Skip to main content
computer-science psychology

​
Anchoring

​
Description

An initial reference value disproportionately influences subsequent estimates, even when the anchor is irrelevant to the judgment at hand. The estimator partially adjusts away from the anchor but typically not enough; the residual pull is the bias. The defining property is irrelevance-tolerance: the anchor pulls the estimate even when the estimator knows or has been told the anchor is arbitrary. Distinct from seeding: seeding determines the emergent topology of a growth process (database schema, model weights, founding team), where the seed’s effect propagates through dynamics over time; anchoring biases a point-estimate magnitude in a single act of judgment. Both are initial-input shapes, but at different granularities.

​
Triggers

User-initiated: User describes a judgment being pulled by an arbitrary first-input, or discusses negotiation tactics around who-anchors-first. Vocabulary cues: “anchoring,” “first number,” “opening offer,” “reference point,” “primed by.” Agent-initiated: Agent notices a prior estimate, prior context, or arbitrary first-value being deferred to disproportionately in subsequent reasoning. Candidate inference: “is this anchor doing real work, or is it just the first number we saw?” Situation-shape signals: Negotiations where the opening move shapes the final outcome. Estimates produced under uncertainty with a salient prior in the room. Prompt-engineering decisions where context order matters. Re-estimation that adjusts insufficiently from a prior baseline.

​
Exclusions

  • Bayesian updating with informative priors — when the “anchor” is actually a well-calibrated prior and the adjustment is appropriately weighted, you have rational inference, not anchoring bias.
  • Domains with strong external grounding — when an objective ground-truth is rapidly available (a measurement, a market price), the anchor is dominated by the signal and the bias dissipates.
  • Re-estimation from scratch — when the estimator deliberately discards prior reference and re-derives from first principles, the anchor’s pull is neutralized (though humans famously struggle to do this even when instructed).

​
Structure

Internal structure of anchoring: a table of its component slots and the concepts that fill them.

​
Relationships

Relationship neighborhood of anchoring: a graph of the concepts it connects to and the concepts it is a part of.
  • seeding — structural cousin: seeding is initial-input-shapes-emergent-topology; anchoring is initial-input-biases-point-estimate. Both ask “what does the first input lock in?”
  • trigger-rule-pair — contrast: trigger-rule-pair is intended coupling between condition and rule; anchoring is accidental coupling between arbitrary anchor and downstream estimate.
  • context-asymmetry — the party setting the anchor has a context advantage; anchoring is a mechanism by which context-asymmetry produces concrete outcome bias.
  • cargo-cult — over-deferring to the first-cited number because “we anchored on this” without re-validating is cargo-cult anchoring; the concept composes with cargo-cult when the anchor’s authority isn’t earned.

​
Examples

Retail pricing · psychology

sticker-price-then-discount is anchoring; the original-price anchors the discounted-price perception of value.

LLM prompt-context priming · computer-science

examples or framings in the prompt anchor subsequent generation toward those values, even when the user explicitly requests divergence.
Galinsky and Mussweiler studied how first offers in negotiation function as anchors, demonstrating that the party who makes the first offer typically captures a disproportionate share of the surplus. The opening number — even when both sides recognize it as a negotiation move — exerts a measurable pull on the final settlement.The paper’s contribution to the catalog: it isolates anchoring as a mechanism, not merely a bias. Counter-anchoring by making your own first offer, or by reframing the focal value before the anchor lands, reliably attenuates the effect. The anchoring shape is the same one that shows in retail pricing and salary negotiations; the Galinsky-Mussweiler result is what lets us treat those as instances of a single primitive rather than three folk-psychology tips.
the first juror to suggest a damages figure anchors deliberation around that magnitude; defense and prosecution both deploy the tactic.
initial list price anchors buyer offers; sellers exploit by listing high, knowing the anchor pulls negotiated price up.
first number stated (by recruiter or candidate) heavily shapes the final compensation; salary-discussion advice often centers on who anchors and how.
Strack and Mussweiler proposed the selective accessibility account of anchoring: the anchor activates target-consistent information in memory, making that information more accessible during the subsequent estimate. The effect appears even when the anchor is implausibly extreme (e.g., “Is Gandhi older than 9 years old?” still shifts age estimates upward), arguing against pure adjust-from-anchor accounts.Inference: anchoring is not just a strategic ploy or a numerical-adjustment artifact — it operates on the substrate of which thoughts come to mind. This generalizes the concept beyond negotiation: any prior framing that activates a subset of memory will skew downstream estimates in that direction, including in code review, design discussions, and incident postmortems.
The 1974 Science article is the founding statement of the heuristics-and-biases program, introducing three heuristics — representativeness, availability, and anchoring and adjustment — as the cognitive shortcuts people use to estimate probabilities and magnitudes under uncertainty. The anchoring section is the original empirical case for the phenomenon: subjects are given a starting value, asked to adjust toward the correct answer, and reliably under-adjust.The wheel-of-fortune demonstration described in this paper is the canonical illustration of the irrelevance-tolerance property: the anchor influences judgment even when subjects watched it being generated by an obviously random process. This is what makes anchoring a bias rather than a rational use of prior information — the structural property that distinguishes it from Bayesian updating.The half-century since has built up overwhelming cross-domain replication: opening offers in negotiation, sticker prices in retail, judicial sentencing demands, jury awards in tort cases, salary expectations seeded by the recruiter’s first number, real estate list prices, and (more recently) the influence of context tokens on LLM completions. The portability is exceptional; few cognitive biases transfer this cleanly across domains where the surface details differ but the structural shape — a salient first value pulling a downstream estimate — recurs.
Subjects in the experiment first watched a wheel of fortune spin to a number between 0 and 100, were asked whether the percentage of African countries in the UN was higher or lower than that number, and then asked to estimate the actual percentage. The wheel was rigged to stop on either 10 or 65 — and groups who saw the lower spin gave substantially lower estimates than groups who saw the higher spin. The spin’s irrelevance was structural and visible to subjects; the pull was robust anyway.Inference: This is the canonical demonstration of the property that makes anchoring a bias rather than rational inference — the anchor influences the estimate even when it is known-to-be-arbitrary. If the pull dissipated when subjects knew the anchor was random, anchoring would just be unsophisticated Bayesian updating; the fact that it doesn’t is what gives the concept its diagnostic edge.