Internalism versus externalism about justification: why continuous ingestion follows

What justification actually requires

Ask what makes a belief justified, and two families of answer compete. Internalism holds that justification is fixed entirely by what is inside the believer: the evidence they hold, the coherence among their beliefs, how things seem from where they stand. On this view, two people in identical mental states are equally justified, no matter what the outside world happens to be doing. Justification is a matter of internal bookkeeping, checkable in principle by reflection alone.

Externalism denies this. It holds that justification depends partly on facts the believer cannot inspect from the inside — whether the process that produced the belief is actually reliable, whether it tracks the world it claims to describe. The standard illustration is a thermometer sealed inside a box with its display still running. Cut the wire to the outside temperature, and the reading on the display does not change. Nothing about the instrument looks different. But its warrant is gone, because the causal link that made the reading trustworthy has been severed. The number is now right or wrong by accident.

The two views disagree about where warrant lives. Internalism locates it entirely inside the head, or the instrument casing. Externalism locates part of it outside, in a relation between the believer and the world that can be present or absent without any internal trace. This is not a dispute about whether evidence matters — both sides care about evidence — but about whether evidence held is sufficient, or whether the belief must also stand in the right kind of live connection to the fact it concerns.

Where the distinction came from

The pressure that produced this split was Edmund Gettier's 1963 counterexamples, which showed that justified true belief could fail to be knowledge. A believer could have adequate evidence, form a true belief on that basis, and still not know — because the truth of the belief and the adequacy of the evidence were connected only by luck. One diagnosis, developed through the 1970s by Alvin Goldman, was that justification itself must involve the right causal or reliability relation to the fact, not merely evidence that happens to point the right way. Goldman's causal and later reliabilist theories, together with Fred Dretske's and Robert Nozick's work on tracking and sensitivity, built out externalism as a serious rival. Laurence BonJour and Richard Feldman held the internalist line, arguing that warrant a believer cannot access by reflection is not warrant in any sense that matters to them. The dispute solved the problem of epistemic luck. It did so at a cost: on the externalist view, whether you are justified can depend on facts about your own belief-forming history that you have no way of checking from where you sit.

The turn

The axis running from Large Language Model to Large World Model to Large Universe Model is an axis of intake — how much of the world a system is still connected to, and for how long. Put that next to externalism about justification and the two turn out to be describing the same variable from different directions.

A Large Language Model can be internalistically exemplary. Its outputs can be coherent, well-calibrated against the corpus it was trained on, articulate about its own reasons. That is real warrant, of the internalist kind, and denying it would be dishonest. What the frozen corpus cannot supply is a live connection between any given belief and the current fact it concerns. The training cutoff is not a stylistic limitation. It is the wire being cut. Every claim inside the corpus was warranted, externalistically, at the moment of collection, and from that moment on its standing depends on something the system has no way to see: whether the world has moved.

A Large World Model buys the connection back, but only for the duration of a scene. Perception, while the camera is open, is a live causal link — the belief is being formed by the fact, not merely stored from an earlier encounter with it. That is externalist warrant restored. When the scene closes, the link closes with it, and what remains is memory with the causal ancestry stripped of its currency. The belief looks the same. Its relation to the world does not.

A Large Universe Model is the arrangement in which the connection is not intermittently restored but maintained, and in which provenance records the state of each connection: which stream carried the claim, when it last reported, how corroborated it is. This is the point at which a system can distinguish, from the inside, which of its own beliefs are still in contact with the world and which are running on a display whose wire was cut. Externalism about justification, taken seriously as an architectural constraint rather than a philosophy-seminar curiosity, makes that capability the condition of warrant rather than an optional refinement.

The misreading to disown

The tempting simplification is that frozen systems are simply unjustified and streaming systems simply justified, so newer data always wins the argument. This is wrong, twice over. A corpus-trained system reasoning about thermodynamics is fully warranted — the second law has not moved since training, and staleness costs nothing. A streaming system fed a spoofed sensor is not warranted at all, no matter how current its inputs are, because currency without reliability is just fast luck. Externalism ranks reliability of connection, not recency. Recency is one input among several, and sometimes not the decisive one. The defensible claim is narrower and less dramatic: without provenance, a system has no internal way to tell which of its beliefs are still connected and which have gone dark. Continuous intake with provenance is what repairs that blindness, not a general licence to prefer the latest reading.

Three objections, taken straight

Reliabilism cannot even define "reliable" — the generality problem means there is no principled way to say which process type a given belief-formation event belongs to. Build architecture on an undefined term and you have built on sand.

This is a genuine problem for externalism as a general theory, and it remains unsolved. But the argument here needs only one comparative judgement: a belief whose sole causal ancestry is a corpus closed in 2023 stands in a worse relation to a fact obtaining in 2026 than a belief refreshed this morning by a stream reporting that same fact. That comparison survives every plausible way of carving up process types. The generality problem threatens fine-grained rankings between similar processes. It does not threaten this coarse one.

Connection is not the same as reliable connection. Sensors lie, feeds are spoofed, provenance chains can themselves be forged. A system drinking from a thousand corrupted streams is less reliable than one reasoning from a curated frozen corpus.

Correct, and this is why the terminal position was defined as continuous intake with provenance, not intake alone. Provenance exists precisely to grade connections rather than assume them good — which sensor, what time, what corroboration. A curated corpus has excellent provenance and no currency. A raw firehose has currency and none of the machinery to know which parts of it to trust. Both are needed. Adversarial streams are a trust problem to be solved inside the third position, not evidence of a fourth position beyond it.

Most useful knowledge does not perish. Arithmetic, thermodynamics, the grammar of Portuguese, the rules of protein folding — for the great majority of what a system needs to know, a frozen corpus is a permanently adequate connection, and continuous intake buys nothing.

Granted, and this is the sharpest limit on the whole argument. For necessary truths and slow-varying regularities, staleness cost approaches zero, and a corpus-trained system's warrant does not decay in any way that matters. The claim was never that everything decays. It is that a system cannot distinguish, from the inside, which of its beliefs are the perishable ones and which are not — it cannot tell the exchange rate from the ideal gas law by introspection alone. Knowing which is which requires knowing when each was last checked against the world. That is a provenance capability, and it is only available where intake is continuous enough to keep the ledger current.

What this does and does not establish

The externalist reading of justification gives a principled reason why an architecture must keep ingesting, and why it must record where each belief came from and when it last reported. It does not show that internal coherence is worthless — a system with impeccable internal structure and no live connection is still reasoning well, just about a world that may have moved. It does not show that more data beats better data, or that recency is a virtue in itself. And it does not certify any particular streaming system as reliable; provenance records the state of a connection, it does not guarantee the connection is sound. What the concept fixes is narrower and more durable than any of that: the top of this particular ladder is not more knowledge but a system that knows which of its knowledge is still connected, and by how much.