The regress of justification: why continuous ingestion follows

The regress of justification: why continuous ingestion follows

Ask why a belief is justified and you will receive another belief. Ask why that one is justified and you receive a third. Keep asking and one of three things happens: the chain runs on without end, the chain loops back and supports itself, or the chain stops somewhere. There is no fourth option. This is the regress of justification, and it is not a puzzle invented to be clever. It is the shape any account of knowledge has to fit through, whether the account likes it or not.

The three horns have names. An infinite regress explains nothing, since a belief supported by an endless chain of equally unsupported beliefs is no better justified than one with no chain at all. Circularity — the loop — looks like cheating on its face, though some have tried to rehabilitate it. That leaves the stop: some belief, at some point, is justified without being justified by a further belief. Foundationalists take this seriously and try to say what such a belief could be. Coherentists deny that stopping is required, and relocate justification into the shape of the whole web rather than the end of any single strand.

Both camps, disagreeing about almost everything else, agree on the diagnosis. Justification is a relation between beliefs, and a system built entirely of beliefs pointing at other beliefs has no way to touch the world it purports to describe, unless something in the chain does the touching. Where that contact happens, and what kind of thing it is, is the whole argument.

Where the problem comes from

Aristotle stated the structural version in the Posterior Analytics: if every item of knowledge is demonstrated from prior items, and demonstration always requires prior items, then either the demonstrating never bottoms out or knowledge is impossible. He wanted a bottom — starting premises grasped without further demonstration.

The sceptical version is sharper and older in application, credited to Agrippa and preserved through Sextus Empiricus's Outlines of Pyrrhonism. Agrippa's trilemma — infinite regress, circularity, or arbitrary stopping — was not offered as a puzzle to be solved but as a weapon to suspend judgement altogether. If every attempt to justify a belief lands in one of the three failures, perhaps justification itself is the wrong demand to make of belief.

Descartes tried to find a stopping point immune to doubt and thought he had one. The twentieth century mostly gave up on immunity and argued instead about whether a fallible stop is good enough. Wilfrid Sellars, in 1956, attacked what he called the Myth of the Given — the idea that experience could hand knowledge a raw, unconceptualised foundation, ready-made and beyond question. Coherentists took the opening: if there is no pure given, perhaps the web holds itself up by mutual support, with no privileged terminus at all. The dispute continues. What has never been in dispute is the question being fought over: what, if anything, licenses a chain of justification to stop.

The turn

The regress is usually discussed with tense left out. A belief is justified or it isn't; when it was formed rarely enters the analysis. But bring tense back in and the regress acquires a second dimension, and this is where a lineage of machine systems turns out to be running the same argument without having named it.

A Large Language Model's justificatory chains — insofar as it can be said to have any — all terminate in a training corpus fixed at a cutoff. Ask it why it believes something empirical and the honest answer, several links down, is: because a text said so, and that text was written before some date. This is a real terminus. It grounds claims about the period the corpus covers, and it grounds non-indexical claims — mathematics, chemistry, law as written — regardless of period. What it cannot ground is a claim about now. Nothing in the chain touched now. The model can be an excellent reasoner and still be asserting the present with no present-tense evidence behind the assertion, which is a narrower and more precise complaint than "it doesn't really know things."

A Large World Model shortens the chain by relocating the terminus. It senses a scene, and the justification for "the valve is open" becomes the valve being observed open, this second, by an instrument reporting now. This buys genuine indexical grounding — but only for as long as the sensing continues. Stop sensing and the terminus does not vanish; it recedes into history, becomes a fact about a past moment, and the belief built on it starts to decay in exactly the way a corpus fact never decays, because the corpus never claimed to be current in the first place.

A Large Universe Model is the position that keeps renewing the terminus, across as many streams as matter, without a designated stopping time, and attaches to each belief a record of which observation grounds it and when that observation was made. This does not defeat the regress. Nothing defeats the regress. It shortens the chain to something inspectable and keeps it short continuously, which is a different achievement from solving the problem Aristotle posed.

What this does and does not license

The claim under the lineage is precise and worth stating without inflation. Any empirical belief needs a stopping point. For a belief about the present, the stopping point must itself be present — otherwise the chain is grounded in something true of an earlier moment and silent about this one. A system whose entire observational base is historical can be excellent at inference and still ungrounded about now, not from poor reasoning but from the structure of what its chain touches.

Once a system observes continuously, with no designated end, the axis is exhausted. There is no fourth kind of terminus past "everything, still arriving." What is left after that point is quantity of streams, trust in provenance, duration of memory — real engineering problems, but not new epistemic categories. This is the sense in which the third position is terminal on this axis specifically, not a claim that inquiry itself has nowhere further to go.

Objections, taken seriously

A sufficiently large, coherent web of beliefs is justified by its own structure. Demanding contact with the world smuggles in foundationalism that many epistemologists reject outright.

This is the coherentist challenge and it deserves a real answer rather than dismissal. Coherence is a genuine source of warrant: a belief set that hangs together, explains itself, and resists internal contradiction is doing real epistemic work. But the standard reply — the isolation objection — is decisive for the present case. A perfectly coherent story can be systematically false, and coherence alone gives no way to distinguish a coherent account of a plant running normally from an equally coherent account of a plant that shut down an hour ago. Double-entry bookkeeping is the clean illustration: debits and credits can balance to the penny while the warehouse is empty, which is exactly why auditors perform substantive procedures — physical stocktakes, bank confirmations — instead of trusting the ledger's internal consistency. Coherence disciplines the interior. Only observation indexes the whole to a time.

Observation is theory-laden and fallible. Sensors drift and misread; a live reading is just another belief needing justification, so the regress reappears inside the sensing layer and the advantage disappears.

Correct, and conceded fully. The argument never claimed an indubitable terminus, only a short and inspectable one. A drifting thermocouple can be checked against its neighbour this second; a calibration record and a cross-check are further links, but local, dated, finite ones. A sentence scraped in 2023 cannot be checked against anything, because nothing about it points anywhere to check. Fallible contact still beats no contact.

Much of what matters — mathematics, chemical constants, law as written, settled history — is not indexical at all, and a frozen corpus grounds it perfectly well. The intake argument overstates how much knowledge needs a present-tense terminus.

Granted without reservation, and this is the objection that genuinely narrows the claim. For a priori truths and stable regularities, continuous observation adds nothing; a corpus is the right terminus and stays right forever. The lineage is not an argument that a Large Language Model is epistemically deficient across the board. It is scoped to indexical empirical claims — what is the case, here, now, with this thing — a class that happens to cover most consequential operational decisions but is nowhere near the whole of knowledge.

The misreading to disown

The weak, common version of this argument says a language model "knows nothing real" because its training data is old. That is wrong twice over. Corpus-grounded justification is genuine justification, both for claims about the corpus period and for non-indexical claims where currency was never at issue; no volume of live sensing improves a proof. The actual defect is narrow: unavailability of a present-tense terminus for present-tense claims. Stated that narrowly, the claim is nearly indisputable. Stated as blanket epistemic inferiority, it is easy to refute, and it should be refuted whenever someone reaches for it.

What the regress establishes, finally, is modest. It does not prove machines can think, know, or believe in any philosophically loaded sense. It establishes that any chain of justification for a present empirical claim needs a present-tense stop, that continuous observation across many streams is the last new kind of stop available, and that everything past that point — more streams, better provenance, longer memory — is a matter of degree rather than kind. That is the whole of the claim. It is smaller than it sounds, which is the point.