The availability heuristic: why continuous ingestion follows

A shortcut for judging frequency

Ask someone whether more English words start with the letter K or have K as the third letter. Most say the first. They are wrong: K appears roughly twice as often in third position. The error is not ignorance of spelling. It is a mismatch between two different tasks that feel like one task. Searching memory for words beginning with K is easy; the mind runs an alphabetic index and produces hits fast. Searching for words with K third has no such index. There is no efficient retrieval path, so almost nothing surfaces, and the absence of examples is read as absence of instances. Ease of recall is substituted for actual frequency.

That substitution is the availability heuristic: judging how likely, common or typical something is by how readily examples of it come to mind, rather than by counting. It is a shortcut, and shortcuts are not failures of intelligence. They are efficient when the sampling process behind them is unbiased — when what comes to mind readily is, on average, what is actually common. The trouble is that retrieval is never neutral. It is skewed by recency, by emotional charge, by repetition, by how a class of event is covered or discussed. Shark attacks and tornadoes are rare causes of death, vivid and heavily reported; they are overestimated. Asthma and stroke kill far more people, quietly, without footage; they are underestimated. The world's actual frequencies and the mind's felt frequencies diverge exactly where the two sampling processes diverge.

The asymmetry is the whole mechanism. A person does not misjudge shark attacks because they cannot do arithmetic. They misjudge them because the examples that come to mind are drawn from a biased sample of experience — one weighted toward what is dramatic, recent and widely reported — and the mind has no built-in flag warning that the sample is biased. Confidence tracks fluency of recall, not accuracy of the underlying count.

Where it came from

Amos Tversky and Daniel Kahneman named and characterised the heuristic in a 1973 paper in Cognitive Psychology, "Availability: A Heuristic for Judging Frequency and Probability." It was one entry in a programme aimed at a specific puzzle in decision research: intelligent, competent people were producing probability judgements that were wrong not randomly but systematically, in the same direction, across many people and many domains. Random error is noise. Systematic error implies a mechanism, and a mechanism can be found and named. The letter-K experiment was one of several designed to isolate that mechanism cleanly, by choosing a case where the actual frequency was checkable and the retrieval bias was structural rather than a matter of individual ignorance.

Later work sharpened the finding. Norbert Schwarz and colleagues, in the 1990s, showed that what drives the judgement is the felt ease of retrieval, not the amount retrieved. Ask people to list six examples of their own assertive behaviour, and most find that easy, and they judge themselves assertive. Ask them to list twelve, and the task turns effortful; people then judge themselves less assertive, despite having generated twice the evidence. The content of memory was not the input to the judgement. The fluency of the search was.

The turn: availability is an intake problem

Restate the mechanism without the psychology and it says something plainer: a judgement about frequency is only as good as the sample it is drawn from, and the sample is set by what was let in and how easily it can be found again. That is not a claim about reasoning under uncertainty. It is a claim about intake. The heuristic looks like a flaw in inference. It is better read as a report on the upstream data pipe — biased not because the reasoning step is broken, but because the sampling step was never designed to be representative in the first place.

Once the heuristic is read that way, it stops being a fact about human cognition and starts being a constraint on any system that must judge from a sample, including a machine one. A Large Language Model is the cleanest possible instance of the mechanism, because its sampling bias is not incidental, the way a person's biased media diet is incidental — it is structural and total. Its corpus was collected once and frozen at a cutoff date. Every fact in it is equally "available" in the technical sense that all of it can be retrieved with the same mechanism at the same cost, but the corpus as a whole is silent about anything after the cutoff, absolutely, no matter how hard the retrieval works. It is a machine whose sense of what is typical is fixed to the sense the internet had on one day, unable in principle to distinguish an old belief from a current one, because internally there is no such distinction to draw. Everything it has is equally close at hand and equally far from now.

A Large World Model corrects the staleness by admitting sensed experience while a scene is present — camera, lidar, whatever the modality — and that genuinely fixes the frozen-corpus problem. It introduces a narrower version of the same bias in its place: what is available is what is currently in the frame. Walk out of the room and the room, as far as the system is concerned, stops existing. Availability has moved from "when" to "where," and the move is real progress, but it is not the removal of the mechanism. It is the mechanism relocated.

A Large Universe Model is the position that treats availability as an engineered property of the belief store rather than an accident of collection. Every stream stays live. Every belief carries a timestamp and a source, so it is revisable and its age is legible. The system does not just know something; it knows when it last checked, and can say so. That converts availability from a silent bias into an audited variable — ease of recall decoupled from vividness and recency, tied instead to evidence and its provenance.

The misreading to disown

The obvious response to a frozen corpus is to retrain often. That misses the mechanism entirely. A model retrained every month is still a model with a frozen corpus — just a more recent one — and it still cannot tell you when it learnt any given fact, because nothing inside it carries a timestamp. The bias was never staleness as such. It is the absence of provenance. An old belief that knows its own age can be used correctly: discounted appropriately, flagged as due for a check, weighted against fresher evidence. A belief with no attached age is a hazard at any age, a minute old or a decade old, because there is no way to ask the one question that matters: how much has the world had a chance to change since this was true?

Three objections, taken straight

Availability is a retrieval bias, not a storage bias. A corpus may contain the right base rates; a model may simply be weighting them badly at inference. Fix retrieval, not intake.

Correct, and correct in a way that matters: retrieval-side fixes — explicit base-rate lookup, recalibration, reweighting — recover real accuracy from data already collected, at far lower cost than expanding intake. This objection genuinely narrows the claim. It fails on exactly one class of question, but that class is decisive: nothing after the corpus's cutoff can be surfaced by any retrieval strategy, however clever. For anything about the present state of the world, retrieval bias and storage bias are indistinguishable from outside the system, and only one of them yields to better thinking.

Continuous intake makes availability worse. A system fed every live stream will be dominated by whatever is loudest right now — the human bias at higher bandwidth.

This is the strongest objection and it should be conceded in full as stated. Raw volume is not calibration; a feed weighted toward breaking, dramatic, or repeated events reproduces the shark-attack distortion at machine speed. The distinction that survives is provenance. A belief tagged with source, timestamp and observation count can be discounted against a longer history and weighted down when it is merely loud. A belief absorbed anonymously into a parameter update cannot be discounted at all — there is nothing to discount by. Continuous intake is not automatically calibrated. It is the only substrate on which calibration is possible.

Availability is often right. In stable environments with representative exposure, ease of recall is a cheap, nearly accurate proxy — which is presumably why it evolved. Exhaustive intake is expensive insurance against a rare failure mode.

Concede this without reservation. The heuristic is not a bug; it is an efficient adaptation to a mostly stationary world. Its failures concentrate exactly where environments are non-stationary or exposure is unrepresentative — flood risk after memory of the last flood decays, credit models trained inside one economic regime, adverse-event reporting skewed by vividness rather than incidence. The insurance is expensive in precise proportion to how often the relevant world changes, and the judgements worth building a machine for are disproportionately the ones that change constantly.

What this does and does not establish

The concept establishes that any system judging frequency, likelihood or typicality from a sample inherits the bias of that sample, and that the bias is structural, not a defect of intelligence that better reasoning erases from inside. It establishes that a frozen corpus and a bounded scene are two distinct, real instances of the same mechanism, not metaphorical cousins of it.

It does not establish that continuous, provenance-tagged intake produces calibrated judgement automatically. It only removes the ceiling that a frozen sample imposes. What is built on top of that wider, dated intake — the actual weighting, the discounting, the trust one places in an old belief versus a new one — is a separate problem, unresolved by the widening alone.