David Deutsch · "A new way to explain explanation"
The ancient Greeks explained the seasons like this. Hades, god of the underworld, kidnaps Persephone, goddess of spring, and negotiates a forced marriage contract requiring her to return regularly. Each year she is magically compelled to go back; her mother Demeter, goddess of the earth, is sad, and makes it cold and barren.
On a stage in Oxford in July 2009, David Deutsch first points out that this myth is testable: if winter is caused by Demeter's sadness, it must happen everywhere on earth simultaneously. "So if the ancient Greeks had only known that Australia is at its warmest when Demeter is at her saddest ..." — laughter — "they'd have known that their theory is false."
Then he does something sharper. He rewrites the myth on the spot. Persephone wasn't released; she escaped, and returns every spring to take revenge on Hades, cooling his domain with spring air, venting heat up to the surface, creating summer.
The new version accounts for the same phenomena and is equally testable — yet what it asserts about reality is close to the opposite. How is that possible? Because the details of the original myth have nothing to do with seasons except by way of the myth itself.
That is his whole argument. What pulled humanity out of a hundred thousand years of stagnation was not the rejection of authority (authorities had been rejected many times without producing a scientific revolution), not the claim that knowledge comes from the senses (you don't see equations carved on the mountainsides), and not the habit of making testable conjectures — any crank claiming the sun will go out next Tuesday has a testable prediction.
The real dividing line is that good explanations are hard to vary. The axis-tilt account of seasons is good because every detail is carrying weight: we know, independently of seasons, that surfaces tilted away from radiant heat are heated less, and that a spinning sphere, in space, points in a constant direction. The tilt also explains the sun's elevation through the year and predicts that the seasons will be out of phase in the two hemispheres. The myth is different: had the Greeks learned about Australia, one sentence would have fixed it — Demeter banishes the heat into the other hemisphere. So being refuted and duly amending your theory is worth nothing by itself. Testability only starts to matter once an explanation is already good.
The talk lands on something blunt. Science is blighted by two bad roads: one famous, untestable theories; the other more important, explanationless ones. Whenever you are told some statistical trend will continue but given no hard-to-vary account of what causes it, you are being told a wizard did it.
What happened next? The demolition held: nobody today argues that testability alone separates science from myth. Thepositive half is still in mid-air. A 2020 arXiv paper that set out to turn the criterion into something computable opens by conceding it has not been formalized, and digs up a circularity: to call an explanation hard to vary you must first fix what it could be varied into, and unless that range is restricted, any explanation can be varied into any other. Others argue it is simply simplicity under a new name. Seventeen years on it has not entered philosophy textbooks — but it has entered machine learning: a 2025 paper on evaluating AI interpretability lists the "Deutschian" view alongside the Bayesian and the Kuhnian as one of four lenses. And there is a more awkward destination. Andreessen's 2023 Techno-Optimist Manifesto names Deutsch a "Patron Saint," quoting him on our duty to be optimistic — a line with nothing to do with the criterion at all. A manifesto listing fifty-six saints, any of whom could be removed without changing a word of the conclusion, is by this talk's own standard easy to vary.
He invents a Greek myth on stage.
A minute earlier he was retelling the real one: Hades kidnaps Persephone, negotiates a forced marriage contract requiring her to return regularly, and lets her go; each year she is magically compelled to return, her mother Demeter is sad, and the earth turns cold and barren. He stops to point out that the myth is testable — if winter is caused by Demeter's sadness, it must happen everywhere on earth at once. "So if the ancient Greeks had only known that Australia is at its warmest when Demeter is at her saddest ..." The room laughs. He waits for it to die down before finishing: "they'd have known that their theory is false."
Then he reaches into the same myth and rebuilds it. Persephone wasn't released — she escaped, and returns every spring to take revenge on Hades, cooling his domain with spring air, venting heat up to the surface, creating summer.
The new version accounts for the same phenomena and stands up to the same tests, yet asserts something close to the opposite. That is not a laugh line; it is a blow to the head. If two contradictory stories both explain the same thing and both survive testing, then testability is not worth what everyone assumed.
David Deutsch was born in Haifa, Israel, on 18 May 1953. His title today is visiting professor in the Department of Atomic and Laser Physics, at the Centre for Quantum Computation in Oxford's Clarendon Laboratory — a strikingly light billing for a man widely credited with founding an entire field.
In 1985 he published "Quantum theory, the Church–Turing principle and the universal quantum computer" in Proceedings of the Royal Society A, volume 400, pages 97–117, defining the universal quantum computer for the first time; quantum computation as a discipline is usually dated from that paper. In 2023 he shared the Breakthrough Prize in Fundamental Physics with Charles Bennett, Gilles Brassard and Peter Shor. The man on the Oxford stage is not a professional philosopher but a physicist who has come to talk about epistemology — the study of knowledge itself — and he does not regard that as trespassing.
The occasion is worth noting too. TEDGlobal 2009 ran 21–24 July in Oxford under the title The Substance of Things Not Seen. Deutsch spoke on the 23rd. His first sentence is about ancestors looking at the night sky and wondering how to explain what they saw in terms of things unseen; his last is that this fact about truth is itself unseen. He pinned the entire talk to the conference theme.
More important is where the idea stood that year: it was unpublished. The Beginning of Infinity appeared from Allen Lane on 31 March 2011, twenty months later. The talk is the skeleton of that book's first chapter. What the room heard was the public premiere of the criterion.
He opens on a very long stagnation. For a hundred thousand years our ancestors also wanted to know how to keep their food supply from failing and how to rest when tired without risking starvation. "I bet those prehistoric cave artists would have loved to know how to draw better." Laughter — they wanted what we want, and they almost completely failed to get it. Then the contrast: the first clue to the origin of starlight arrives as late as 1899, radioactivity, and within forty years physicists have the whole explanation, delivering along the way the transmutation of elements that had eluded alchemists for centuries. So what were people doing for the first time in those forty years? He calls this "the most important universal truth that it's possible to know," notes that worryingly there is no consensus about what it is, and adds: "So, I'll tell you." Big laugh.
What follows is a chain of eliminations — that is the talk's real structure. Not the rejection of authority: authorities had been rejected many times before and that rarely, if ever, caused anything like the scientific revolution. Not empiricism either: knowledge comes from the senses — in what language? "Look at the world. You don't see equations carved on the mountainsides. If you did, it would be because people had carved them. By the way, why don't we do that?" Laughter. "What's wrong with us?" More laughter. The serious reason is that scientific theories explain the seen in terms of the unseen, and the unseen does not reach us through the senses — the clinching evidence for curved space-time was a photograph of an eclipse, a dot there rather than there; the evidence for evolution, some rocks and some finches; parallel universes, again dots there rather than there on a screen. Our connection to reality, as Popper put it, is always theory-laden.
And not testable conjectures. This is the step that catches everyone, because it sounds like science itself. His counterexample takes one sentence: any crank claiming the sun will go out next Tuesday has got a testable prediction. So he brings out Persephone and performs the rewrite. It works because the details of the original myth are unrelated to seasons except via the myth itself — and if so, infinitely many other ad hoc entities could play the role the marriage contract plays.
This easy variability is the sign of a bad explanation, because, without a functional reason to prefer one of countless variants, advocating one of them in preference to the others is irrational.— David Deutsch, A new way to explain explanation, 2009
The positive example follows, and it is his only prop of the evening: the tilt of the Earth's axis. His slide carries the line "Not to scale!", which he reads out — "Better put that up" — for one more laugh.
He walks through what each detail is doing: we know, independently of seasons, that surfaces tilted away from radiant heat are heated less, and that a spinning sphere, in space, points in a constant direction; the tilt further predicts that the seasons will be out of phase in the two hemispheres. Then the sharpest cut of the evening. Had the Greeks learned about Australia's seasons, they could easily have varied the myth to predict them — when Demeter is upset she banishes heat into the other hemisphere. So being proved wrong by observation and changing the theory accordingly still wouldn't have got them one jot closer to understanding seasons; whereas if axis-tilt were refuted, its defenders would have had nowhere to go. That inverts the usual common sense: it's only when an explanation is good that it even matters whether it's testable.
The landing is aimed at someone. Science is blighted by two false approaches: one well known, untestable theories; the other more important, explanationless ones. Whenever you are told that some existing statistical trend will continue but aren't given a hard-to-vary account of what causes it, you are being told a wizard did it. When someone announces the nature-nurture debate is settled because a given percentage of our political opinions is genetically inherited, but doesn't explain how genes cause opinions — "They're saying that our opinions are caused by wizards, and presumably, so are their own." Then the close:
That the truth consists of hard-to-vary assertions about reality is the most important fact about the physical world. It's a fact that is itself unseen, yet impossible to vary.— David Deutsch, A new way to explain explanation, 2009
This was not a new idea in 2009; it had been in the field for twelve years. In The Fabric of Reality (1997) Deutsch fights the same fight, but there the enemy has a name: instrumentalism. He describes the opposing position in the book — to them the basic purpose of a scientific theory is not to explain anything but to predict the outcomes of experiments; any consistent explanation a theory may give is as good as any other, or as good as no explanation at all, so long as the predictions are true; explanations are "mere psychological props: a sort of fiction which we incorporate in theories to make them more easily remembered and entertaining." The book's ballast is five words: But prediction is not explanation. The minute on stage about explanationless theories is the compressed version of a twelve-year war — and what got compressed out includes the naming of names: in the book he writes that the Nobel laureate Steven Weinberg was "in instrumentalist mood" when he made a particular remark. On stage, not one name.
The half-sentence that is actually his entire worldview. Look again at the evidence chain: "And parallel universes? Again: dots there rather than there, on a screen." The tone is identical to "some rocks and some finches" — as casual as settled fact. Behind that half-second sit whole chapters of The Fabric of Reality: from a four-slit interference experiment he forces out the conclusion that the interfering entities have passed through the other slits, and calls these invisible interferers shadow photons — undetectable, knowable only through their interference effects on the tangible ones. The book even counts them. The largest area you could conveniently illuminate with a laser is about a square metre, and the smallest manageable hole about a thousandth of a millimetre, so there are about 10¹² (one trillion) possible hole-locations on the screen; therefore at least a trillion shadow photons accompany every tangible one. That is the full price of the half-sentence.
And this is not the consensus of physics — the stage doesn't say so. From 3 to 7 July 2011, at a conference on quantum foundations held in a former monastery on Lake Traunsee in Austria, Maximilian Schlosshauer, Johannes Kofler and Anton Zeilinger handed 33 participants (physicists, philosophers, mathematicians) a questionnaire of 16 multiple-choice questions. On "What is your favorite interpretation of quantum mechanics?" the results were: Copenhagen 42%, information-based approaches 24%, Everett (many worlds and/or many minds) 18%, objective collapse 9%. The limits matter: 33 people is a snapshot of one conference, not a census, and the authors themselves note that the particular set of participants shapes the outcome. Even so, 18% is not consensus — and on stage, "parallel universes" is said in the voice of something everybody knows.
One more qualification goes unsaid, and it is the source of all the later trouble: "hard to vary" is a comparative, not a threshold. How hard is hard enough? Which alterations count as producing a "variant" at all? In sixteen minutes, neither question is touched.
PARTLY HELD UP · as of 2026-09
The demolition held: that testability alone cannot separate science from myth is barely disputed today. The positive criterion stalled halfway — the most serious attempt to formalize it opens by conceding it has not been formalized, and unearths a circularity; a second line of argument holds it may be simplicity under another name. And on its furthest-travelling journey, what got carried off wasn't the criterion at all.
The Beginning of Infinity appeared on 31 March 2011. The following year, in Philosophy Now issue 89, the former Royal Society vice-president Sir Alistair MacFarlane praised it as "a coherent and comprehensive exposition of Karl Popper's philosophy" while delivering a heavy verdict on its second half: when Deutsch hits his stride contemplating our infinite future development, "conjectures proliferate but refutations fade away." Popper's own vocabulary, turned back on Popper's most famous student.
16 December 2020. Daniel C. Elton posted "Applying Deutsch's concept of good explanations to artificial intelligence and neuroscience" to arXiv. It is a sympathetic paper — the author wants to turn the criterion into something a machine-learning researcher can compute — and that is exactly why its candour carries weight: it states up front that so far the hard-to-vary principle has not been formalized. The circularity it then uncovers is worse. To say an explanation is hard to vary, you must first treat it as a member of an equivalence class of possible explanations it could be varied into; and, the paper notes, that class must be restricted, otherwise every explanation can be varied into any other (with enough possible additions and subtractions) and all explanations are equally variable. But restricting the class means settling in advance which alterations count — which is very nearly the same as already knowing what counts as an explanation. The work the criterion was supposed to do has been quietly smuggled into its premises.
20 January 2021. Richard Ngo published "Deutsch and Yudkowsky on scientific explanation" on LessWrong, making a different objection: hard-to-vary sounds like a new criterion but is really an implication of simplicity. A version of "the gods did it" that could actually predict a given set of data would have to encode all that data, so it is not simple at all; in Kolmogorov terms, bad explanations were already complex explanations and no new word is needed. Deutsch had pre-empted the move: he grants that models with more parameters are more variable but refuses the converse, since "there are plenty of simple explanations that are highly variable, such as 'Demeter did it'." The argument has not been settled.
It never entered philosophy textbooks; it entered machine learning instead. On 2 May 2025 Kola Ayonrinde and Louis Jaburi published "Evaluating Explanations: An Explanatory Virtues Framework for Mechanistic Interpretability" on arXiv, setting the "Deutschian" perspective alongside the Bayesian, the Kuhnian and the Nomological as one of four lenses for judging interpretability research. A criterion delivered on a TED stage in 2009 became, sixteen years later, a tool for deciding whether a given account of a neural network is a good explanation. That is probably its healthiest destination.
The most awkward one is dated 16 October 2023. Marc Andreessen published "The Techno-Optimist Manifesto," listing David Deutsch among the "Patron Saints of Techno-Optimism" and quoting him: "We have a duty to be optimistic. Because the future is open, not predetermined and therefore cannot just be accepted: we are all responsible for what it holds." That line has nothing to do with hard-to-vary. What was carried off was Deutsch's optimism, not his criterion — and the entire force of this talk goes into teaching people how to tell whether an assertion is worth believing. A manifesto that lines up fifty-six patron saints, any one of whom could be struck out without moving the conclusion a millimetre, is by the talk's own standard easy to vary. Not refuted. Circumvented.