Does the anchor have to be obviously unreasonable?

No. A plausible number can still receive more influence than its evidence warrants.

Can an AI explanation remove anchoring automatically?

No. Its effect depends on the explanation, task, and review process; it should be tested rather than assumed.

Is changing an AI estimate proof of independence?

No. A small change may leave the final judgment strongly tied to the initial suggestion. The relevant question is how the evidence supports the final answer.

AI Glossary / Deep learning / Beginner

Anchoring Bias

Definition

Anchoring bias is the tendency for an initial value or reference point to exert disproportionate influence on a later judgment. People may adjust away from a starting number yet remain too close to it, even when that number has little evidential value. In AI-assisted work, a suggested score, estimate, or draft can become an anchor that shapes the human review that follows.

What it means

An anchor supplies a starting point. The bias concerns how much influence that starting point receives relative to the evidence that should guide the decision. Using a well-supported estimate is not automatically biased; treating an arbitrary or weakly supported suggestion as a privileged baseline can be.

This is primarily a cognitive and decision-making concept, not a neural-network layer or training algorithm. The supplied deep-learning category is retained, but the practical relevance here is how people interact with AI outputs. An AI recommendation can be useful information and still influence judgment through the way it is presented.

How it works

A person encounters an initial number, interpretation, or proposed answer. Subsequent reasoning begins near that reference point, and adjustments may fail to reflect the available evidence adequately. The starting point can come from another person, an interface default, or one's own preliminary calculation.

In an AI workflow, showing a prediction before the reviewer forms an independent view may make the prediction the center of the discussion. Explanations can help scrutiny, but they can also provide additional material around the same starting answer. Whether a particular interface creates harmful anchoring requires empirical testing; agreement with the AI alone does not establish the bias.

Three practical examples

1. Project estimates: a generated plan suggests ten days. The team debates whether to add two days instead of building an estimate from the actual tasks and dependencies.

2. Content review: an assistant labels a customer comment as “angry.” A reviewer reads ambiguous wording through that label and overlooks a plausible neutral interpretation.

3. Demand planning: a dashboard pre-fills next month's forecast. An analyst makes a small adjustment without examining whether the starting forecast used a relevant comparison period. The issue is insufficient evidence-based revision, not the mere existence of a default.

Think of it this way

Think of placing a movable marker on a ruler before estimating an unknown length. Even when you know the marker can move, it may become the point around which you negotiate the answer. The analogy captures the pull of a reference point. It does not imply a fixed mechanical force: people, tasks, evidence, and presentation conditions affect whether and how strongly anchoring occurs.

Worked example

Consider a fictional estimation exercise. A task inventory supports 24 hours of work under a stated set of assumptions. Before examining that inventory, one reviewer sees an AI suggestion of 12 hours and another sees 36 hours.

Suppose their final estimates are 18 and 30 hours. Each moved six hours from the initial suggestion, but both remain six hours away from the inventory-based estimate in the direction of the anchor. The final estimates differ by 12 hours despite the same task evidence.

These numbers illustrate the pattern; they are not experimental findings. A real test would need suitable assignment of participants, comparable information, repeated observations, and a justified evaluation standard. One disagreement or inaccurate estimate cannot, by itself, diagnose anchoring bias.

Illustrative case study

Fictional case: Draft Compass, a support operations team, introduces AI-generated issue-priority labels. Reviewers must confirm a label before opening the underlying customer messages. During a workflow review, the team notices that discussion tends to revolve around defending or changing the preselected category.

It pilots an alternative in which reviewers record a brief evidence-based assessment before seeing the AI suggestion. The team compares decision quality, review time, and disagreements across the two workflows. It does not assume that hiding the suggestion must improve every case. The example reports no measured outcome; it illustrates a testable design change that makes independent judgment observable.

Common misconceptions and mistakes

Anchoring is not identical to automation bias. Automation bias concerns inappropriate reliance on automated assistance; anchoring concerns the influence of a starting reference. They can occur together, but a human-provided estimate can also anchor judgment.

Anchoring is not the same as statistical model bias or a learned neural bias parameter. Sharing the word “bias” does not make these concepts interchangeable.

It is also mistaken to conclude that simply warning people about anchoring eliminates it. A practical safeguard needs evaluation in the actual task. Finally, rejecting every AI suggestion is not evidence of sound judgment; the goal is appropriately weighting evidence, including useful machine output.

Business use

Businesses can review where suggested values enter a decision process: pre-filled budgets, risk categories, first drafts, ranked options, and meeting summaries are possible starting points. This review should focus on decisions where the initial suggestion can dominate later evidence.

Possible experiments include collecting independent estimates first, showing the supporting evidence alongside suggestions, or requiring a reason when the final decision conflicts with clear facts. These are design hypotheses, not guaranteed remedies. Compare accuracy and operational effort, and inspect cases where the AI is wrong as well as cases where it is right. A workflow can show high human - AI agreement while still handling misleading suggestions poorly.

Check your understanding

A reviewer adjusts an unsupported AI estimate from ten days to eleven days without examining the task list. Has the adjustment established that anchoring was avoided?

Answer: No. Moving away from a starting value is not enough. The final estimate still needs support from relevant evidence and assumptions.

Related terms

Automation bias
Confirmation bias
Cognitive Bias
Human-in-the-loop
Calibration
decision support
independent review
overreliance on AI

Learn next

Identify one recurring decision that begins with an AI suggestion. Record what evidence should determine the answer, then design a small comparison of suggestion-first and independent-assessment-first review. Measure decision quality rather than agreement alone.

Sources

Primary sources: Tversky and Kahneman, “Judgment under Uncertainty: Heuristics and Biases,” https://fbaum.unc.edu/teaching/articles/Science_1973_JudgmentUnderUncertainty.pdf ; Rastogi and colleagues, “Deciding Fast and Slow,” https://arxiv.org/abs/2010.07938 . The original anchoring section and the latter authors' abstract were read. Proposed workflow experiments and numerical examples are original illustrations, not universal empirical claims about every AI system or user.

Previous: Anchor texts Next: Anchoring effect