Chapter 8 — Machine Abduction

Chapter 8 — Machine Abduction

There is now a Sherlock Holmes for every human being alive, and it never sleeps, and it works for whoever pays it.

It does not live at 221B Baker Street. It lives in server farms, and it performs Holmes's exact trick — clue in, best explanation out — billions of times a second, on a scale that would have stopped the great detective's cold heart. It reads your traces and guesses your secrets: what you'll buy, whether you'll repay the loan, whether the lump is malignant, whether you're about to commit a crime. Machine learning is, when you strip away the mathematics, industrialised abduction — inference to the best explanation, mechanised and mass-produced. And because it is the purest embodiment of Holmes's method ever built, it also embodies, with perfect and terrifying fidelity, Holmes's fatal flaw: it is magnificently, catastrophically confident, and it has no author to keep it from being wrong.

The method, mechanised

Recall the shape of abduction from Chapter 1: a surprising observation, and a leap to the hypothesis that best explains it. Now describe how a machine-learning system works. It is shown a vast pile of examples — images labelled "tumour" or "not tumour," borrowers labelled "repaid" or "defaulted" — and it learns the patterns that best explain the labels. Then, shown a new case, it leaps to the most probable explanation: this shadow is best explained by a tumour; this applicant is best explained by "will default." That is not a metaphor for abduction. It is abduction, executed in arithmetic. The clue goes in; the best explanation comes out.

The deep machinery even bears the name of the minister from Chapter 5. The computer scientist Judea Pearl built much of modern AI on Bayesian networks — webs of probabilities that update belief across many linked clues exactly as Bayes' theorem prescribes, the honest engine running at industrial scale. A medical-diagnosis system reasoning from symptoms to disease is doing, formally and correctly, what Holmes only gestured at over the corpse. For the first time in history, the reasoning of the great detective exists as an actual, inspectable, buildable thing. The dream is real. You can download it.

And here — just as you are ready to celebrate — is where the machine reveals that it has inherited not just Holmes's genius but his disease, and inherited it in a form far more dangerous than a Victorian short story, because the machine's mistakes come with handcuffs.

The oracle that launders prejudice

Pearl himself is the sharpest source of the warning, because he insists on a distinction the machines mostly cannot make: the difference between correlation and causation. A machine-learning model sees that two things go together and leaps to an explanation — but it cannot, on its own, tell whether A causes B, B causes A, or both are caused by some hidden C. It reads the shadow on the wall and does not know it is a shadow. And so it faithfully reproduces every bias woven into the world that produced its data, then hands that bias back to us wearing the white coat of mathematics, laundered into what looks like objectivity. "Garbage in, garbage out" is the old programmer's phrase; the modern version is crueller, because the garbage comes out polished, bearing a probability score, impossible to argue with, and we are psychologically primed — the phenomenon is called automation bias — to trust the machine's confident number over our own doubt.

Watch it happen in the highest-stakes arena there is. Across American courtrooms, algorithms score criminal defendants on their likelihood of reoffending, and judges consult those scores when deciding bail and sentencing. One such tool was called COMPAS. In 2016 the investigative newsroom ProPublica analysed its output and reported that it was racially biased: among defendants who did not go on to reoffend, Black defendants were far more likely than white ones to have been wrongly flagged as high-risk — the machine's false alarms fell disproportionately on one race. The company that made COMPAS fired back with its own statistic: the score was equally calibrated across races — a given score meant the same actual reoffense rate whether you were Black or white — and by that measure the tool was fair.

Here is the part that turns a news story into something close to a mathematical tragedy. They were both right. And a trio of computer scientists soon proved why they were both right, and why the argument could never be settled: it is mathematically impossible, except in trivial cases, for a predictive system to satisfy both notions of fairness at once. You cannot have equal false-alarm rates across groups and equal calibration across groups when the base rates differ. The fairness criteria are not merely hard to reconcile; they are provably incompatible. There is no algorithm, however clever, that is "fair" in all the ways we intuitively demand — the demands contradict each other. Holmes's dream of a perfectly objective reasoning machine runs aground not on bad engineering but on a theorem. The machine cannot be neutral, because neutrality, rigorously defined, does not exist. Somebody has to choose which kind of fairness to sacrifice, and that somebody is making a moral decision while pretending to make a technical one.

The men in the ditch have names again

And when the machine is simply, flatly wrong — not subtly biased but wrong — the consequences are no longer a footnote. On a January morning in 2020, police arrived at the home of a man named Robert Williams, in front of his wife and two young daughters, and arrested him for a theft he did not commit. A facial-recognition system had matched his driver's-licence photo to grainy surveillance footage of the actual thief, and that algorithmic guess — the machine's confident abduction — was treated as fact. Williams, who is Black, spent roughly thirty hours in a cell before the obvious became undeniable: he looked nothing like the man in the footage. His is believed to be the first documented case of a wrongful arrest in the United States caused by a face-recognition match. It was not the last; several others, overwhelmingly Black Americans, have since been wrongly arrested on the say-so of the same category of technology, which study after study has shown to be markedly less accurate on darker-skinned faces.

Look hard at what happened in that driveway, because it is the whole book compressed into a single scene. A system performed a Holmesian inference — this face is best explained as that face — and delivered it with the flat confidence Holmes always had. But there was no Arthur Conan Doyle standing behind the machine to guarantee the guess was right. There was only a man in handcuffs in front of his daughters. This is precisely the horror Part I promised: abduction is the fertile, unsafe leap, and when you industrialise the leap without industrialising the humility — when you build the confidence of Sherlock Holmes but not his one honest whisper of "Norbury" — you get a machine that arrests the innocent at scale and issues each mistake with a probability score attached.

The verdict on the mechanical detective

So the machine is not the villain of this chapter, and it is not the hero. It is the mirror. It shows us that Holmes's method really can be built — that inference to the best explanation, run on enough data with enough arithmetic, is genuinely one of the most powerful forces ever assembled, curing disease and catching killers and reading the world's traces at a scale Locard could not have dreamed. And in the very same motion it shows us that the method, divorced from humility, inherits every flaw of Part I and amplifies them to industrial scale: the base-rate blindness, the confirmation of existing bias, the seductive confidence that outruns the accuracy. The machine is Sherlock Holmes with the certainty dial welded to maximum and the author removed. That is a thrilling thing to have built. It is also, without the missing ingredient, a dangerous one.

The missing ingredient is not more data or faster chips. It is the thing every chapter of the resurrection has been circling: calibrated humility — the discipline of knowing exactly how often you are right, and saying so, and holding your conclusions with a grip proportioned to the evidence rather than to your confidence. It turns out some human beings have built precisely this discipline, without any machine at all, and they quietly out-predict both the algorithms and the geniuses. They are the true heirs of the science Holmes only dreamed of, and they are the least Holmes-like people you will ever meet. We turn to them now.