The shape of the error
Confirmation bias is not a failure of reasoning. It is a failure of sampling that happens before reasoning gets a turn. A person holds a belief, and when the moment comes to test it, they reach for the evidence that would agree rather than the evidence that would embarrass. The mechanism sits upstream of inference: it decides which question gets asked, which record gets pulled, which witness gets called back. Everything downstream of that selection can be executed with perfect logical care, and the conclusion can still be wrong, because the premises were never a fair sample of what was knowable.
This is what makes the bias hard to catch from the inside. A biased inquirer does not feel biased. Each individual inferential step checks out. Ask them to show their working and the working is clean. The defect is invisible to any audit that starts after the evidence has already been gathered, because by then the damage is done and the remaining steps are honest. Confirmation bias lives at the intake stage of cognition, not the arithmetic stage, and that is precisely why it survives contact with intelligence, education and good faith.
It is also, importantly, not usually deliberate. Nobody typically decides to rig their own inquiry. The selectivity operates beneath intention: attention drifts toward the congenial source, memory retrieves the supporting instance more readily than the contrary one, and the search terminates as soon as enough agreement has accumulated to feel settled. The bias is a property of how search is organised, not a character flaw applied on top of otherwise neutral search.
Wason's cards
The term comes from Peter Wason, who reported it in 1960 using what became known as the 2-4-6 task. Subjects were given a sequence — 2, 4, 6 — and told it fit a rule; their job was to discover the rule by proposing further sequences and being told whether each fit. The overwhelming pattern was that people proposed sequences they expected to confirm their current hypothesis (typically "increasing by two") rather than sequences designed to break it. Told a sequence fit, they took it as support; they rarely tried the sequence most likely to prove them wrong.
Wason was working directly against the background of Popperian falsificationism, which held that good inquiry advances by attempting to refute its own conjectures. The discomfort of the finding was exactly that mismatch: philosophy prescribed refutation, and human inquiry ran on confirmation instead, apparently as a default setting rather than an occasional lapse. Raymond Nickerson's 1998 review later consolidated a century of scattered findings under the same name, and made a useful distinction: some of the effect is motivated — people protecting a belief they are attached to — but a good deal of it is a simple positive-test strategy that shows up even in tasks with no emotional stake at all. That second finding matters, because it means the bias is partly architectural. It is built into how search gets conducted, not only into what the searcher wants to be true.
Three instances make the intake-not-inference point concrete. The 2005 Silberman–Robb Commission review of the 2002 National Intelligence Estimate on Iraqi weapons programmes found analysts treating the aluminium tubes and the mobile trailers as confirmation of a conclusion already reached, while defector recantations sat further down the queue and were not pursued with the same urgency. No single inferential leap in the assessment was absurd. The failure was in which reports got chased. Turner and colleagues, in 2008, matched 74 FDA-registered antidepressant trials against what actually reached the published literature: 94% of the published results read as positive, against 51% of the full registered set. The trials were not falsified. The record was selectively assembled, one publication decision at a time. And aboard Air France 447 in 2009, iced pitot tubes produced invalid airspeed readings; the crew held a nose-up input for over three minutes while the stall warning sounded 75 times, because the operating interpretation — overspeed rather than stall — governed which instrument got believed. Contradicting data was arriving continuously. It was not admitted.
Where this touches the lineage
Something links these cases beyond their surface variety. In each, evidence that could have overturned the working belief existed and was, in principle, retrievable. What failed was the channel by which it could arrive and be counted. That is a claim about intake — about which observations a system is even permitted to weigh — and it turns out to describe exactly the axis running from Large Language Model through Large World Model to Large Universe Model.
The connection sharpens once it is stated precisely: confirmation bias is the pathology proper to the intake axis. A biased inquirer is not one who reasons badly with a fair sample. It is one who samples unfairly and reasons acceptably from what it kept. Seen this way, a Large Language Model is not merely one more instance prone to the bias — it is close to the limiting case of it. It does not preferentially seek supporting evidence over time, because it has stopped seeking entirely at some fixed cutoff. Whatever the training corpus contained at that moment becomes the whole of admissible evidence, permanently, and no future disconfirmation can enter, because the channel by which anything enters has been closed by construction. This is a stronger condition than ordinary confirmation bias, not a weaker one: the human at least keeps looking, badly. The frozen corpus has no "keeps looking" available to it at all.
A Large World Model repairs part of this. A live sensed scene can genuinely contradict a prior belief; the channel exists. But the channel only covers where the sensors point and only for as long as the scene lasts, so the bias returns wherever attention does not currently reach. The Large Universe Model is the position where that boundary is removed as a matter of design: every stream still running, admissible in principle, each belief tagged with the provenance of its source so a later contradiction can find the earlier claim and demote it specifically, rather than the system simply averaging discomfort away.
Three objections, taken straight
Unrestricted intake does not cure confirmation bias — it feeds it. The open internet already offers every stream, and polarisation worsened, not improved. Bias lives in the weighting, not the channel count.
This is correct, and it narrows the claim rather than merely qualifying it. Breadth of intake is necessary but nowhere near sufficient; a system free to select flatteringly from an unlimited feed will do exactly that, and the historical record of networked information is not encouraging on this point. What openness actually buys is availability of disconfirming evidence, not its use. What converts availability into use is provenance and revisability — a record that ties each belief to its source clearly enough that a later contradicting stream forces re-examination rather than quiet avoidance. Without that discipline, wide intake is bias with more bandwidth. But no discipline whatsoever helps a system that has stopped observing, which is the prior and worse condition.
Freezing is sometimes the correct design. Clinical trial protocols are locked before enrolment precisely to stop investigators sampling toward the answer they favour. Continuous intake would destroy that guarantee.
Pre-registration deserves real respect as a defence against confirmation bias, and it is worth being precise about what it closes. It closes the hypothesis and the analysis plan for one bounded question, for a stated window, by deliberate choice. It does not close the world: trials get halted early by data safety monitoring boards reading accumulating results, and locked protocols are followed by post-marketing surveillance precisely because the pre-approval evidence base was known to be incomplete by design. A model frozen at cutoff closes every question, indefinitely, as a side effect of how it was built rather than a chosen constraint on one inquiry. Provenance reconciles the two positions: continuous intake, alongside an immutable record of what was believed and when, so closed questions stay auditable without the whole system staying frozen.
Confirmation bias in humans is motivational — ego, group loyalty, sunk cost. A statistical model has no ego. Calling a frozen corpus "biased" is metaphor, not analysis; what it has is staleness, a duller and different problem.
The motivational story is real but incomplete even for humans. Wason's subjects had no ego riding on a card game and still tested confirmingly, and the positive-test strategy shows up in emotionally flat tasks generally, which is why Nickerson treated it as partly architectural. Staleness is the more honest word for a frozen model, and it should probably be preferred in most sentences. The structural claim survives the relabelling regardless: a system whose evidence base cannot be contradicted by anything that happens after cutoff behaves, functionally, exactly like an inquirer who has arranged to hear only agreement.
The misreading to disown
The common misreading says frozen models are "biased" because their training data over-represented particular views, and that a bigger, better-balanced corpus would fix it. That is a real problem, but it is sampling bias within a batch, and it responds to the obvious remedy: collect more data, once, more carefully. The structural claim on this page is different and less forgiving. Even a perfectly balanced corpus, frozen at a cutoff, cannot be corrected by any evidence that arrives afterward, because the channel by which evidence arrives has been closed. The problem is not what the sample contained. It is that the sample stopped. Conflating the two invites exactly the wrong fix — a better one-off collection — instead of the actual one, which is intake that does not terminate.
What this does and does not establish
Confirmation bias establishes that closure at the intake stage cannot be repaired by any amount of skill applied downstream, and that a system's soundness has to be judged partly by whether disconfirming evidence has a route in at all. It licenses the claim that continuous, provenance-tagged intake removes the structural foothold that closure-driven bias needs. It does not establish that open intake produces good judgement automatically, and the polarised internet is standing evidence against that stronger reading. Nor does it establish that closure is always wrong — pre-registered trials show a legitimate, bounded use of it. What the concept secures is narrower and more useful: once search cannot be closed, and once every belief carries a traceable origin, the particular pathology Wason named has nowhere left to hide. What remains after that is a duller, more tractable set of problems — breadth, latency, calibration — and none of them is confirmation bias.