Is this the same as automation bias?

They can overlap, but automation bias specifically concerns favoring automated advice; interaction bias includes other user-driven patterns.

Can explicit ratings be biased too?

Yes. Asking directly does not remove expectations, inconsistent criteria, or selective participation.

Does correcting for result position remove every problem?

No. Position is only one influence on behavior, and any correction depends on assumptions.

AI Glossary / Governance, ethics and safety / Advanced

User Interaction Bias

Definition

User interaction bias arises when people's existing biases or patterned behavior shape their interaction with data, AI outputs, or results. Those interactions can distort judgments and the feedback recorded by a system. In the supplied ontology, the term emphasizes the user's contribution while recognizing that presentation and workflow can influence behavior.

What it means

People do not approach AI systems as neutral recording devices. They bring expectations, habits, preferences, and constraints. These can affect which results they inspect, which suggestions they accept, and which outputs they rate or correct.
The resulting activity record is evidence of behavior in a particular setting, not automatically an objective measure of quality. A click may reflect visibility or curiosity. Accepting an answer may reflect agreement with an existing belief rather than checking its factual support.
This does not mean that every preference is a harmful bias or that user feedback is useless. The issue is whether behavior systematically distorts the conclusion or learning signal for the task being attempted.

How it works

A user encounters information through an interface. Their expectations and behavior influence attention, interpretation, selection, and feedback. If the organization treats those actions as direct evidence of correctness or relevance, the user's pattern can enter evaluation or training data.
The system can then reinforce that pattern by favoring outputs associated with positive feedback. This feedback loop is possible, but it depends on whether and how the system learns from interactions; not every product updates itself from every click.
Research on search behavior shows that clicks can be informative while also reflecting ranking and surrounding results. The lesson is to interpret interaction evidence in context. Separate the observation, such as a click or approval, from the stronger claim it is supposed to support.

Three practical examples

A reviewer approves an AI explanation more readily when it supports their existing view. If approvals are treated as factual-quality labels, the resulting evaluation may reward agreement rather than evidence.
Users open highly ranked search results partly because they expect the first results to be useful. A learning system that treats every click as an unbiased relevance judgment can inherit effects of the earlier ranking.
A staff member repeatedly asks an assistant to revise a report until it supports a preferred recommendation. The final output may appear independently generated even though the interaction deliberately selected for a particular conclusion.

Think of it this way

Think of a suggestion box in which people submit comments after seeing a prominently displayed example. The comments may still contain useful information, but they reflect who responded, what they saw, and how they interpreted the invitation. Interaction with an AI system likewise produces situated evidence rather than an untouched record of underlying truth.

Worked example

Consider an invented evaluation of 40 AI answers. A reference assessment classifies 20 as factually correct and 20 as incorrect under the task's stated criteria.
Twenty answers support a reviewer's prior opinion; ten of these are correct and ten incorrect. The other 20 conflict with that opinion and have the same split: ten correct and ten incorrect.
Suppose the reviewer approves every answer that supports their opinion and rejects every answer that conflicts with it. Approval rates are 20/20 = 100% for agreeing answers and 0/20 = 0% for disagreeing answers, although both groups are 50% correct.
Only ten of the 20 approved answers are correct, so approval precision is 50%. Ten correct answers are rejected. Using these approvals as correctness labels would encode the stipulated agreement preference rather than the reference criterion.
This construction illustrates a mechanism, not a claim about typical users. In a real study, differing approval rates could have several explanations. Establishing the cause would require evidence about the task, answers, presentation, and reviewers' reasoning.

Illustrative case study

Fictional company Cedar Insights asks staff to rate an AI assistant's draft analyses. A thumbs-up button is described informally as “good answer,” leaving unclear whether staff should judge factual support, writing style, or agreement with the conclusion.
The evaluation team reviews a sample of interactions and finds that those meanings differ between reviewers. Some approve fluent drafts that reinforce their expectations, while others reject supported findings they find surprising.
The company separates its evaluation questions, supplies relevant reference material, and examines a sample of ratings independently. It also documents whether feedback is used for product monitoring or model improvement.
This fictional case illustrates investigating what feedback actually represents. It does not assume that independent reviewers are infallible or that clearer rating instructions remove every source of bias.

Common misconceptions and mistakes

“Users' actions reveal objective quality.” Actions can reflect several influences besides quality.
“Every user preference is a bias to eliminate.” Preferences may be relevant; their role must match the task.
“The model alone determines the outcome.” Users can shape prompts, selections, edits, and final decisions.
“All clicks immediately retrain the system.” Learning depends on the actual product design.
“A human review step automatically removes bias.” Reviewers can introduce or reinforce their own patterns.

Business use

Define what each feedback signal is meant to measure. Avoid combining factual correctness, usefulness, style, and agreement into an unexplained approval count. Document who can provide feedback and what information they have when doing so.
Compare interaction measures with other relevant evidence, such as independent task assessments or final outcomes. Examine how presentation influences attention and whether important users or cases are missing from the recorded activity. Before using feedback for training, test the assumptions connecting behavior to the desired label. Retain uncertainty where those assumptions remain weak.

Check your understanding

Why are the worked-example approval labels unreliable for correctness?

Answer: They track agreement even though correctness is balanced across both sets.

Why distinguish monitoring from model training?

Answer: Feedback affects the system differently depending on how it is used.

What should accompany a click count?

Answer: Context about exposure, presentation, and the conclusion being drawn from it.

Related terms

Learn next

Choose one interaction signal from an AI product. List three reasons someone might produce it, identify the intended interpretation, and design a check that could reveal whether the signal supports that interpretation.

Intended interpretation

Sources

Artificial Intelligence Ontology, User Interaction Bias: https://raw.githubusercontent.com/berkeleybop/artificial-intelligence-ontology/main/aio.owl NIST SP 1270, user interaction and related bias definitions: https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.1270.pdf Joachims and colleagues, Evaluating the Accuracy of Implicit Feedback from Clicks and Query Reformulations in Web Search, 2007: https://www.cs.cornell.edu/People/tj/publications/joachims_etal_07a.pdf

Previous: User-defined function Next: User matrix