The Anchoring Effect in AI-Assisted Decisions: Why the First Number AI Shows You Wins

2. Displaying Recommendations

The second pattern is the early display of recommendations. If the AI’s assessment appears before the user has examined the underlying data, the anchor has already shaped the user’s reading of that data. Users then attend more to evidence that supports the AI’s conclusion and less to evidence that contradicts it. This is anchoring compounded by confirmation bias. The AI’s recommendation tells the user what to look for and the user finds it.

3. Narrowing Option Sets

The third pattern is narrow option sets. When an AI presents just one recommended option or ranks options by a clear favorite, users consider a smaller range of alternatives than they would have if they approached the problem without any suggestion. The AI’s recommendation defines the center of the possibility space, and the user’s thinking stays close to that center. Options that are far from the AI’s recommendation receive less consideration—even though they might be superior.

The Adjustment Problem

The anchoring effect works through what Epley and Gilovich have identified as insufficient adjustment. When people start from an anchor, then adjust away from it, they stop adjusting too soon. Each unit of adjustment from the anchor requires cognitive effort, so people stop when they reach a value that feels good enough rather than continuing to the value they would have reached independently.

This effect applies to AI-assisted decisions in business-to-consumer (B2C) settings. For example, when the AI suggests a price of $150 but the user believes the right answer is closer to $120, he is unlikely to adjust all the way to $120. The user settles somewhere between the anchor and his independent estimate—perhaps around $135 or $140. The AI’s recommendation has pulled the user’s final answer toward it by an amount his unaided judgment would not have added.

Now consider a scenario that is common in enterprise information-technology (IT) service management. An incident manager receives an alert regarding a complex database outage. Before the manager can fully assess the diagnostic logs, the AI assistant presents a prominent recommendation: an estimated resolution time of two hours.

The AI’s two-hour projection immediately anchors the assessment of the manager—an experienced professional who might have initially estimated a much more conservative six-hour resolution time for a similar outage. Instead of independently evaluating the problem, the manager unconsciously adjusts his estimate from the AI’s anchor, perhaps ultimately allocating three hours of engineering time. The anchor pulled the manager’s expert judgment downward, which could potentially lead to under-resourcing, delayed communications to stakeholders, and missed service-level agreements (SLAs). In high-stakes enterprise software, a single anchor could cascade into a systemic operational failure.

For designers of AI systems, the most significant finding from anchoring research is that their awareness of this effect does not eliminate it. Participants who are told about anchoring, who know the anchor is arbitrary, and who are explicitly instructed to ignore it still show substantial anchoring effects. This bias operates below the level of conscious correction and has a direct implication for how AI teams typically respond to anchoring concerns. Training users to think independently or to trust their own judgment will not overcome anchoring when an AI’s recommendation is visible during the decision-making process. The cognitive process behind anchoring runs beneath such instructions. One solution is to design user interfaces that do not produce this anchoring condition in the first place.

When Anchoring Helps and When It Harms

For a routine, low-stakes decision, if an AI’s recommendation is significantly more accurate than the average user’s assessment, a user’s anchoring on the AI’s recommendation might not be a problem. In fact, anchoring on an AI’s output could actually improve average decision quality. The user who would have independently guessed $120 but settles on $140, which is closer to the AI’s correct answer of $150, has made a better decision because of the anchor. In such cases, the user interface works well. However, a problem arises in the following three situations:

1. The AI is wrong. An incorrect recommendation that anchors the user’s thinking produces a worse outcome than no recommendation at all because the user adjusts his assessment from a flawed starting point instead of generating his own estimate. In this case, the AI’s error becomes the floor from which the user adjusts his assessment rather than a ceiling he pushes against.
2. The task requires wide consideration of alternatives. Strategic decisions, creative problem-solving, and novel situations can all benefit from a broad search of the possibility space. Anchoring compresses that search. For example, a product team that is evaluating potential markets would consider a narrower range of alternatives if the AI has already flagged its top recommendation. Options that deserve the most consideration precisely because they are outside the AI’s suggested range would receive the least attention.
3. Independent human judgment is the safety check on AI output. This is the situation most relevant to UX researchers and designers who are working on oversight-intensive workflows. The purpose of having a human review an AI output is to catch errors, but direct anchoring undermines that purpose. The reviewer’s judgment is pulled toward the AI’s recommendation, making him less likely to identify the very errors his goal is to catch. The oversight mechanism and the anchoring effect are working against each other, and the anchoring effect tends to win.

Four Design Patterns for Reducing Unwanted Anchoring

The following four design patterns would not be appropriate for every workflow. The question to ask before applying any of these patterns is: Should the user’s judgment be independent of the AI’s recommendation at this decision point? If so, these patterns can help preserve that independence. If not, displaying the AI’s recommendation first is reasonable and most efficient.

Pattern 1: Elicit the user’s assessment before displaying a recommendation.

Ask the user to provide his initial assessment before showing the AI’s recommendation. This creates an internal anchor, the user’s own estimate, which competes with the AI’s anchor. Research shows that self-generated anchors are stronger than externally provided anchors. A user who has committed to his own estimate of $120 is more resistant to the AI’s $150 than a user who encounters the AI recommendation of $150 before forming any estimate at all.

The design trade-off for this pattern is adding a step to the workflow. This might feel slower to users who are accustomed to seeing the AI’s answer immediately. For high-stakes decisions where independent judgment genuinely matters, this trade-off could be worthwhile. In routine decisions where the AI is reliably accurate and human variation adds noise rather than value, the benefit might not justify the added friction. The key is making that determination explicitly for each decision type rather than defaulting to whichever is easier to build.

Pattern 2: Present recommendations as ranges rather than point estimates.

A range such as $130 to $170 preserves more of the user’s judgment space than an explicit amount of $150. The range communicates uncertainty, reducing the anchoring effect by making the recommendation feel less definitive. Users tend to treat a range as a region to explore rather than a point from which to make adjustments. The range also communicates something true about the AI’s actual confidence, which is rarely the certainty that a single number implies.

Ranges work best when they are grounded in something meaningful such as the distribution of outcomes in comparable historical cases rather than acting as arbitrary buffers around a point estimate. A range that reflects real uncertainty teaches calibrated trust. An arbitrary range teaches users to ignore it.

Pattern 3: Show multiple alternatives rather than a single recommendation.

When an AI presents three ranked options rather than one recommendation, the user’s thinking engages with the differences between the options rather than anchoring on a single value. The comparison process activates more independent evaluation than the accept-or-reject process that a single recommendation triggers. Users considering three options must articulate why one is preferable, which requires engaging their own domain knowledge rather than simply adjusting from a starting point.

Multiple options also make the AI’s reasoning more transparent. If the AI ranks Option A above Option B, users can examine what distinguishes the options and form a view about whether the ranking reflects their own priorities. This is a more engaged form of AI review than single-recommendation acceptance.

Pattern 4: Separate the recommendation from the primary decision user interface.

Place the AI’s recommendation in a secondary panel or expandable section that the user can consult rather than in the primary workflow where it becomes the default starting point. The physical separation creates psychological separation. The recommendation becomes available information rather than the opening bid in a negotiation with the user’s own judgment.

This pattern is most appropriate in expert-driven workflows where the user has sufficient domain knowledge to form an independent view and that view is part of what makes the oversight valuable. It is less appropriate for users who are new to a domain and would genuinely benefit from seeing the AI’s recommendation as a learning scaffold.

What UX Researchers Should Look For

The anchoring effect is invisible to the people experiencing it. A user who makes an anchored recommendation believes he has exercised independent judgment, arrived at his answer through his own reasoning, and is unaware that his reasoning started from a point the AI chose. This invisibility makes anchoring particularly difficult to detect through standard product metrics or self-reporting.

In moderated research sessions, the most direct test is a sequence comparison. Run two versions of the same task: one in which the AI’s recommendation is visible from the beginning; another in which the user must form an initial estimate before seeing the AI’s recommendation. Compare the final decisions and the reasoning that the user articulates. In the anchored condition, users often use the AI’s framing as the scaffolding for their explanation without realizing they are doing so. In the unanchored condition, users tend to generate their own framing first, then compare it to the AI’s.

Directly ask users: What would it take for them to move significantly away from the AI’s recommendation? Users who are anchored struggle to answer this question concretely. They know, in the abstract, that they could override the AI. But when asked to describe the evidence or reasoning that would lead them to recommend $90 instead of the AI’s $150, they are often unable to do so. The anchor has narrowed their estimate, as well as their sense of what is possible.

Look at the distribution of user decisions relative to AI recommendations in behavioral data. In a workflow without anchoring effects, user decisions should show meaningful variance from the AI’s output in cases where the AI is uncertain or the user has relevant context that the AI lacks. If the distribution of user decisions clusters tightly around the AI’s recommendations even in high-uncertainty situations, anchoring is likely shaping the output.

Pay particular attention to override patterns over time. New users, who have not yet calibrated their trust in the AI, tend to deviate more from the AI recommendations than experienced users. Some of that convergence reflects genuine learning—that users have discovered that the AI is reliable in particular domains. But some of it is anchoring that has deepened with familiarity. Distinguishing between these two requires a qualitative investigation. Ask experienced users: What is your reason for staying close to the AI’s recommendation? If the answer is “It’s usually right,” that is calibrated trust. If the answer is some version of “I wasn’t sure I had a good reason to change it,” anchoring has done its work.

Designing for Independent Thought

The anchoring effect is a documented feature of human cognition that activates whenever a reference point is present during a judgment task. When an AI user interface displays a recommendation, it provides such a reference point. The question for UX designers is whether anchoring on the AI’s output is the desired outcome at a particular decision point.

For low-stakes, routine decisions where the AI is reliably more accurate than individual users, anchoring on the AI’s output is a reasonable and efficient design choice. For high-stakes decisions, those that require broad exploration of alternatives, or those for which human judgment is a safety check on the AI’s output, anchoring is a design problem that requires a design solution.

Most AI user interfaces display recommendations first because this is a simpler design solution that feels more helpful. Showing the user the answer immediately does feel helpful. But the purpose of keeping a human in an AI-assisted loop is to have someone who can catch the AI when it is wrong. Anchoring makes that person less likely to catch anything at all.

The design investment that is necessary to preserve independent judgment is not large. None of the common solutions is technically complex: eliciting assessments before displaying recommendations, presenting ranges rather than point estimates, offering multiple options, or separating the recommendation from the primary user interface. Each of these design solutions is a deliberate choice to treat the user’s judgment as something worth protecting. Making that choice consistently across high-stakes decision points is the difference between human oversight that functions well and human oversight that seems to, but doesn’t. 

References

Amos Tversky and Daniel Kahneman. “Judgment Under Uncertainty: Heuristics and Biases.” Science, 1974. Retrieved June 10, 2026.

Nicholas Epley and Thomas Gilovich. “The Anchoring-and-Adjustment Heuristic: Why the Adjustments Are Insufficient.” Psychological Science, 2006. Retrieved June 10, 2026.