Researching AI hallucination perception with end users
A practical guide for UX researchers on studying how real users detect, react to, and attribute AI hallucinations in the products they use.
Researching AI hallucination perception with end users
End users do not experience AI hallucination the way researchers measure it. Accuracy benchmarks tell you the error rate; user perception research tells you whether people catch those errors, how they respond when they do, and what happens to trust when they do not. Both questions matter, but only the second one directly predicts product outcomes.
This guide is for UX researchers and mixed-methods teams building AI-assisted products who need a structured approach to studying how users perceive, attribute, and recover from hallucinated output.
Why perception matters more than the error rate alone
A language model that hallucinates five percent of the time sounds manageable in the abstract. In practice, the impact depends entirely on user behaviour. A highly sceptical user who verifies every factual claim will catch most errors before acting on them. A user who has learned to trust the product implicitly may act on a hallucination and only realise the mistake after consequences have occurred.
Research from Nielsen Norman Group on AI chatbot interactions consistently shows that users develop trust heuristics based on early experiences. If the first few responses feel authoritative and accurate, users calibrate upward and become less likely to question later outputs, even when those outputs are wrong. This is not a failure of intelligence; it is a normal human response to apparent reliability.
The goal of hallucination perception research is to map those calibration patterns across different user segments, product contexts, and hallucination types before they become support tickets or reputational incidents.
Three research questions to anchor your study
Before choosing methods, define which question you are actually trying to answer:
- Detection: Can users identify hallucinated output when they encounter it, and what cues do they use?
- Attribution: When users notice something wrong, do they blame the AI, themselves, or the source data?
- Behavioural response: After a hallucination event, how does user behaviour change? Do they verify more, use the product less, abandon it, or continue as before?
Most early-stage teams prioritise detection because it feeds directly into UI decisions (confidence indicators, source citations, error messaging). Attribution and behavioural response studies are higher-value but require longer engagements and more complex research designs.
Research methods by use case
Think-aloud usability testing with injected scenarios
This is the most direct method for detection research. You present users with an AI-powered prototype or live product and ask them to complete representative tasks. Within the session, you seed one or more hallucinated responses: a confidently stated but incorrect fact, a plausible-sounding fabricated reference, or a numerical error in a summary.
The key protocol decisions are:
- Naturalistic vs prompted disclosure. Decide in advance whether to say nothing and observe, or to ask “does anything seem off to you?” after each AI response. The first approach gives you honest detection rates; the second surfaces the reasoning behind detection or failure to detect.
- Severity grading. Design hallucinations across a severity spectrum. An obvious logical inconsistency tests whether users are paying attention at all. A subtle factual error in a domain where the user has expertise tests the quality of their verification behaviour.
- Debrief protocol. Immediately after the session, explain every injected error and why it was included. Give participants the opportunity to ask questions. This is not optional; it is an ethical requirement for any study involving deception.
For B2B AI products, think-aloud sessions are most valuable when participants are recruited by job function and domain expertise, because detection rates differ sharply between, say, a financial analyst reviewing an AI-generated brief and a sales rep using the same tool for a different workflow.
Diary studies for longitudinal perception shifts
Detection studies capture a snapshot. Diary studies capture drift. Over two to four weeks of real product usage, participants log moments when they questioned AI output, verified a claim, or felt surprised by an error. They also record moments when they chose not to verify, and why.
This method is especially valuable for products that users interact with daily because it reveals how trust calibration shifts over time. A participant who starts the diary sceptical and gradually stops verifying is showing a pattern your design team needs to act on before the product is at scale.
Diary prompts for hallucination perception research should include:
- “Did any AI output surprise you today? What was your response?”
- “Did you check any AI-generated information against another source? What prompted that?”
- “Was there a moment today where you felt uncertain whether to trust the AI’s response?”
Keep prompts short and daily. Long weekly reflections tend to miss the low-salience moments, which are often the most revealing.
Semi-structured interviews after exposure sessions
Post-exposure interviews let you probe the mental models users hold about why AI produces errors. These mental models directly shape attribution patterns. A user who believes the AI is “just retrieving web content” will attribute hallucinations to bad source data. A user who understands that language models generate rather than retrieve will have a different attribution pattern and, typically, a more calibrated trust level.
Interview questions to include:
- “When the AI gave you that incorrect information earlier, what did you think had happened?”
- “If you had to explain to a colleague why AI sometimes gives wrong answers, what would you say?”
- “Has a moment like that changed how you plan to use the tool going forward?”
For more on structuring qualitative interview sessions for AI products, see user research for AI products in 2026.
Quantitative trust scales
For validation at scale, pair your qualitative findings with a validated instrument. The Trust in Automation scale developed by Jian et al. is widely used in HCI research and provides a repeatable baseline you can compare across product versions, user segments, or time points. The Semantic Differential method, where users rate the AI on paired adjective scales such as reliable-unreliable and accurate-inaccurate, is faster to administer and easier to parse for product teams without research backgrounds.
Run the scale before and after a structured exposure session to measure trust change as a direct result of a hallucination event. This gives you a quantifiable signal that complements the qualitative “why” from interviews.
Screener design for this research
Who you recruit matters as much as the method you use. The following screener criteria are specific to hallucination perception studies.
| Criterion | Rationale |
|---|---|
| Weekly AI tool usage in the relevant category | Ensures participants have a baseline mental model of AI behaviour |
| Domain expertise level (novice, intermediate, expert) | Detection and attribution patterns differ significantly by expertise |
| Self-rated confidence in evaluating AI output | Identifies users most at risk of uncritical acceptance |
| Verification habit (do they fact-check AI responses?) | Baseline behaviour before any hallucination exposure |
| Professional context (B2B: role, tenure, stakes of decisions) | High-stakes users show different recovery behaviours |
For consumer AI products, also screen by how frequently participants share AI output with others. Social propagation of hallucinated content is a downstream risk that differs from individual decision-making.
Ethical considerations
Hallucination perception research involves a form of deception: participants do not know in advance that specific errors have been seeded. Three practices protect participant wellbeing and research integrity:
- Consent without full disclosure. Consent forms should state that the study involves interacting with an AI system and that you are studying how people engage with AI output. They should not mention injected errors specifically, as this would invalidate the study.
- Full debrief. Every participant receives a complete explanation of the seeded errors and the research rationale before leaving the session.
- Harm avoidance. Never seed hallucinations in domains where acting on false information could cause real harm to the participant. This includes medical symptoms, financial decisions, legal questions, and safety-critical information.
For broader guidance on ethical research design for AI products, the ACM Code of Ethics and Belmont Report principles provide the foundational framework.
What to do with your findings
Hallucination perception research typically surfaces three types of actionable output:
UI and copy changes. If detection rates are low for a particular output type, the product team can add source citations, confidence indicators, or explicit “AI may make mistakes” prompts at the relevant touchpoints. If attribution is skewed toward self-blame, the UI language may be inadvertently creating false confidence.
Onboarding design. Users who receive structured onboarding about AI limitations show better calibrated trust at four weeks than users who self-discover the same limitations. This is consistently one of the highest-leverage findings from hallucination perception studies.
Participant segmentation for follow-up research. Perception studies almost always reveal that two or three user segments behave very differently. The segment of uncritical acceptors and the segment of over-sceptical abandoners both represent product risks worth separate research tracks.
For related reading on the researcher-facing side of this problem, see AI hallucination in research analysis: real risks.
Recruiting participants for AI perception research
Recruiting the right participants is the single biggest constraint for this type of study. General-purpose panels often lack the filtering depth needed to isolate AI tool users by category, expertise level, and verification behaviour in a single screener.
CleverX’s panel of 8 million-plus verified B2B and B2C participants across 150-plus countries supports multi-variable screening that fits hallucination perception research well: filter by AI tool category, professional role, domain expertise, and self-reported trust levels in a single pass. The platform also supports both moderated interview recruitment and survey distribution, so teams can run a qualitative discovery phase and a quantitative validation phase without switching vendors.
For tool selection in the moderated session phase, AI usability testing tools in 2026 covers the current options.
For more on structuring qualitative interview sessions generally, see how to conduct effective user interviews.
Frequently asked questions
What does ‘AI hallucination perception’ mean in UX research?
AI hallucination perception refers to how end users notice, interpret, and respond when an AI system produces incorrect, fabricated, or confidently wrong output. Unlike researcher-facing hallucination risk (where the tool distorts analysis), this field of study focuses on the user experience: does the person catch the error, trust the output anyway, lose confidence in the product, or change their behaviour? Understanding this perception is essential for any team designing AI-assisted products.
Why should UX researchers study hallucination perception separately from accuracy metrics?
Accuracy metrics tell you how often the AI is wrong, but they do not tell you how users respond to being misled. Two users who encounter the same hallucination may react very differently: one verifies every response, another accepts it uncritically. Perception research uncovers the mental models, trust thresholds, and recovery behaviours that shape real-world risk. Without it, product teams are optimising for benchmark performance while missing the human factors that determine adoption and harm.
What research methods work best for studying AI hallucination perception?
Think-aloud usability testing with injected hallucination scenarios is the most direct method. Diary studies capture how perception shifts over extended real-world use. Semi-structured interviews after exposure sessions surface the mental models and attribution patterns users hold. Survey instruments such as the Semantic Differential Scale or adapted trust scales like the Trust in Automation scale provide quantifiable comparison data. Most teams combine at least two methods: a qualitative session to build hypotheses and a quantitative instrument to validate at scale.
How do I design hallucination scenarios for usability sessions without deceiving participants unethically?
The standard approach is to disclose in the consent form that the session includes an AI system and that you are studying how people interact with AI output, without specifying that some outputs may contain errors. After the session, conduct a full debrief explaining exactly what was injected and why. This is consistent with accepted standards in HCI research. Avoid scenarios where a false output could cause real-world harm to the participant, such as medical or financial decisions, unless your IRB protocol specifically addresses those conditions.
What screener criteria should I use when recruiting participants for hallucination perception research?
Screen for regular AI tool usage (at least weekly interaction with a relevant category: writing assistants, search copilots, customer-facing chatbots, coding tools), the level of domain expertise (novices and experts respond differently), and self-reported confidence in evaluating AI output. For B2B products, also screen by job function and tenure so you can compare high-stakes professional users with casual adopters. Avoid over-indexing on tech-savvy participants; low-familiarity users often show the most consequential trust patterns.
How does CleverX help teams recruit participants for AI perception studies?
CleverX provides access to a verified panel of 8 million-plus B2B and B2C participants across 150-plus countries. For AI hallucination perception research, you can filter by AI tool usage, professional role, domain expertise, and tech comfort level in a single screener pass. The platform supports both moderated interview recruitment and survey distribution, so teams can combine a qualitative discovery phase with a larger quantitative validation phase without switching vendors or re-recruiting from scratch.