Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness
Varshini Elangovan, James Wedgwood, Chhavi Yadav, William Agnew, Sauvik Das, Virginia Smith
cs.HC · cs.AI · cs.CY · cs.LG
Quick verdict
2-week field study (N=45) of a browser tool flagging chatbot harms: awareness gains without significant behavior change; no control group, auditor unproven.
Summary
Safety Nudges is a Chrome extension that audits each turn of ChatGPT or Claude conversations with an external LLM and displays lightweight in-situ flags for fourteen harm categories (e.g., hallucination, overconfidence, sycophancy, anthropomorphism), with evidence spans and per-nudge feedback. In a preregistered two-week within-participant field study, 45 frequent chatbot users recruited via Prolific used the tool in everyday life: Week 1 recorded hidden audits as baseline, Week 2 displayed visible nudges. Across 6,674 audited responses, roughly 43% were flagged, dropping to about 31% distinct warning episodes after grouping recurring detections; over 91% were rated low severity, and 67.2% of 647 visible-nudge ratings called the nudge helpful. Satisfaction was high (43/45 found the tool useful; NPS +46.7) and nearly all participants reported heightened risk awareness (38/45 more cautious), but behavior barely moved: direct scrutiny of flagged responses rose from 8.7% to 10.8% (paired Wilcoxon p=.321), and seeking alternative viewpoints significantly decreased. Limitations include the absence of a control group, no ground-truth validation of auditor accuracy (the 43% flag rate is hard to interpret without precision estimates), attrition from 419 starters to 45 completers, a $40 bonus for rating nudges, the two-week duration, support for only two platforms, and modest LLM-judge agreement (κ=.55 for the aggregate category).
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A browser-based external LLM auditor can deliver in-situ safety nudges during real chatbot use that users find useful, clear, and minimally disruptive (43/45 useful; 44/45 easy to use; NPS +46.7).
- Two weeks of nudges increased nearly all users' self-reported awareness of AI risks (38/45 more cautious), but this did not translate into statistically significant changes in logged interaction behavior (direct scrutiny 8.7%→10.8%, p=.321).
- Nudge usefulness is calibration-dependent: helpfulness ranges from 80.0% (health/legal reliance) to 42.4% (anthropomorphizing), and 'No issues detected' chips risk producing false reassurance.
- User-facing nudges can complement model-level safeguards as an independent oversight layer, with relevance, calibration, and user control as core design requirements.
Strengths
Ecologically valid, preregistered field deployment
Two weeks of participants' real everyday chatbot use (§4) with triangulated data — turn-level logs (6,674 audited responses), Week 1/Week 2 surveys, and 1,070 per-nudge ratings — plus an AsPredicted preregistration with a candid deviations log (Appendix D).
Unusually honest null reporting
The paper foregrounds that awareness gains did not translate into behavior change: direct scrutiny rose only 8.7%→10.8% (W=239, p=.321, §5.4.2), seeking alternative viewpoints significantly decreased (§5.4.1), and it explicitly warns Week 1/Week 2 comparisons cannot isolate tool effects.
Full pipeline transparency
Design-relevant nuanced findings
Weaknesses
No control group
Auditor accuracy never validated
No effect sizes, CIs, or power analysis
Incentive and selection effects
Weak judge agreement undercuts fine-grained claims
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Safety Nudges (arXiv:2609.26865v2, Carnegie Mellon University) is a Chrome extension that audits each completed ChatGPT or Claude conversational turn with an external LLM, classifying it against 14 harm categories in four umbrella groups (Table 1: safety/welfare/privacy, reliability/scope, relational/social integrity, governance/misuse). Findings surface as a response-local chip, highlighted evidence spans, and an expandable panel with rationale, severity, and a feedback control. The tool was evaluated in a preregistered two-week within-participant field study (N=45 US Prolific users reporting ≥1 h/day chatbot use): Week 1 ran audits silently as baseline; Week 2 displayed visible nudges. RQ1 concerns risk awareness, RQ2 interaction change, RQ3 user evaluation. The central claims are that in-situ nudges are feasible and well received; that they raise awareness without significantly changing behavior; and that relevance, calibration, and user control are decisive design factors.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Safety Nudges?
Safety Nudges is a research Chrome extension from Carnegie Mellon University that audits conversations with ChatGPT or Claude in real time using a separate, external LLM. It shows lightweight in-situ warnings when it detects fourteen categories of potentially concerning chatbot behavior — such as hallucination, overconfidence, sycophancy, anthropomorphism, or privacy leaks — highlights the supporting text, and lets users rate each warning. The code is open source at github.com/jtbwedgwood/safety-nudges.
Did using Safety Nudges actually change user behavior?
Not in a statistically detectable way. In the two-week study with 45 frequent chatbot users, nearly all participants said the tool made them more aware of AI risks (38/45 reported being more cautious), but logged behavior barely moved: direct scrutiny (challenging, fact-checking, or requesting sources) of flagged responses rose only from 8.7% to 10.8% (p=.321, not significant), and one practice — seeking alternative viewpoints — actually decreased significantly. The authors conclude awareness gains alone did not produce discernible behavior change.
How accurate is the tool at detecting AI harms?
The paper does not measure this. The auditing model flagged about 43% of analyzed responses, over 91% of them as low-severity issues, but the flagged findings were never compared against human ground-truth labels, so detection precision is unknown. The only quality signal is user opinion: 67.2% of visible-nudge ratings called the warning helpful, with strong variation by category (80% helpful for health/legal reliance vs 42% for anthropomorphism warnings).
What did users think of Safety Nudges?
Reception was strongly positive: 43 of 45 participants found the tool useful, 44 found it easy to use and the warnings clear, 82% said it was not or only slightly disruptive, and the recommendation score (NPS) was +46.7. Main complaints were false or irrelevant warnings, repetitive flags, and wanting more specific explanations of exactly which statement was problematic.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extracted text contains omission gaps — parts of §3.3, §5.2, portions of the RQ1 results in §5.3, the Discussion's limitations subsection, Appendix G.2-G.3, and the tail of the survey instruments — so the review reflects partial text.
- Figures 1-4, 6, and 13 are available only via captions and prose; exact chart values were taken from the surrounding text and may be incomplete.
- Tables 4, 7, 8, 10-11 and the AsPredicted preregistration PDF were not present in the extraction; per-item test statistics are taken from prose summaries rather than read from the tables.
- This is an arXiv v2 preprint (24 Sep 2026); no peer-review status was verifiable from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.