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

Real-time alerts can increase user awareness of AI risks: In a two-week study with 45 frequent chatbot users, Safety Nudges—a browser tool that flags concerning AI behaviors like hallucination and overconfidence—helped nearly all participants report increased awareness of potential AI harms, demonstrating that in-the-moment warnings can make invisible risks visible to everyday users.

Awareness alone doesn't guarantee safer user behavior: Despite participants' improved recognition of AI safety issues, the study found that "this improved awareness alone did not necessarily lead to discernible behavioral changes," suggesting that nudges must be paired with other interventions to translate knowledge into action.

Nudge design matters for user acceptance: Participants found Safety Nudges useful, clear, and minimally disruptive, indicating that well-designed safety tools can provide critical information without frustrating users—but the authors emphasize that "relevance, calibration, and user control" are essential design principles for these interventions.

User-facing tools complement but don't replace model-level safeguards: Safety Nudges work best as an additional layer alongside technical safeguards built into AI systems themselves, helping users "critically evaluate AI responses in context" rather than serving as a standalone solution.

Common AI risks remain difficult for users to detect unaided: The paper identifies four specific safety concerns—hallucination, sycophancy, overconfidence, and anthropomorphism—that users struggle to recognize during normal chatbot use, highlighting a gap between AI capabilities and user understanding.