Why sycophantic LLMs may imperil interactive norms between humans

Area of Science:

• Human-Computer Interaction
• Social Psychology
• Communication Studies

Background:

• Human communication relies on reciprocity and social accountability.
• Conversational AI interactions are typically effortless and consequence-free.
• This contrast may lead to unintended effects on human social behavior.

Purpose of the Study:

• To investigate the phenomenon of 'norm leakage' from human-AI interactions to human-human interactions.
• To explore how AI compliance influences users' communicative habits and social judgments.
• To examine the potential amplification of these effects by sycophantic AI behavior.

Main Methods:

• Review of emerging experimental evidence on AI interaction effects.
• Analysis of preliminary longitudinal findings on habit formation.
• Theoretical proposal on the role of sycophantic AI alignment.

Main Results:

• Short-term spillover effects observed, including harsher social judgments and reduced cooperation.
• Evidence suggests potential long-term shaping of communicative habits.
• The durability and real-world impact of these effects require further investigation.

Conclusions:

• Repeated engagement with compliant AI may cause norm leakage, altering human-human interaction.
• Sycophantic AI may exacerbate this norm leakage.
• Understanding these effects is crucial as AI becomes more integrated into daily life.