When humans are asked to make numerical estimations or complex decisions, they rarely calculate an absolute value from first principles. Instead, they rely on a deeply ingrained cognitive shortcut: they identify an initial reference point—an anchor—and adjust away from it until they reach a plausible answer.
However, this adjustment process is systematically flawed. Across diverse domains, human decision-makers consistently fail to adjust far enough from their starting points. This phenomenon, known as insufficient adjustment, leaves final judgments heavily biased toward whatever arbitrary or structured information was encountered first.
Understanding why this bias occurs, where it manifests, and how we can design systems to mitigate it is essential for building robust human-in-the-loop systems.
The Stopping Rule: Why Adjustment Fails
To understand why adjustment is insufficient, we must look at how the mind decides to stop thinking. As Yufan Liu explains in A literature review of a cognitive heuristic: The anchoring effect, the adjustment process does not aim for an objective center of truth. Instead, a decision-maker moves away from the anchor step-by-step and terminates the search the moment they hit the nearest boundary of a subjective "range of plausibility." Because the search stops at the very edge of what seems acceptable, the final estimate remains disproportionately close to the original anchor.
[Low Anchor] ----> (Plausible Range Boundary) . . . . [True Value] . . . . (Plausible Range Boundary) <---- [High Anchor]
^ ^
Search Stops Here Search Stops Here
This premature termination is not a sign of random human irrationality; rather, it is a highly structured, resource-rational strategy. Falk Lieder et al. demonstrate in The anchoring bias reflects rational use of cognitive resources that the mind acts as a bounded agent balancing a trade-off: the cognitive cost of mental effort versus the cost of estimation error. When high precision yields diminishing returns, stopping at the boundary of plausibility is mathematically optimal for conserving limited mental energy. Supporting this view, Taylor A. Burke shows in Mental control of uncertainty that such cognitive biases are not fixed biological flaws, but strategic adaptations that adjust dynamically based on incentives for accuracy.
The Illusion of Expertise
A common misconception is that professional expertise or domain knowledge shields individuals from this bias. Empirical evidence strongly refutes this. The authors of The Trap of Anchoring Effect: A Study on the Impact of Anchoring Effect on Decision-Making emphasize that even highly trained practitioners in high-stakes environments subconsciously anchor on initial figures, leading to severe market distortions and skewed expert opinions.
This vulnerability spans several critical fields:
- Software Engineering: In a study on project planning, Jorge Aranda found in Anchoring and adjustment in software estimation that providing estimators with an initial suggestion—even an arbitrary one—strongly pulled their final estimates. Notably, high anchors exerted a significantly more powerful distorting effect than low anchors, and experienced professionals were just as susceptible as novices.
- Clinical Medicine: In healthcare, Gil Alterovitz et al. note in Predicting ward transfer mortality with machine learning that clinicians frequently fall into anchoring traps, maintaining an initial diagnosis for a patient even when subsequent clinical data points to a worsening, entirely different condition.
- Academic Peer Review: The peer review process relies heavily on a rebuttal phase where authors correct reviewer misconceptions. However, Yiqiao Jin et al. utilize LLM simulations in AgentReview: Exploring Peer Review Dynamics with LLM Agents to show that initial impressions create a cognitive anchor that author rebuttals struggle to overcome. Interestingly, a randomized controlled trial by Ryan Liu et al. (Testing for Reviewer Anchoring in Peer Review: A Randomized Controlled Trial) suggests that when corrections are strictly objective, reviewers can successfully update their scores, highlighting that the nature of the counter-information dictates whether the anchor holds.
Because anchors are so powerful, they can also be intentionally deployed to manipulate behavior. For example, Lara Gauder et al. developed a protocol in A protocol for collecting speech data with varying degrees of trust that successfully manipulated user trust in a virtual assistant simply by priming them with a high or low initial performance rating. Similarly, in opinion dynamics, Abdullah Al Maruf et al. (Learning Dissemination Strategies for External Sources in Opinion Dynamic Models with Cognitive Biases) demonstrate that external actors must explicitly model these subjective reference points to successfully nudge public opinion.
Designing Around the Bias
If insufficient adjustment is an inescapable feature of human cognition, how do we build reliable systems that involve human input? Researchers have proposed two primary engineering paradigms: algorithmic de-biasing and structural choice architecture.
1. Algorithmic De-Biasing (Interactive ML)
When humans interact with machine learning models, they are often asked to provide feedback on the model's current predictions. This creates a feedback loop: the user sees the model's output, gets anchored by it, and provides a biased correction. This "double use of data" causes the model to overfit to its own errors.
To break this loop, Pedram Daee et al. proposed a solution in User Modelling for Avoiding Overfitting in Interactive Knowledge Elicitation for Prediction. They constructed a Bayesian user model that mathematically formalizes the user's anchoring behavior. By explicitly modeling how much the user's response was pulled by the model's current state, the system can back-calculate and infer what the user actually knew before the machine influenced them, drastically improving prediction accuracy.
[Model Prediction] ---> (Anchors User) ---> [Biased User Input]
|
[Bayesian User Model]
| (Subtracts Anchor Effect)
v
[True, Independent Knowledge]
2. Structural Choice Architecture (Iterative Local Voting)
Instead of trying to mathematically remove bias after the fact, we can redesign the user interface so that the anchoring effect works for us rather than against us.
In large-scale collective decision-making (such as determining a complex national budget), asking participants to propose absolute values from scratch triggers severe anchoring vulnerabilities. Nikhil Garg et al. address this in Collaborative Optimization for Collective Decision-making in Continuous Spaces by introducing Iterative Local Voting (ILV).
Instead of asking for absolute values, ILV presents voters with a current candidate solution and asks them only to make small, local adjustments (e.g., "should this specific budget item be slightly higher or lower?"). By transforming a global estimation task into a series of local, incremental choices, the algorithm drastically lowers the cognitive burden on participants. Remarkably, the authors prove that this local adjustment process mathematically converges to socially optimal outcomes, turning a human cognitive limitation into a reliable engine for consensus.
Key Foundational Work
Anchoring and adjustment in software estimation (2005) — contributed 8 concept elements to this direction.
The anchoring bias reflects rational use of cognitive resources (2017) — contributed 1 concept elements to this direction.
Collaborative Optimization for Collective Decision-making in Continuous Spaces (2017) — contributed 1 concept elements to this direction.