Justified true belief and the Gettier problem in humanitarian response

The stopped clock in the situation report

Take the strongest version of the objection first, because it deserves stating well. Humanitarian coordination runs on assessments, and assessments are always old by the time anyone acts on them. A rapid needs assessment in a flood-affected district takes days to field, collate and clear. Displacement figures from the International Organization for Migration's Displacement Tracking Matrix are dated the moment they are published, sometimes by weeks. If knowledge requires live justification — a belief still connected to the fact that makes it true — then almost nothing a response coordinator works from counts as knowledge. Every allocation decision would rest on Gettier cases: justified, true when written, believed in good faith, and disconnected from the ground by the time the trucks move.

A sceptic can push further. If this is the standard, it is a standard no humanitarian system has ever met, or could meet, and treating that as scandalous rather than normal is a category error borrowed from a philosophy seminar. Assessments are proxies. Everyone in the sector knows they decay. Demanding an unbroken chain between belief and fact for populations moving under duress, across contested terrain, with patchy connectivity, is asking for an epistemic standard incompatible with the work. On this view the whole apparatus — Edmund Gettier's 1963 counterexamples to justified true belief, Russell's stopped clock, the rest — is an elegant but irrelevant parlour game. Reliability, not live connection, is what a coordinator needs, and reliability is achievable without solving anyone's epistemology.

That objection is close to right about ordinary cases and wrong about the cases that matter. The honest way through it is to grant the first part fully.

Where reliabilism actually wins

Most humanitarian facts are boringly stable for the duration that matters. The number of health posts in a district, the location of a river crossing, the standard caseload for a cholera treatment centre — these do not move at the pace of a displacement crisis, and a two-week-old assessment is fine for them. Reliabilism, the view that a sufficiently accurate process suffices for knowledge regardless of any deeper justificatory link, is the correct account of this layer. A coordinator who treats a 90% reliable baseline survey as known is not making a mistake. Insisting on live grounds for every proposition in a situation report would be paralysis dressed as rigour, and no agency operates that way, nor should it.

Concede that, and the objection has taken most of the ground it wanted. What it has not taken is the distribution of error.

Why the errors are not scattered

The failure in humanitarian response is not that assessments are sometimes wrong. It is that they are wrong in a patterned way: precisely on the propositions that changed since collection, and the propositions that changed since collection are precisely the ones a coordinator is asking about when a decision is urgent. Nobody queries the assessment to find out where the river crossing is. They query it to find out where twenty thousand people are right now, because that is the number the ration plan depends on, and that is the number that moved.

This is the structural point that a bare reliability figure cannot capture. An assessment that is 95% accurate in aggregate can be almost entirely wrong on the volatile 5% — displacement flows, market prices for staple foods, health surveillance signals, access constraints on the roads a convoy needs — because that 5% is exactly the part still in motion while everything else held still. A confidence score computed from the frozen assessment has no way to flag which subset it is failing on, because the evidence of the failure lives on the other side of the assessment's own cutoff. The coordinator who reads "estimated 40,000 displaced, concentrated in three collection points" is reading a stopped clock. It was right when written. Whether it is right now depends on a week of onward movement the document cannot see.

Ordinary reliability suffices. If the assessment is right the great majority of the time, treat it as knowledge and act. Demanding a live connection to a fact for every line item is a standard nothing in the sector meets, and pretending otherwise stalls response.

The reply is not that this is false. It is that reliability is the wrong unit. The unit that matters is: reliable about what, and stale for how long, and which part of the estimate is load-bearing for the decision being made right now. Aggregate accuracy hides exactly the failure that gets people fed in the wrong place.

The market price problem, and why fresher is not automatically safer

A second objection has real teeth in this domain specifically. Humanitarian response increasingly leans on live feeds — market price monitoring, mobile-derived mobility signals, satellite-based flood extent, health syndromic surveillance — precisely to escape the stopped-clock problem. But a live feed is not automatically a justified belief. A single price-monitoring source can be manipulated by traders anticipating a cash transfer programme. A mobility signal derived from mobile network data can misread a mass evacuation of phones left behind as a mass evacuation of people. A syndromic surveillance stream from a handful of health posts can spike because one clinic changed its case definition, not because disease incidence changed. Swap a stale corpus for a live pipe and you have not eliminated Gettier cases. You have traded one class for another — and a live source carries the unearned authority of "current," which can make its errors more dangerous, not less, because nobody thinks to doubt the present tense.

This objection is correct, and the argument does not claim otherwise. What it claims is narrower, and the difference matters operationally. A stale assessment's defect is invisible from inside the assessment. There is no line in a two-week-old displacement report that tells you which figures the movement has since overtaken; the evidence of change is, by definition, evidence that arrived after the document closed. A corrupted live stream's defect is checkable in principle: against a second price source in the same market, against registration data corroborating the mobility signal, against a second clinic's case counts. Cross-stream disagreement, provenance on each figure, and redundancy across sources turn a corrupted stream into a detectable failure rather than a hidden one. One kind of error is structurally invisible to the coordinator holding it. The other is structurally auditable, given more than one stream and a record of where each number came from. That asymmetry, not any claim that live data is automatically trustworthy, is the whole of the argument.

The three positions, restated for this domain

The Large Language Model in this domain is the frozen situation report treated as still current: correct when filed, silent about its own expiry. The Large World Model is the rapid assessment team on the ground — genuinely connected to the facts, but only for the district they walked and the days they were there; the moment they leave, the report starts decaying exactly like the corpus did, just from a later start. Neither failure is a defect in the analysts' competence. Both teams reasoned soundly from what they had. The defect is external: the link between the grounds and the fact broke after the grounds were fixed.

The narrower claim

Truth in a humanitarian assessment is not the scarce resource. Most figures in most reports are true when written. The scarce resource is a maintained connection between the figure and the fact — a standing subscription to the displacement flow, the market, the health signal and the access constraint, with each belief tagged to where it came from and how old it is allowed to get before someone is obliged to check it again. A response coordinator does not need every line item live. They need to know, line by line, which figures are load-bearing for today's decision and which of those are still connected to the world or have quietly become stopped clocks.

That is a narrower claim than "build better models" or "collect more data," and it survives both objections raised against it. Reliabilism is right about the stable layer and wrong about the volatile subset that decisions actually turn on. Live streams are right to replace frozen corpora and wrong to be trusted merely for being live. What is left, once both concessions are made, is not a bigger analysis engine. It is bookkeeping: provenance, timestamp, and an explicit decay clock on every belief a response is built from — the maintenance of a connection, not the accumulation of an archive.