Must an anchor be deliberately manipulative?

No. An accidental default or a person's own initial thought can become an anchor.

Is every starting estimate bad?

No. Relevant evidence can justify using one; the concern is excessive or unjustified influence.

Does the effect have a fixed size?

No. Its measurement depends on the task, people, presentation, and comparison method.

AI Glossary / Graphs and symbolic AI / Beginner

Anchoring effect

Definition

The anchoring effect is a tendency for a judgment, especially a numerical estimate, to be influenced by an initial value or reference point. Later estimates may remain closer to that starting point than the relevant evidence warrants. An anchor can be supplied by another person, displayed by a system, or generated during one's own thinking.

What it means

Suppose an assistant shows a rough completion estimate before a planner examines the work. That number can become the starting point for discussion, even if it was based on incomplete information. People may revise it while still giving it too much influence.

Anchoring is relevant to AI because interfaces frequently present scores, forecasts, rankings, and suggested answers before a person makes a judgment. The number's presence can affect how the person interprets later evidence.

This does not mean every initial estimate is misleading. A relevant, well-supported starting point can be useful. The concern is disproportionate influence, particularly when the initial value is arbitrary, weakly justified, or difficult to evaluate.

How it works

Classic research by Tversky and Kahneman described estimates shifting toward starting values, including values unrelated to the quantity being estimated. This establishes an important distinction between information that appears in a task and information that actually deserves weight.

One explanation involves starting from a value and adjusting too little. Subsequent research distinguishes this process from other mechanisms, including the way a supplied anchor can make anchor-consistent information more accessible. It is therefore too simple to say that all anchoring is conscious calculation followed by insufficient correction.

In an evaluation, compare judgments under controlled conditions. If groups receive different irrelevant starting numbers but otherwise comparable evidence, a systematic shift in their estimates can support an anchoring interpretation. A single person agreeing with one number does not establish the mechanism.

For an AI interface, record when the suggestion appeared and what evidence accompanied it. These details help distinguish useful information sharing from an avoidable influence on independent assessment.

Three practical examples

1. A project assistant displays an early duration guess before task owners estimate their own work. A team can examine whether that display changes estimates without improving their accuracy.

2. A quality-review screen shows a preliminary score before the reviewer reads the material. The reviewer may interpret borderline evidence in relation to that score.

3. A researcher tests whether a chatbot's numerical answer changes when an irrelevant starting number is included in a prompt. Any observed effect describes that test setup, not proof of a human-like mental process.

Think of it this way

Think of setting a movable marker on a measuring line before examining the object. Later movements may remain organized around that initial marker, even when it was placed without evidence.

Unlike a physical anchor, a psychological anchor does not exert a fixed force. Its influence depends on the person, task, information, and presentation. The metaphor helps explain a starting-point effect but cannot predict its size or guarantee that it will occur.

Worked example

Consider an original illustrative review exercise. Two groups receive the same description of a fictional archive and estimate how many folders it contains. Group L first sees an irrelevant comparison number of 200; Group H first sees 800.

Suppose their resulting median estimates are 380 and 620. The difference in medians is 240 folders, while the anchor difference is 600. A simple descriptive ratio is 240/600 = 0.40.

These invented numbers demonstrate how a shift could be summarized. They are not experimental findings, an estimate of a universal anchoring strength, or evidence that a particular person is biased.

To investigate causation, a real study would need an appropriate assignment procedure, comparable groups, enough observations, and analysis of uncertainty. Knowing the true folder count would additionally allow assessment of accuracy. A shift between groups and an error relative to truth are different quantities; one should not substitute for the other.

Illustrative case study

Fictional case: Larch Planning notices that staff often revise an AI-generated estimate by only a small amount. The team introduces a trial workflow in which staff first record an independent estimate and its reasoning, then see the assistant's proposal.

It compares the workflows for accuracy, disagreement, time, and user understanding. It does not assume that hiding the suggestion must improve every decision. This invented case demonstrates a testable interface change, not a proven cure or a reported productivity outcome.

Common misconceptions and mistakes

Anchoring is not simply remembering a previous number. The issue is how that number influences judgment relative to the evidence.

Nor is every agreement with an AI estimate a bias. The estimate might be accurate and well supported. Establishing inappropriate influence requires examining the task and comparison conditions.

Avoid claiming that awareness alone eliminates anchoring or that a particular percentage adjustment solves it. Such rules can introduce another arbitrary reference point. Also distinguish human anchoring research from experiments on language models: similar output patterns do not demonstrate identical underlying cognitive mechanisms.

Business use

The concept helps teams inspect how decision-support interfaces shape human evaluation. Preliminary ratings, default quantities, and automatically generated forecasts should have clear meanings and evidence trails.

Possible design tests include collecting an independent estimate before revealing a suggestion, separating supporting evidence from the final number, or comparing alternative display orders. These are hypotheses to evaluate for the actual workflow, rather than universally superior designs. The objective is a better-grounded decision, not disagreement with AI for its own sake. Preserve useful suggestions while making their influence easier to examine.

Check your understanding

A reviewer sees a model score of 70 and independently finds evidence supporting 70. Does agreement alone prove anchoring?

Answer: No. One needs evidence about the score's influence and whether it exceeded what the relevant information justified.

Related terms

Cognitive Bias
Automation bias
Human evaluation
Confirmation bias
anchoring and adjustment
judgment under uncertainty
decision support

Learn next

Next, identify the first number displayed in a decision workflow. Document its basis, then design a controlled comparison that tests whether displaying it changes accuracy or merely shifts estimates.

Sources

Tversky and Kahneman, Judgment under Uncertainty: Heuristics and Biases, original paper reproduction: https://tualatinpsychology.weebly.com/uploads/5/7/2/8/5728202/anchoring_effect-_tversky___kahneman__1974_.pdf

Epley and Gilovich, Are Adjustments Insufficient?, primary abstract: https://journals.sagepub.com/doi/10.1177/0146167203261889

Original-paper anchoring passages and indexed publisher abstract consulted. Examples are original; interface suggestions are proposals to test, not established universal remedies.

Previous: Anchoring Bias Next: And-Elimination