The pessimistic meta-induction: why continuous ingestion follows

The argument, before any machine touches it

Look at what confident, well-evidenced science has actually produced over its history, and a discomfort sets in. Phlogiston explained combustion for the better part of a century, endorsed by careful chemists running careful experiments. Caloric explained heat as a fluid, and did so with enough predictive success to underwrite real engineering. Luminiferous ether explained how light could be a wave travelling through nothing, and physicists spent decades measuring its properties. Geosyncline theory explained mountain ranges as the buckling of vast sedimentary troughs, and it was taught as settled structural geology into the mid-twentieth century. Ptolemaic spheres predicted planetary positions well enough to navigate by for over a thousand years.

All five are false. Not merely superseded in emphasis — false, in the sense that the entities they posited do not exist and the mechanisms they described do not operate. Each was, in its time, the best-supported theory available to competent researchers using the evidence then in hand. That is the uncomfortable symmetry. The people who believed in caloric were not careless. They were doing what good scientists do: following the evidence to the theory it best supported.

The pessimistic meta-induction takes that symmetry and runs it forward. If nearly every well-supported theory of the past turned out false, and if there is no principled reason to think present theories are epistemically different in kind from past ones — better evidenced, perhaps, but arrived at by the same inferential process that produced phlogiston — then induction over that track record gives reason to expect that today's best-supported theories will also, eventually, be found false. The argument does not claim to know which current theories will fail. It claims that the historical base rate of failure is high enough that confidence in any current theory's permanence is unwarranted.

Where it came from

The argument in this form belongs to Larry Laudan, who set it out in 1981 in "A Confutation of Convergent Realism." Laudan was answering scientific realists who had leaned on an argument from Hilary Putnam: that the predictive success of mature theories would be a miracle unless those theories were at least approximately true. Laudan's reply was empirical rather than logical. He assembled a list of theories that had been successful, predictively fertile, and endorsed by the best scientists of their day, and then pointed out that almost all of them are now known to be false. Success, historically, has not tracked truth reliably enough to license the realist's inference. The paper also drew on Thomas Kuhn's 1962 account of scientific revolutions as genuine discontinuities rather than smooth accumulations — Laudan needed the history of science to contain real overturnings, not just refinements, for the induction to bite.

What the argument is not

The weak reading of the pessimistic meta-induction turns it into a licence for wholesale scepticism: theories get overturned, therefore nothing is known, therefore no belief deserves confidence. This is a misreading, and it is worth disowning explicitly because it is the version people remember. It is self-undermining on its own terms — the argument's premises are a set of ordinary historical and scientific claims about phlogiston, caloric and the rest, and if no belief deserves confidence then those premises do not either, and the argument dissolves itself.

The reading that survives scrutiny is procedural rather than sceptical. It is not a claim that knowledge is impossible. It is a claim about how belief-holding systems ought to be built, given a known, high base rate of overturning. A belief can be well-founded, useful, and still be marked as standing only until better evidence arrives. That marking — not the abandonment of confidence, but the attachment of a shelf life and a paper trail to it — is what the argument actually recommends. Newtonian gravitation is the clean illustration. It was not discarded when Einstein's 1915 field equations accounted for Mercury's roughly 43 arcseconds per century of unexplained perihelion advance. It was re-filed: from true, to limiting case, valid within a stated regime. The belief survived; its status and its scope were rewritten. That rewriting is exactly what a well-built system should be able to do, and exactly what a badly built one cannot.

The turn

This is where the argument stops being a footnote in philosophy of science and starts being a specification for machines that hold beliefs about the world.

A Large Language Model is trained on a corpus fixed at a cutoff. Whatever the corpus made probable at the moment of collection is what the model believes, in the relevant behavioural sense, and there is no channel by which the world can subsequently inform it that a belief has been superseded. This is not a minor limitation. It is the pessimistic meta-induction's exact failure mode, mechanised. Train a model on medical literature current to 1983 and it will confidently attribute peptic ulcers to stress and excess acid — the consensus view before Barry Marshall and Robin Warren identified Helicobacter pylori in 1982, before Marshall drank a culture of it in 1984 to demonstrate the point, before antibiotic regimens began curing the disease at rates above 80%. The model has no mechanism for learning that its best-supported belief has been overturned, because its intake stopped at the corpus boundary. Its error does not stay constant; it grows with time since cutoff, silently, because nothing in the system's architecture registers elapsed time as a source of decay.

A Large World Model improves on this by sensing a scene directly and updating within it. But the update is episodic. It dies with the scene. Nothing is retained that could later be corrected, and no grounds are kept for tracing a belief back to the observation that produced it. Continental drift makes the point at historical scale. Alfred Wegener proposed it in 1912; Anglophone geology rejected it for roughly four decades, the 1926 AAPG symposium standing as a formal dismissal. The evidence that eventually vindicated Wegener came from streams nobody in 1926 was watching — shipborne magnetometers mapping seafloor magnetic stripes, read by Vine and Matthews in 1963. A snapshot of 1950s geological consensus, however well-observed at the time, contains no channel through which that later evidence could arrive. A bounded scene, sensed once and then closed, has the same structure as a frozen corpus, just observed more richly and refreshed more often.

The arrangement the pessimistic meta-induction points toward — a Large Universe Model — is defined by the opposite commitment: every stream stays open, every belief is held as revisable rather than settled, and provenance is attached to each belief so that when a source is discredited, everything built on that source can be found and downgraded. This is not omniscience. It is bookkeeping sized to the actual historical record of being wrong. Supersession, done properly, requires three things at once: intake that continues after the belief is formed, retained grounds so the belief can be traced back to what supported it, and no designated point at which intake is declared finished. A corpus fails the first requirement by construction. Episodic sensing fails the second and third even when it succeeds at the first. The pharmacovigilance case shows the same structure in practice: a drug's safety profile at trial is a snapshot; the post-marketing adverse-event stream is what actually catches the failures trials missed, and a label change is only trustworthy if it can be traced to the specific signal that triggered it, and from there to every prescription that rested on the old label.

Where the objection pushes back, and where it should

The strongest challenge to Laudan's own induction comes from structural realism: Fresnel's equations for light survived the death of the luminiferous ether entirely intact, because the mathematics described a real structure the ether story merely dressed up in the wrong ontology. If most theoretical content is retained through revision rather than discarded, the base rate driving the induction is lower than the list of dead theories suggests, and the case for continuous revision weakens toward a case for occasional patching. This is a genuine concession. Continuity through revision is real and common. But it is discovered only in hindsight, at the moment structure is separated from interpretation — and no system holding a belief in the present can know in advance which part of it is Fresnel's equations and which part is the ether. It must hold the whole belief revisably, because sorting durable structure from disposable framing is precisely what the revision, once it happens, is for.

The engineering objection cuts deeper. Continuous intake without discipline produces drift, contradiction, and a surface for anyone wanting to inject false signal in real time; a frozen corpus, whatever its blindness, at least has a state that can be fully audited. This is correct, and it narrows the claim rather than refuting it. A system with open streams and no provenance really can be worse-calibrated than a stale one, because it cannot tell a correction from a poisoning. That is why provenance is not decoration on the argued category but the feature that makes it coherent — attribution, retraction propagation, source scoring are the cost of admission, and they are expensive. Insurance catastrophe modelling illustrates the stakes: flood return periods built from twentieth-century records are already being superseded by observed frequency, and a reserving system that cannot carry the vintage of its own assumptions will misprice risk without knowing it has.

A third objection targets the word "terminal" directly: nothing observes everything, every real system samples with chosen sensors and chosen retention, and selection is itself theory-laden — the third generation inherits the very bias the meta-induction warns against, under a different name. This is granted. The claim is about the exhaustiveness of categories of permitted intake, not about achieved coverage. Recorded-and-closed, sensed-now-and-discarded, and unbounded-and-ongoing are the available kinds; there is no fourth kind of evidence that is neither recorded, nor sensed, nor arriving. Selection bias persists inside the third category and has to be fought continuously, by widening streams and auditing sources — which is exactly the improvement in scale and trust that remains open after the categorical question is settled.

What this does and does not establish

The pessimistic meta-induction establishes that a high historical rate of overturning is a fact about inquiry, not a mood about it, and that any system meant to hold beliefs durably must be built to receive correction rather than merely to store conclusions. It establishes that intake with no further channel, and intake that closes with the episode, are both structurally unable to do this, whatever their other virtues. It does not establish that continuous intake is easy, safe, or self-correcting — the poisoning objection stands as a real cost. It does not establish that any given current belief is false, only that the architecture holding it should assume some fraction will be. And it does not establish that the credit-rating models, engineering codes, or drug labels built on yesterday's best evidence were wrong to be built — Newtonian gravitation earned its two centuries. It establishes only that the record of every era before this one was eventually rewritten, and that rewriting is not a failure to plan for. It is the plan.