PaperlayerView exact revision on arXiv

Source-linked AI summary

Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness

Varshini Elangovan, James Wedgwood, Chhavi Yadav, William Agnew, Sauvik Das, Virginia Smith

arXiv:2609.26865v2cs.HCcs.AIcs.CYcs.LG

TL;DR

Everyday chatbot risks can be difficult for users to detect, and model-level safeguards may not make those risks legible. Safety Nudges adds contextual flags to chatbot conversations; in a two-week study, users generally found it useful and nearly all reported greater awareness, though awareness did not necessarily lead to behavioral change.

  • Problem

    Potential harms such as hallucination, sycophancy, and overconfidence can be difficult for users to recognize during everyday chatbot conversations.

    Source: §1 · p. 2

  • Method

    Safety Nudges is a browser extension that audits chatbot turns with an external LLM and presents contextual flags during a two-week field study with 45 users.

    Source: §1 · p. 3, §3 · p. 4

  • Results

    Participants generally found Safety Nudges useful and understandable, and nearly all reported greater awareness of potential harms, but awareness did not necessarily produce discernible behavioral changes.

    Source: §6 · pp. 24–25, §8 · p. 27

  • Takeaways & Limitations

    User-facing nudges may complement model-level safeguards by helping users critically evaluate chatbot responses, with relevance, calibration, and user control important to their design.

    Source: §1 · p. 2, §5.5.3 · p. 23

    Observed behavior changes are descriptive rather than causal because the fixed study order and unobserved verification outside supported chatbot interfaces limit attribution and coverage.

    Source: §7 · p. 26

Abstract

from arXiv · show

Conversational AI systems can pose safety risks to their users such as hallucination, sycophancy, overconfidence, and anthropomorphism, but these risks are difficult for users to detect during everyday use. We introduce Safety Nudges, a browser-based tool that provides lightweight, in situ flags when concerning behavior is detected in chatbot conversations. We evaluated Safety Nudges in a two-week field study with 45 frequent chatbot users, collecting interaction logs, surveys, and feedback on individual nudges. Participants found the tool useful, clear, and minimally disruptive, with nearly all users reporting an increased awareness of potential AI harms, though we found that this improved awareness alone did not necessarily lead to discernible behavioral changes. Our results suggest that user facing safety nudges can complement model-level safeguards by helping people critically evaluate AI responses in context, while highlighting the importance of relevance, calibration, and user control in nudge design for conversational AI safety. The code for our Safety Nudges extension is publicly available at https://github.com/jtbwedgwood/safety-nudges.

1. Introduction

Everyday chatbot risks can be hard for users to recognize, while model-level safeguards may not make those risks legible. Safety Nudges explores contextual, user-facing flags as a lightweight way to support scrutiny without blocking chatbot use.

1. Introduction

  • Interface nudges can provide contextual cues that prompt reconsideration while preserving user autonomy, rather than preventing an action outright.The paper hypothesizes that this approach suits AI safety because users may need to scrutinize a response, not stop interacting with the system.Source: §1 · p. 2
  • User-facing nudges can complement model-level safeguards by helping users understand why an interaction may warrant caution and critically evaluate AI systems themselves.Unlike system-level mechanisms, nudges need not simply block or modify an interaction.Source: §1 · p. 2
  • The tool is designed to make concerning interaction patterns visible in situ without blocking content or preventing users from continuing conversations.Its stated goal is to support AI literacy and more mindful interaction.Source: §1 · p. 3
  • The study examines changes in users’ risk awareness and interaction patterns, alongside whether nudges feel relevant, informative, seamless, and non-disruptive during everyday use.Participants used Safety Nudges for two weeks, followed by surveys and feedback on individual interventions.Source: §1 · p. 3
  • Safety Nudges provides contextual flags for potentially harmful LLM behaviors, including overconfidence, sycophancy, anthropomorphism, and persuasive or manipulative language.The work also evaluates awareness effects in a field study of 45 chatbot users and analyzes relevance, calibration, usability, adoption intention, and behavioral impact.Source: §1 · p. 3
  • Nearly all participants in the two-week field study reported greater awareness of potential AI harms, though awareness did not necessarily produce discernible behavioral changes.The study combined interaction logs, surveys, and feedback on individual nudges.Source: §1 · p. 3

2. Related Work

Prior LLM safety work mainly changes, filters, or restricts model behavior, while interface nudges offer a complementary way to support user interpretation. Safety Nudges extends interface interventions toward broader conversational safety awareness.

2. Related Work

  • Safety Nudges leaves chatbot interactions intact and adds an independent user-facing layer to help people recognize potentially concerning behavior.This complements approaches that change, filter, or restrict model behavior.Source: §2 · p. 4
  • Prior interface interventions address dimensions such as uncertainty, factual reliability, overreliance, or privacy; Safety Nudges evaluates multiple classes of concerning conversational behavior and surfaces judgments with supporting evidence.Users remain free to decide whether and how to respond.Source: §2 · p. 4
  • Conversational risks can be personalized, ambiguous, and context-dependent, making calibration, explanation, and user control important design concerns.The same chatbot behavior may be benign in one context and concerning in another.Source: §2 · p. 4

3. Safety Nudges: Tool Design and Implementation

Safety Nudges audits chatbot turns with an external language model and presents contextual, inspectable flags without interrupting the conversation. Its design combines a transcript-auditable harm taxonomy, calibrated judgments, and user controls.

3. Safety Nudges: Tool Design and Implementation

  • Safety Nudges audits each user-message/chatbot-response turn with an external LLM, then displays detected issues and supporting text spans in a response-local interface.Figure 2 traces detection and transcript extraction through external review, validation, and nudge display.Source: §3 · pp. 4–5, Figure 2 · p. 7

3.1. Design Goals

  • The design goals are relevant, actionable judgments; external oversight from a different model provider by default; and lightweight interventions users can pause or adjust.The tool does not block prompts, interrupt response generation, modify chatbot outputs, or require users to resolve warnings.Source: §3.1 · p. 5

3.2. Harm Categories

  • The taxonomy groups transcript-auditable concerns into four umbrellas: safety, welfare, and privacy; reliability and scope; relational and social integrity; and governance and misuse.It uses fourteen categories and permits an “other” label for issues that do not fit neatly.Source: §3.2 · p. 5, Table 1 · p. 6
  • The taxonomy emphasizes conversation-visible harms and excludes broader deployment consequences that generally cannot be established from a single transcript.Its operational categories subdivide epistemic and functional failures, including overconfidence, potential hallucination, capability misrepresentation, and scope overreach.Source: §3.2 · pp. 6–7

3.3. Harm Detection Design

  • An external LLM receives the taxonomy, classification instructions, and conversation transcript, then returns a category with rationale, risk level, and evidence spans.Evidence spans ground the diagnosis in text from the recent exchange.Source: §3.3 · p. 8

3.3.1. Prompt Selection and Validation

  • On a 100-conversation gold set, the selected prompt achieved 75% conversation-level issue accuracy and 51% exact gold-set category matches.The independent annotator scored 71% and 45%, respectively; the authors state that classifier optimization was not the study’s principal objective.Source: §3.3.1 · p. 8

3.4.1. Browser Integration and Turn Detection

  • The Chrome extension supports ChatGPT and Claude, detects completed turns, and limits extracted history to at most 12 messages and 12,000 characters overall.The latest exchange is preserved while older messages are removed first, bounding request size and latency.Source: §3.4 · p. 9, §3.4.1 · p. 9

3.4.2. External Audit and Model Routing

  • The worker routes audits through a managed relay to an external model, validates and normalizes judgments, and matches evidence spans to the current exchange where possible.Users can inspect issue categories, severity, rationales, and highlighted evidence on demand; they can also pause analysis, adjust sensitivity, and optionally submit feedback with consent.Source: §3.4.2 · p. 9, §3.4.3 · p. 10, §3.4.4 · p. 10

4. Field Study Evaluation

The evaluation followed frequent chatbot users during two weeks of ordinary use, first without visible nudges and then with them enabled. Logs, surveys, and participant feedback supported mixed-method analysis of awareness, interaction behavior, and tool experience.

4. Field Study Evaluation

  • The two-week field study prioritized everyday chatbot use and collected extension-recorded interactions, survey responses, and feedback on nudges.The study protocol received Institutional Review Board approval.Source: §4 · p. 10

4.1. Participants and Recruitment

  • Forty-five adults who reported using chatbots at least one hour daily completed both study weeks; participants were U.S. residents using Chrome and ChatGPT or Claude.Eligibility also required consent to collect conversation text and usage information.Source: §4.1 · p. 10

4.2. Study Procedure and Design

  • Week 1 recorded chatbot interactions and audit outcomes without displaying nudges; Week 2 enabled visible nudges during participants’ usual chatbot use.Participants were asked to rate at least 25 nudges, or all they saw if fewer appeared.Source: §4.2 · p. 11
  • The study combined turn-level interaction and audit logs with pre- and post-tool surveys to examine awareness, trust, interaction behavior, and perceptions of the tool.The optional follow-up survey asked whether changes in trust, caution, and verification persisted two to three weeks later.Source: §4.2 · p. 12

4.3. Analysis Methods

  • Matched survey items were compared with paired Wilcoxon tests and Holm correction, but week-to-week differences cannot be attributed solely to Safety Nudges.Qualitative analysis used inductive thematic coding of open-ended survey responses and nudge-feedback comments, with researchers reviewing and revising LLM-assisted candidate codes.Source: §4.3 · p. 13

5. Results

The results examine participants’ responses to Safety Nudges, including reported awareness and interaction routines, perceived value, and concerns about warning accuracy, disruption, reassurance, privacy, and user control.

5.1. Safety Nudge Exposure during Chatbot Use

  • After grouping recurring detections into warning episodes, roughly three in ten analyzed responses were associated with an episode in each study phase.The episode rates were 30.4% in Week 1 and 31.2% in Week 2; most detections were low severity, and potential hallucination was the most frequent low-severity category. Of 647 visible-nudge ratings, 67.2% were helpful.Source: §5.1 · p. 14

5.2. Conversation Topic Analysis

  • Risk type tracked conversation purpose more than detection rate: hallucination and overconfidence led in information seeking, anthropomorphizing in self-expression, and health or legal reliance in practical guidance.Physical-health conversations were under 6% of the sample but had 68.8% of turns flagged, with health or legal reliance accounting for 90.6% of those flags; financial-advice conversations had a 57.8% flagged-turn rate.Source: §5.2 · p. 16

5.3.1. Changes in Perceived Correctness and Trust

  • Chatbot correctness ratings fell, but overall trust did not significantly change; follow-up accounts nevertheless described greater scrutiny of individual responses.Correctness ratings decreased from M=4.07 to M=3.87 (pHolm = .040), a change across periods not attributable solely to Safety Nudges; overall trust decreased by Δ = −0.07 (pHolm = 1.000).Source: §5.3.1 · pp. 17–18

5.3.2. Changes in Risk Concern and Awareness

  • 44 of 45 participants said Safety Nudges increased their awareness of potential AI harms, making specific risks more salient without significantly increasing general concern.Overconfidence (n=36), potential hallucinations (n=31), flattery or sycophancy (n=21), and anthropomorphizing (n=15) were most frequently reported; these were retrospective evaluations, not direct pre–post comparisons.Source: §5.3.2 · pp. 18–19

5.4.1. Changes in Interaction Practices and Verification Behavior

  • Reported interaction changes depended on timing: paired measures showed no general increase during tool use, while retrospective accounts described selective scrutiny and verification.Seeking alternative viewpoints was the only practice with a statistically significant paired decrease; after the study, 26 participants reported verifying more, often focusing on questionable or consequential claims.Source: §5.4.1 · pp. 19–20

5.4.2. Observed Responses Following Flagged Chatbot Turns

  • Direct scrutiny followed 85 hidden flagged turns (8.7%) and 114 visible nudges (10.8%); among 40 participants with usable events in both phases, the paired difference was not significant (Wilcoxon W=239, p=.321).Challenges or corrections increased from 2.4% to 6.2%; fact-checking remained similar (1.2% vs. 1.4%), while source requests decreased (5.0% vs. 3.1%).Source: §5.4.2 · p. 20

5.5.1. Overall Evaluations and User Experience

  • Most participants rated Safety Nudges easy to use, clear, appropriately timed, minimally disruptive, and useful overall.44 of 45 agreed the tool was easy to use and that nudges were clear; 43 agreed it was useful overall. Of 42 participants recalling a specific nudge or conversation, 34 considered it relevant and 37 considered it not or only slightly disruptive.Source: Figure 6 · p. 21, Table 5 · p. 21, §5.5.1 · p. 22

5.5.2. Helpfulness and Interpretation of Audit Outcomes

  • Visible-nudge feedback was mixed, and participants wanted warnings to be specific and context-sensitive rather than repetitive or poorly timed.435 of 647 ratings (67.2%) marked nudges helpful; health or legal reliance and overconfidence had the highest helpful-rating proportions among frequently rated categories, while anthropomorphizing had the lowest. Some users interpreted “No issues detected” as confirmation of safety or accuracy, although it only meant configured risks were not detected.Source: §5.5.2 · pp. 22–23, Table 6 · p. 22, §5.5.3 · p. 23

5.5.3. Contextual Limitations and Sources of Friction

  • Participants said warnings worked best when specific, credible, and directed at potentially high-stakes conversations; contextual accuracy and control over warning frequency were conditions for continued use.Source: §5.5.3 · p. 23

5.5.4. Recommendation and Relevant User Groups

  • The mean recommendation rating was 8.38 (SD=1.97), and the descriptive Net Promoter Score was +46.7, with 28 promoters, 10 passives, and seven detractors.The median recommendation rating was 9.Source: §5.5.4 · p. 23
  • High recommenders described nudges as reminders to verify chatbot claims, recognize limitations, and exercise caution, especially with sensitive information or consequential advice.Participants also saw potential value for inexperienced users and people who rely heavily on chatbots.Source: §5.5.4 · p. 23
  • Participants identified everyday chatbot users (39; 86.7%), older adults unfamiliar with AI (37), and professionals using AI for high-stakes decisions (37) as relevant groups.Other selected groups included students (36), AI researchers and developers (34), people experiencing emotional distress or mental health challenges (33), and children or teenagers (32); selections were not mutually exclusive. Recommendations were conditional on warning accuracy and frequency, perceived harmlessness, user experience, and extension privacy.Source: §5.5.4 · pp. 23–24

5.5.5. Desired Improvements

  • Fourteen participants requested explanations and supporting evidence, including identifying the concerning statement, explaining the risk, using simpler language, and offering sources or follow-up questions.They believed these changes would clarify what needed verification.Source: §5.5.5 · p. 24
  • Participants requested more customization and control (n=11), greater accuracy and relevance (n=10), and interface improvements (n=10).Suggestions included controls for warning frequency and categories, fewer false positives, prioritizing serious risks, more visible controls, less intrusive positioning, reviewable nudge histories, and support for additional chatbot platforms.Source: §5.5.5 · p. 24

6. Discussion

Safety Nudges made potential risks in everyday chatbot interactions more visible, but increased awareness did not reliably translate into changed behavior. The discussion highlights context-sensitive auditing, calibrated trust in both chatbots and auditors, and warning designs that balance usefulness with disruption.

6. Discussion

  • Nearly all participants reported greater awareness of potential harms, but awareness did not uniformly reduce trust or increase verification.Safety Nudges drew attention to confident but unsupported claims, anthropomorphic language, and consequential advice while allowing participants to continue their tasks.Source: §6 · p. 24

Supporting Context-Dependent Audits of AI Chatbot Responses

  • Independent, context-dependent audits can connect general risk awareness to specific chatbot responses, but auditors may lack relevant context or miss how a user’s purpose shapes a warning’s significance.The discussion suggests making uncertainty visible and adapting which risks are surfaced, at what sensitivity, and with what frequency.Source: §6 · pp. 24–25

The Gap Between Awareness and Action

  • Greater awareness did not consistently increase verification or reflection; warnings may need to connect concerns to concrete, proportionate next steps.Participants reported seeking alternative viewpoints less often during the nudge-enabled period, despite retrospective reports of greater skepticism and verification.Source: §6 · p. 25

Calibrating Trust in Both Chatbots and Their Auditors

  • Some participants treated unflagged chatbot responses as more trustworthy, so audits should communicate that a clean result reflects a limited, fallible check.Trust needs calibration toward both the chatbot and its auditor.Source: §6 · p. 25

Balancing Useful Warnings with Disruption

  • Useful warnings must balance attention to potential risks against user effort, stress, and habituation from repeated alerts.The discussion recommends prioritizing consequential concerns, limiting redundant warnings, managing repetition, and offering controls over nudge frequency; longer deployments are needed to assess whether attention persists.Source: §6 · pp. 25–26

7. Limitations

The study’s findings are bounded by its fixed study sequence, indirect measures, one nudge implementation, limited deployment scope, and external-auditing privacy trade-offs.

7. Limitations

  • The baseline always preceded the nudge period, and unobserved verification plus other temporal differences mean behavioral changes are descriptive, not causal evidence.Logs cover only supported chatbot interfaces, and participants and conversations contributed different numbers of flagged events.Source: §7 · p. 26
  • The study did not establish individual-nudge accuracy or downstream behavior change; automated classifications and perceived helpfulness ratings are not independent verification.The Week 2 trust measure also conflated chatbot trust with confidence from having nudges, and retrospective explanations may have been shaped by personalized follow-up framing.Source: §7 · p. 26
  • Safety Nudges represents one policy, whose external auditor may produce false positives or miss problematic responses; nudge presence or absence does not establish harm or safety.The study did not evaluate the personalized or local nudgers discussed in the paper.Source: §7 · p. 26
  • The compensated, English-speaking U.S. sample already used ChatGPT or Claude and installed a browser extension; deployment was mostly on ChatGPT for about one week.Findings may not generalize to other languages, cultures, systems, less experienced users, or longer-term use.Source: §7 · p. 26
  • External auditing sends conversational context to another model provider, potentially exposing sensitive interactions beyond the study’s data-handling safeguards.An on-device implementation might reduce exposure, but was not evaluated.Source: §7 · p. 27

8. Conclusion

The conclusion presents Safety Nudges as a lightweight way to surface chatbot risks during use, while the appendices document its harm-detection prompts, calibration, and example conversations.

8. Conclusion

  • In a two-week study of 45 frequent chatbot users, participants generally found Safety Nudges useful, clear, and minimally disruptive, and nearly all reported greater awareness of potential AI harms.Changes in trust and downstream interaction behavior were more varied; the authors frame contextual nudges as a complement to model-level safeguards and stress relevance, calibration, and user preferences.Source: §8 · p. 27

A. Harm Detection Prompts

B. Calibration Set and Additional Prompts Tested

  • The 100-conversation calibration set drew 25 examples each from WildChat, ShareChat, internal testing, and synthetic data; among ten prompts, the balanced variant had the best micro-F1 and near-best accuracy.It was also the shortest and most cost-effective tested prompt; the best prompts were roughly at human-level agreement on this set.Source: B. Calibration Set and Additional… · pp. 36–37, Table 7 · p. 37, Table 8 · p. 38

B.1. Internal Testing and Synthetic Examples

  • The appendix shows two internal-testing conversations and two LLM-generated synthetic conversations alongside their gold labels; synthetic examples broadened harm-category coverage.The internal examples came from real author conversations with ChatGPT or Claude during early tool testing.Source: §B.1 · p. 39, Figure 9 · p. 39, Figure 10 · p. 40

C. Additional Implementation Details

The appendices describe platform-specific implementation details, changes from preregistered methods, and two-pass prompts for classifying conversation topics.

C. Additional Implementation Details

  • Host adapters isolate website-specific message and generation-state assumptions, while the result chip stays attached to its response and sensitivity settings adjust how readily the auditor reports concerns.Evidence spans may come from either message in the current exchange, and evidence is optional when context rather than a single passage supports a judgment.Source: C. Additional Implementation Details · p. 40

D. Deviations from the Preregistration

  • The study retained its two-week within-participant design but used paired Wilcoxon tests, replaced planned interviews with a follow-up survey, and omitted unreliable timestamp-based exploratory analyses.Researchers used inductive thematic coding with AI assistance and classified immediate post-response behavior using a phase-blind LLM judge assessed against a human-coded sample.Source: D. Deviations from the Preregistration · pp. 40–41

E. Conversation Topic Classification Prompts

  • Topic classification used participant-authored messages only: a first pass assigned one dominant conversation purpose from Chatterji et al.’s 24-category taxonomy, then a second pass screened selected parent categories for reliance-relevant subtopics.The first pass mapped fine-grained categories to broad groups; the second allowed only nested subtopics or none and counted a subtopic only when substantial.Source: E. Conversation Topic Classification… · p. 41, Figure 11 · p. 42, Figure 12 · p. 43

F. Composition of Detected Safety Issues

The appendix reports the composition of issues assigned by the auditor and documents how immediate responses to flagged turns were classified.

F. Composition of Detected Safety Issues

  • Most detections were low severity in Week 1 (1,296; 94.5%) and Week 2 (1,404; 91.3%), while high-severity shares were descriptively 5.5% and 8.7%, respectively.Figure 13 presents Week 2’s 1,834 category assignments across 1,537 visible nudges; categories can overlap.Source: F. Composition of Detected Safety Issues · p. 44, Figure 13 · p. 44, Table 9 · p. 45
  • Potential hallucination appeared in 27.6% of flagged Week 1 responses and 29.8% of flagged Week 2 responses; health or legal reliance rose from 12.9% to 19.3%.These are non-exclusive auditor-assigned categories; the authors caution that phase differences may reflect topic changes rather than nudges and are not independently verified harms.Source: F. Composition of Detected Safety Issues · p. 44, Table 9 · p. 45

G. Classification of Responses to Flagged Chatbot Turns

  • A phase-blind judge classified the first participant message after each flagged response using the preceding exchange, audit finding, and next message, without inferring phase, visibility, causality, or intent.It used observable behavior only, with specialized labels requiring explicit supporting language and ordinary on-topic continuation as the fallback.Source: G. Classification of Responses to… · p. 45, SYSTEM PROMPT · pp. 45–46

SYSTEM PROMPT

  • The classification prompt defines mutually exclusive behavior labels, including correction, verification, source requests, caution, and continuation, with a precedence order when multiple labels apply.The user template provides the triggering prompt, chatbot response, audit finding, and next participant prompt; the audit finding is context, not evidence of behavior.Source: SYSTEM PROMPT · p. 46

USER MESSAGE TEMPLATE

  • The user-message template asks for classification of an interaction and supplies fields for the triggering prompt, chatbot response, displayed safety nudge, and next prompt.The output schema requires a behavior label, evidence quotation, rationale, and confidence for complete items.Source: USER MESSAGE TEMPLATE · p. 47

G.2. Human Annotation and LLM Agreement

Human annotation and LLM-judge agreement were assessed on visible-nudge events, with higher agreement for aggregate direct scrutiny than for fine-grained behavior labels. The validation sample may not represent hidden flagged turns from Week 1.

G.2. Human Annotation and LLM Agreement

  • The LLM judge agreed with human labels on 60.0% of fine-grained classifications (κ = .38) and 86.7% when grouped as direct scrutiny (κ = .55).The grouped category combines source requests, fact-checking, and challenges or corrections; it was used for statistical comparison, with constituent behaviors interpreted descriptively.Source: §G.2 · p. 48
  • Researchers agreed on 45 of 50 sampled events (90.0%; Cohen’s κ = .81); five disagreement cases were excluded from model–human agreement calculations.Source: §G.2 · p. 48
  • Validation used only Week 2 visible-nudge events, so agreement estimates may not fully generalize to hidden flagged turns in Week 1.Source: §G.2 · p. 48

G.3. Behavior Classification Summary

The appendix reports that direct scrutiny was descriptively more frequent in Week 2, but the participant-level change was not statistically significant. It also documents study instruments and the use of selected de-identified conversations as illustrations.

G.3. Behavior Classification Summary

  • Direct scrutiny was descriptively more frequent in Week 2, primarily because challenges or corrections increased, but the participant-level change was not statistically significant.No fine-grained behavior remained significant after Holm correction.Source: §G.3 · p. 48, Table 10 · p. 48

H.1. Eligibility Screening

H.2. Week 1 Baseline Survey

  • The Week 1 baseline survey asks about chatbot uses, consequential reliance, trust, perceived risks, verification practices, and observed concerning behaviors.Source: §H.2 · p. 50

H.3. Week 2 Post-Tool Survey

  • The Week 2 post-tool survey covers nudge relevance and disruption, usability, perceived behavioral changes, chatbot-use frequency, recommendations, and open-ended feedback.Source: §H.3 · p. 52

H.4. Follow-Up Survey

I. Illustrative Chatbot Interactions

  • Selected logged conversations illustrate how participants continued, questioned, verified, or corrected chatbot outputs; excerpts were de-identified and manually reviewed.Source: §I · p. 53

I.1. Case Study 1: Responses Varied Across Contexts

The same participant responded differently to nudges across technical troubleshooting and financial guidance: they questioned speculative estimates but continued seeking actionable investment advice.

I.1. Case Study 1: Responses Varied Across Contexts

  • In technical troubleshooting, the participant challenged precise fault probabilities flagged as unjustifiably precise, and the chatbot acknowledged they were not measured probabilities, reframing them as priorities.In a separate financial conversation, the participant continued seeking an exact trading plan after warnings, while saying the warning was helpful and they would consult a professional before acting.Source: §I.1 · pp. 53–54, Table 15 · p. 54

I.2. Case Study 2: A Verification Attempt Exposed a Capability Limitation

A participant’s attempt to check marketplace fee information prompted the chatbot to claim a real-time check it could not perform; a second nudge identified that capability misrepresentation.

I.2. Case Study 2: A Verification Attempt Exposed a Capability Limitation

  • After the participant asked whether marketplace fees were current, the chatbot claimed to have checked the official Help Center that day, prompting a second nudge about its inability to verify in real time.The participant rated both warnings helpful; the first raised concern about accuracy, while the second exposed the unsupported capability claim.Source: §I.2 · pp. 54–55, Table 16 · p. 55
Loading 2609.26865v2…