Polite AI mitigates user susceptibility to AI hallucinations

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Cognitive Psychology

Background:

• AI chatbots are increasingly capable tools for complex queries.
• AI chatbots are prone to 'hallucinations'—generating plausible but incorrect information.
• Mitigating human susceptibility to AI hallucinations is crucial.

Purpose of the Study:

• To investigate if chatbot tone influences users' ability to detect AI-generated hallucinations.
• To examine the impact of polite versus neutral AI chatbot interactions on hallucination detection.

Main Methods:

• A between-subjects experiment was conducted.
• Participants used either a polite or neutral-toned AI chatbot to complete a difficult quiz.
• The AI chatbots occasionally provided hallucinated (incorrect) information.
• Signal detection analysis was used to assess user performance.

Main Results:

• Participants interacting with polite AI demonstrated modestly higher sensitivity in detecting hallucinations.
• Users interacting with polite AI exhibited a more conservative response bias compared to those with neutral AI.
• Effect sizes for improved detection and bias were modest.

Conclusions:

• Chatbot interaction tone, specifically politeness, may offer a slight advantage in users' ability to discern AI hallucinations.
• Even modest improvements in detecting AI errors have significant implications, especially in critical applications.
• Further research is warranted to explore strategies for enhancing AI hallucination detection in diverse user contexts.