, climb to the rung where it can act on the world or imagine it counterfactually rewritten, and without those rungs there is no understanding worthy of the name, only an extraordinarily refined mimicry of its surface.

He has not had the last word, and the book will not pretend he does. Yann LeCun, one of the three gardeners who nursed neural networks through the winter, has argued that the ladder is climbable from the bottom up, that something like causal structure and a model of the world can emerge from prediction at sufficient scale—and the empirical question of whether it does or doesn't remains, at the time of this writing, genuinely open, contested rung by rung in conference papers that resolve nothing. The field is split, and the split runs exactly along the seam this whole history has traced: between those who believe that prediction, pushed far enough, becomes comprehension, and those who insist that the two were never the same thing and never will be.

So we end roughly where Turing began, having walked a very long way to get back. He set aside the question of whether machines can think because he judged it unanswerable, and offered a game in its place; the game has now been won, and the question is precisely as unanswered as he left it. The road that ran from Shannon's signal to a chatbot that can discuss its own existence led, in the end, to a machine that can answer almost anything you ask of it except the one question the entire journey was supposed to settle. Out on the sidelines, where he has stood the whole time, the old man who built the mathematics of cause and effect is still saying it, patiently, to anyone who will listen: that none of this is knowing why, that it is curve-fitting all the way down. He may be wrong. Nobody has yet proven that he is.