Anchoring Bias

Whoever hears a number first is influenced by it — even if it has nothing to do with the actual question at hand. This effect is called anchoring bias. The first value acts like an anchor: the judgment that follows does drift away from the anchor, but rarely far enough. If someone asks whether Australia’s population is greater or smaller than 500 million, and then asks for one’s own estimate, most people name a figure that is far too high — because the 500 million lingers in the mind as a starting point. This happens even though the original number was clearly wrong. The effect has been well documented in research on human decision-making since the 1970s.

Anchoring Bias in AI Systems

In AI development, the anchoring effect is relevant for two reasons. First, it influences the people who evaluate and train AI systems. When a model delivers an initial answer, human evaluators tend to take it as a reference point — later answers are not judged in absolute terms, but relative to it. This distorts the feedback the model receives during training.

Second, language models themselves exhibit anchor-like behavior. Several studies have shown that a model gives different answers when it is first given a particular number or claim — even when it is subsequently asked for an independent assessment. This is not consciousness and not intent. It is a statistical pattern: the model has learned from human text, and human text is full of anchoring effects.

How an Anchor Distorts Thinking

The mechanism behind the effect is comparatively well understood. When an initial value is introduced, the brain starts searching from that point instead of estimating independently. It adjusts step by step — and stops too soon. This systematically causes estimates to be skewed toward the anchor, even when the person knows the anchor was arbitrary.

Interestingly, even obviously absurd anchors have an effect. In a classic experiment by Amos Tversky and Daniel Kahneman, two psychologists who systematically studied the phenomenon, test subjects spun a wheel of fortune. The outcome of the wheel — a completely random number — subsequently influenced their estimates of historical facts. Conscious deliberation can weaken the effect, but it rarely disappears entirely.

In machine learning, no such mental process takes place. Nevertheless, similar patterns emerge because training data is full of human texts in which anchor values shape judgments. The model learns these statistical correlations along with everything else — without questioning them.

Where the Term Appears in Practice

Anchoring bias is frequently encountered in financial reporting. Analysts who first hear a price forecast rarely deviate strongly from it in their own assessment. The same applies to AI-powered valuation tools that first show users a reference value — that reference value is then rarely corrected significantly, even when the underlying data would suggest it should be.

In AI research, the term plays a growing role in so-called prompt engineering — the deliberate formulation of queries to a language model. Anyone who asks a model to estimate a price while mentioning an example value receives systematically different results than without that hint. Developers of AI applications need to take this into account when designing input fields and user interfaces.

Anchoring bias is also a topic in AI safety and fairness. When historical data — such as past court rulings or credit decisions — serves as the basis for training, reference values set once in the past can permanently act as anchors and systematically distort new decisions.