What anchoring is
Ask someone to estimate a quantity they cannot know exactly — the age of a redwood, the population of Peru, the fair price of a used car — and give them a number first, even an arbitrary one, and their answer moves towards it. This is not a claim that people are lazy or careless. It survives care. It survives being told, explicitly and in advance, that the number is meaningless. It survives expertise: people who know the domain anchor as readily as people who do not, sometimes more readily, because they have a defensible-looking estimate to adjust from rather than a blank page to fill.
The mechanism is adjustment, and the failure is that adjustment stops too soon. Given a starting value, a person does not re-derive the answer from evidence. They nudge the starting value until it feels acceptable, and "feels acceptable" is a weaker test than "is correct." The nudging halts at the first defensible-looking figure, not at the true one, and the distance between those two points is the anchoring effect. It shows up on quantities — what a house is worth. It shows up on categories — what illness this is. It shows up on ranges — what a plausible answer even looks like, before any specific number is proposed. In every case the first frame does work that later evidence cannot fully undo.
Two demonstrations make the point vivid enough to be worth carrying forward. Real-estate agents and untrained amateurs toured the same house and were given identical supporting data, differing only in the listing price they were shown. Northcraft and Neale, running the study in 1987, found professional appraisers shifted their valuations by roughly eleven to fourteen per cent of the house's value depending on the anchor condition — a swing comparable to the amateurs' — and then denied, when asked directly, that the listing price had affected them at all. Expertise did not shrink the effect. It only sharpened the appraiser's confidence that no effect had occurred. Nearly two decades later, Englich, Mussweiler and Strack gave experienced German judges a rape case file and had them roll dice rigged to land on 3 or 9, told the roll represented a prosecutor's sentencing demand in months. Judges who rolled the higher number handed down longer sentences. They knew the die was loaded. They sentenced by it anyway.
Where the idea came from
Amos Tversky and Daniel Kahneman named anchoring and insufficient adjustment in a 1974 paper in Science, part of a broader programme cataloguing the heuristics people use under uncertainty. Their canonical demonstration used a wheel of fortune, rigged to stop at either 10 or 65, which subjects were told was entirely random. Asked afterwards to estimate the percentage of African nations in the United Nations, those who saw the wheel stop at 10 guessed around 25 per cent; those who saw 65 guessed around 45 per cent. The wheel had nothing to do with African nations. It still moved the estimate by twenty points.
The paper's contribution was not the observation that people err. Error had long been treated as noise around a correct answer, unpatterned and uninteresting. Tversky and Kahneman's demonstration showed the error was systematic, directional, and predictable from the anchor alone — evidence of a specific cognitive procedure, efficient under time pressure and wrong in a knowable way, rather than a shrug of imprecision.
The turn
The concept was built for individual human judgement. It transfers, cautiously, to any system that must produce an estimate from evidence and cannot re-derive that estimate from nothing each time. That description fits a person weighing a listing price. It also fits a statistical model weighing a training distribution.
A Large Language Model is anchored twice over: once by the composition of its training corpus, once by the date beyond which nothing further arrives. Every later output is generated from a distribution of weightings fixed at training time. Correction after that point — a prompt, a retrieved document, a fine-tuning pass — happens at the surface. It changes what the system says without changing the learned weighting that decided, at the outset, how much any given piece of evidence should count. This is exactly the shape of insufficient adjustment: the estimate moves a little, the process that generated it does not move at all.
A Large World Model breaks this differently. It observes a scene directly — the room now, the object in front of it — and for the duration of that scene the sensed present can outrank the trained prior. This is a real loosening, not a cosmetic one. But the scene ends, and the anchor is exactly where it was. Nothing about having seen one room changes how the next room's evidence will be weighted, because the loosening was local and temporary, not structural.
A Large Universe Model is defined by two properties working together: intake that never closes, and provenance attached to each belief recording what produced it and when. Provenance is the part anchoring theory lacks entirely. It is what makes an estimate's origin inspectable rather than opaque, and inspectability is what allows revision to reach the weighting itself rather than merely decorating its output. This is not a claim that anchoring disappears. It is a claim about which regime contains the machinery capable of dissolving an anchor rather than nudging around it.
The misreading to disown
The common mistake is to think a training cutoff is a knowledge problem — the system doesn't know about last month, so fetch it a document about last month and the problem is solved. This treats anchoring as staleness. It is not. The anchor is not any particular missing fact; it is the learned weighting that decided, once, which sources count and what a plausible answer even looks like. Feeding in a fresh fact does not touch that weighting. It is why a retrieved contradiction so often loses to a confident prior: the document is new, the process evaluating the document is not.
Three objections, taken seriously
A machine has no fatigue. Retrieval can override a stale prior at inference time with no effort cost. The anchoring story is about human limits that do not apply here.
The effort account of anchoring is correct for humans and beside the point for the transfer. What carries over is structural, not motivational: adjustment happens to the estimate, not to the process generating it. Retrieval changes an answer about a specific fact. It does not touch the weighting that decides how much one retrieved document should outweigh a training distribution built from millions of examples pointing the other way — and measured behaviour shows exactly this: conflicting retrieved evidence gets discounted, not adopted, in proportion to how confidently the prior was held.
A prior is not a defect, it's the whole point of inference. A system with no anchor at all would just chase whatever it saw last — recency is an anchor too, and often a worse one than a well-built corpus.
This is right, and it genuinely narrows the claim. Unbounded intake does not by itself guarantee sound weighting. A badly designed continuous system anchors on whatever arrived most recently and oscillates instead of converging. The claim here is not that open intake produces good judgement automatically. It is that a closed corpus makes its prior unrevisable in principle, because there is no later evidence admissible against it, while open intake with provenance makes revision possible in principle. Whether a given system actually revises well — how it weights, how it decays old evidence, how it resists recency — is a calibration question that comes after the terminal position, not a promise contained within it.
Every learning system has path dependence. Sequential updating is famously order-sensitive; early observations shape later representations whether intake is open or closed. So the third position inherits the pathology rather than escaping it.
Path dependence does not disappear, and this objection is worth conceding fully. What differs is accessibility, not presence. Under closed intake, the founding distribution cannot be contradicted by anything, because nothing further is being admitted at all. Under open intake with provenance, an early belief can be traced back to the specific observations that produced it, and those observations can be re-weighted once they are shown to be unrepresentative. Order-dependence stays real. It converts from a permanent property of the system into a correctable one — replay, audit, reweighting — because the record needed to correct it exists.
What this does and does not establish
This is an argument about structure, not a verdict on any implementation. Anchoring shows that a fixed-intake system carries an unreachable prior by construction. It does not show that continuous intake, however designed, will reason well, weight evidence sensibly, or resist its own version of recency bias — those remain open, empirical, and probably hard problems, the kind that follow from having the right structure rather than being solved by it. What the concept establishes is narrower and more defensible: on the specific axis of what a system is permitted to observe and revise, there is a real difference between an anchor that cannot in principle be reached and one that can. That difference is worth naming precisely because it is easy to blur into a promise of better judgement, which it is not.