When AI Says "I Am Unable to Answer": Understanding User Responses to AI Refusals

1 Introduction

• LLMs can produce plausible but incorrect information, commonly called hallucinations, which remain a persistent problem despite progress in capabilities and mitigation strategies.The passage reports 53% population adoption within three years, faster than the PC or internet [61].Source: §1 · p. 2

2.1.1 Mitigating Hallucinations through Refusal.

• Refusal can reduce hallucination risk, but reliable behavior also requires answering correctly and recognizing when to refuse.Prior work describes training systems to answer confident questions and refuse uncertain ones.Source: §2.1.1 · p. 4

2.1.2 Tradeoff Between Hallucination and Over-Refusal.

• Lower hallucination rates can coincide with false refusals of answerable questions, creating a trade-off between reducing hallucinations and over-refusal.One cited benchmark reported that Llama-3.1-405B-Instruct falsely refused 56.77% of answerable questions.Source: §2.1.2 · p. 4

2.1.3 User Evaluation of Hallucinations.

• Users may recognize hallucinations as less accurate yet still be influenced by plausible incorrect answers in their judgments.Prior work found users accepted an incorrect hallucinated answer in more than half of cases.Source: §2.1.3 · p. 5

2.1.4 User Evaluation of Refusals.

• Existing work has largely studied hallucinations and refusals separately, leaving limited evidence on how users compare correct answers, incorrect definitive answers, and refusals across repeated interactions.This study addresses the gap by comparing genuine answers, hallucinations, and refusals within one sitting.Source: §2.1.4 · p. 5

2.2.1 Communicating and Explaining Refusals.

• More informative explanations can improve perceptions of refusals, motivating study of how uncertainty communication affects responses.Prior findings distinguish brief unexplained refusals from more informative responses and report improved perceptions with explanations.Source: §2.2.1 · p. 5

2.2.2 User Perceptions of Repeated Refusals.

• Because users encounter sequences of responses, occasional refusals may signal caution while repeated refusals may signal incompetence; refusal frequency has received little research attention.The work tests whether explanation benefits differ for infrequent versus frequent refusals.Source: §2.2.2 · p. 5

2.2.3 Refusal as a Signal of Caution.

• Occasional refusals might make a system’s answers seem more trustworthy, whereas frequent refusals might instead signal limited capability.Prior work motivates testing whether users judge answers from refusing systems as more accurate, and whether frequency matters.Source: §2.2.3 · p. 6

2.3 Need for Cognitive Closure (NFCC); 2.3.2 NFCC and AI Refusals.

• Need for cognitive closure (NFCC) captures preference for definite knowledge, and unresolved refusals may affect higher-NFCC users more negatively.The study examines whether explanations reduce uncertainty about why a refusal occurred, even though they do not answer the original question.Source: §2.3.1 · p. 6, §2.3.2 · p. 6

3.1 Stimuli Generation and Presentation.

• The study generated matched genuine and hallucinated responses to TruthfulQA questions using GPT-5 and Gemini 2.5 Pro.The first 84 questions were selected, excluding polarizing or obscure topics; both models contributed equally within each response type.Source: §3.1 · pp. 6–7

3.1.1 Prompt Engineering and Stimuli Generation.

• Rule-based prompts used reference answers for genuine responses, generated parallel hallucinations, and standardized refusals with or without epistemic explanations.Explanation reasons were insufficient relevant information or conflicting information.Source: §3.1.1 · p. 7

3.1.2 Stimuli Validation.

• Hallucinations were retained only after manual comparison and agreement by two language models that they did not entail their verified genuine counterparts.The final stimulus set comprised the first 36 questions to pass both checks.Source: §3.1.2 · p. 7
• Participants evaluated 18 questions, with low-refusal conditions containing 2 refusals and high-refusal conditions containing 4, alongside genuine and hallucinated responses.The 36 questions were divided into two groups of 18, and response assignments were counterbalanced across participants.Source: §3.1.2 · p. 7

3.1.3 Stimuli Presentation.

• Responses were pre-generated, anonymized, and revealed after participants clicked an arrow in a randomized Q/A interface.A typewriter animation simulated real-time interaction while keeping responses identical across participants.Source: §3.1.3 · p. 8

3.2 Experiment Method; 3.2.1 Participants.

• The online experiment accepted 599 participants who were US-based adults fluent in English, after excluding one participant who failed an attention check.Participants were recruited through Prolific; conditions were allocated to promote demographic balance.Source: §3.2.1 · p. 8

3.2.2 Procedure.

• Each participant completed 18 randomized Q/A trials, with refusal counts and genuine and hallucinated responses varying by condition.Low-refusal conditions included 2 refusals; high-refusal conditions included 4, with explanation types counterbalanced across questions and models.Source: §3.2.2 · p. 9

3.2.3 Measures.

• Measures covered six-item NFCC, perceived accuracy, satisfaction, perceived AI confidence, overall system evaluation, and exploratory negative perceptions of refusals.Overall system evaluation used 13 items, while the negative-perception score averaged the first six of nine refusal-perception items.Source: §3.2.3 · pp. 9–10

3.3 Data Analysis.

• Repeated response ratings were analyzed with linear mixed-effects models, while overall evaluations were analyzed with linear models and NFCC was tested as a continuous moderator.The analyses used participant- and item-level random intercepts and adjusted prespecified contrasts for multiple comparisons.Source: §3.3 · p. 11

4.1.1 Effects of Response Type, Refusal Frequency, and Explanations

• Across refusal-present conditions, genuine responses were most satisfying, followed by hallucinations and refusals, and refusal exposure did not change satisfaction with definitive responses.All within-condition response-type comparisons were significant; genuine-response and hallucination satisfaction remained stable across conditions.Source: §4.1.1 · pp. 11–12, Figure 5 · p. 12
• For infrequent refusals, explanations raised refusal satisfaction from 1.23 to 1.91 (Δ = .68); for frequent refusals, the increase from 1.24 to 1.41 was nonsignificant.With explanations, satisfaction was higher for low than high refusal frequency (Δ = .50); without explanations, frequency did not affect satisfaction.Source: §4.1.1 · pp. 12–13, Figure 5 · p. 12
• Mean item-level satisfaction was slightly lower with low refusals and explanations than with no refusals (M = 3.21 vs. M = 3.38; Δ = −.16).The reported 95% CI was [−.32, −.001], with p = .049.Source: §4.1.1 · p. 13

4.1.2 Effects of NFCC

• At mean and high NFCC, explanation conditions were associated with higher satisfaction, but not at low NFCC; NFCC did not alter the response-type satisfaction pattern.Neither Response Type × NFCC nor Condition × Response Type × NFCC was significant.Source: §4.1.2 · p. 13
• As NFCC increased, the genuine–hallucination satisfaction gap narrowed from Δ = .84 to .75 to .67, while genuine responses remained more satisfying.This pattern did not significantly vary across experimental conditions.Source: §4.1.2 · p. 13

4.2.1 Effects of Response Type and Refusal Condition

• Genuine responses were perceived as more accurate than hallucinations in all five conditions, with no perceived-accuracy differences across conditions.The accuracy gap also remained similar across conditions, providing no evidence that refusal tendency changed perceived accuracy of definitive answers.Source: §4.2.1 · p. 13, Figure 6 · p. 14

4.2.2 Effects of NFCC

• The genuine–hallucination accuracy gap narrowed from Δ = .93 at low NFCC to .82 at mean NFCC and .72 at high NFCC, while genuine responses remained more accurate.This pattern did not significantly vary across experimental conditions.Source: §4.2.2 · pp. 13–14, Table 3 · p. 14, Figure 6 · p. 14

4.3.1 Effects of Refusal Frequency and Explanations

• Neither refusal presence, refusal frequency, nor explanations significantly affected overall evaluation.The planned contrasts between no-refusal and refusal-present conditions and between low and high refusal frequency were nonsignificant.Source: §4.3.1 · p. 14

4.3 Overall Evaluation

• Overall evaluation did not significantly differ across the five experimental conditions.The results are summarized across all five conditions in Table 4 and Figure 7a.Source: §4.3.1 · p. 14, §4.3 · p. 14, Table 4 · p. 15

4.3.2 Effects of NFCC

• At high NFCC, participants evaluated the no-refusal system more favorably than refusal-present systems (Δ = .41, p = .003), but refusal frequency and explanations did not distinguish evaluations.The no-refusal contrast was nonsignificant at low and mean NFCC.Source: §4.3.2 · p. 15, Figure 7 · p. 15

4.4 Additional Analysis: Perceived AI Confidence

• Perceived confidence ranked genuine responses above hallucinations and hallucinations above refusals; refusal frequency and explanations did not affect refusal confidence.Confidence in genuine responses and hallucinations also did not differ across conditions.Source: §4.4 · pp. 15–16

4.5 Exploratory Post-Session Perceptions of AI Refusal

• Explanations were associated with less negative refusal perceptions, while high-NFCC participants perceived frequent refusals more negatively than infrequent refusals.No significant condition differences were found at low NFCC.Source: §4.5 · p. 16

5.1.1 Satisfaction and Perceived Accuracy Are Distinct Judgments.

• Participants preferred hallucinations to refusals despite rating hallucinations as less accurate, showing that satisfaction and perceived accuracy are distinct judgments.The authors caution that preference or satisfaction should not be treated as interchangeable with accuracy.Source: §5.1.1 · p. 17

5.1.2 Refusals Signal Local Uncertainty, Not a Stricter Answering Threshold.

• Refusals signaled lower confidence in individual responses, but participants did not view definitive answers from refusing systems as more accurate or confident.They therefore did not appear to infer that those systems used a stricter threshold for answering.Source: §5.1.2 · p. 17

5.1.3 Response-Level and System-Level Evaluations Are Distinct; 5.2.5 Evaluating Refusal Policies Across Reliability and User Experience.

• Dissatisfaction with individual refusals coexisted with no direct condition effect on overall system evaluations.The authors recommend measuring response-level and system-level perceptions alongside answer correctness and refusal appropriateness.Source: §5.1.3 · p. 17, §5.2.5 · p. 19

5.1.4 Benefits of Explanations Depend on Refusal Frequency.

• Explanations increased satisfaction with infrequent refusals, but not frequent refusals, and did not directly improve overall system evaluation.The benefit did not reliably increase perceived confidence in the refusal, and low-frequency explanations did not eliminate the satisfaction cost relative to no refusal.Source: §5.1.4 · pp. 17–18

5.1.5 Need for Cognitive Closure and Aversion to Refusals.

• Higher-NFCC participants rated refusal-present systems less favorably than the no-refusal system, although NFCC did not widen the satisfaction gap between definitive answers and refusals.They also showed a smaller perceived-accuracy gap between genuine and hallucinated responses, while participants at all NFCC levels rated genuine responses more accurate.Source: §5.1.5 · p. 18

5.2.1 Using Refusals with Discretion.

• Explanations may make occasional refusals more acceptable, but explanation alone is unlikely to offset the cost of refusing too frequently.The authors recommend using refusals with discretion.Source: §5.2.1 · p. 18

5.2.2 Aligning Refusal Policies With Application Scenarios.

• Refusal policies should reflect application-specific error costs: general-purpose systems should balance reliability and usefulness, while consequential settings may need to prioritize refusal when evidence is insufficient.Users may favor answers even when they are less reliable, so dissatisfaction alone is not sufficient reason to relax a refusal policy.Source: §5.2.2 · p. 18

5.3 Limitations and Future Work

• The study used educated, technologically experienced U.S. Prolific workers and general-purpose questions, experimentally selected refusal rates, randomized stimulus order, and constrained refusal wording and explanations.The authors also note that the pre-generated responses excluded adaptive, multi-turn interaction; explanations were assigned and should not be treated as verified reports of model internals.Source: §5.3 · pp. 19–20

• Users may prefer an incorrect answer to a refusal, while explanations help more when refusals are infrequent.The authors frame refusal as an interaction-level design problem involving frequency, uncertainty communication, and how prior refusals shape later perceptions.Source: §6 · p. 20
• Refusals did not lead users to view a system’s definitive answers as more accurate, and reactions differed with need for cognitive closure.Source: §6 · p. 20
• Optimizing primarily for immediate satisfaction may pressure systems toward incorrect answers rather than appropriate refusals.Source: §6 · p. 20

• The study protocol was approved by the authors’ institutional review board, and participants gave informed consent.The reference number and approval date were to be provided upon acceptance.Source: Ethics and Privacy Statement · p. 20
• Participants were not told in advance that some AI-generated responses were intentionally hallucinated or which responses were hallucinations.Pre-generated responses used a typewriter animation to simulate real-time interaction.Source: Ethics and Privacy Statement · p. 20
• Participants were fully debriefed about the deception and generated responses and given links to support resources.The statement also notes that participant privacy was protected throughout the study.Source: Ethics and Privacy Statement · p. 20

• The authors developed the conceptual ideas, study design, analyses, and interpretations on their own.Source: §A.1 · p. 31

A.2 Prompts for Stimuli Generation

• Hallucinations were generated from each question and its genuine answer by changing important information while keeping the response believable.The prompt also requested two short sentences and prohibited warnings that the response might be incorrect.Source: §A.2 · pp. 31–32
• Refusals were generated from the question, with the same refusal statement used to control for wording differences.The prompt asked for a refusal statement.Source: §A.2 · pp. 31–32
• Genuine responses used the TruthfulQA answer and prompts requiring two short sentences, unrestricted questions, and compliance with the prompt rules.The generated answers were checked and modified to meet those rules.Source: §A.2 · pp. 31–32
• Across response prompts, instructions prohibited irrelevant information and required checking that outputs followed the rules.Source: §A.2 · pp. 31–32

A.3.1 Onboarding.

• Participants rated 18 response sets for accuracy and satisfaction on 1–5 scales and perceived AI confidence on a 0–100 scale.They also provided demographic information and completed a post-session evaluation of the AI system.Source: §A.3.1 · p. 32

A.3.2 Debrief.

• The study initially described its purpose as evaluating answer accuracy, then debriefed participants that it examined perceptions of refusals and used deception.The debrief explained that responses had been pre-generated rather than provided in real time.Source: §A.3.2 · p. 33

A.4.1 Demographic Questions.

• Participants reported age, gender, language, racial or ethnic category, education, field of study, computer expertise, AI expertise, and AI-use frequency.The AI expertise and use questions used five-point scales.Source: §A.4.1 · pp. 33–34

A.4.3 Attention-Check Question.

• Participants answered two instructed-response attention checks, and the study automatically terminated if either answer was incorrect.Source: §A.4.3 · p. 35

A.4.4 Overall Evaluation Measures.

• Overall system evaluations covered warmth, competence, trust, usefulness, and future-use intention.The post-session refusal-perception composite averaged six items selected after factor analysis, covering perceived failure, reliability, frustration, disengagement, task interruption, and preferring an answer to none.Source: §A.4.4 · p. 35, §A.4.5 · p. 35

A.5.1 Satisfaction.

• Genuine responses were more satisfying than hallucinations in all five conditions, and all within-condition response-type comparisons were significant.The genuine-versus-hallucination comparisons had Holm-adjusted p < .001; all within-condition response-type comparisons had Tukey-adjusted p < .001.Source: §A.5.2 · p. 36, Table 6 · p. 36

A.5.4 AI’s Confidence in its Response.

• Across conditions, definitive responses were perceived as 45.58 points more confident than refusals, and genuine responses as 9.51 points more confident than hallucinations.Both contrasts were significant with Holm-adjusted p < .001.Source: §A.5.4 · p. 37
• For definitive responses, no pairwise condition comparison in perceived AI confidence was statistically significant.The supplementary analysis also reported no systematic differences across conditions for definitive responses.Source: §A.5.4 · pp. 38–39
• The Condition × NFCC interaction for overall evaluation was significant, F(4, 589) = 3.28, p = .011.The analyses included all 599 participants, with NFCC treated continuously in the moderation model.Source: Table 16 · p. 40, §A.5.4 · p. 40

A.5.5 Exploratory Post-Session Perceptions of AI Refusals.

• In the refusal-only model, refusal frequency, explanation, and their interaction were not significant predictors of perceived AI confidence.The tests were F(1, 474) = 1.18, p = .277 for refusal frequency; F(1, 474) = 1.17, p = .281 for explanation; and F(1, 474) = 0.44, p = .506 for their interaction.Source: Table 18 · p. 41, §A.5.5 · p. 41

A.6 Questions for Stimuli Generation.

• The appendix lists questions selected from TruthfulQA for stimulus generation.The listed prompts cover topics including health, law, finance, and science.Source: §A.6 · pp. 41–42

A.5.5 Exploratory Post-Session Perceptions of AI Refusals.

• The Condition × NFCC interaction for post-session perceptions of AI refusal was significant, F(4, 589) = 3.04, p = .017.Higher scores indicated more negative perceptions; the note defines negative With −Without estimates as less negative perceptions with explanations and negative Low −High estimates as more negative perceptions under high refusal frequency.Source: Table 21 · p. 42, §A.6 · p. 42

A.5.6 Correlation Analyses.

• The weighted Fisher-z mean correlation between perceived confidence and satisfaction across 22 response instances was r = .558 (95% CI [.523, .595]).All 22 response-instance correlations remained significant at p < .001 after Holm correction.Source: Table 22 · p. 43, §A.6 · p. 43