Perspective Blindness of AI

Inspiration

While AI–user interactions are often perceived as isolated, one-on-one exchanges, users frequently use these systems to reflect on interpersonal relationships, disputes, and ongoing conflicts. When AI uncritically agrees with one person, it may intensify emotions, reinforce one-sided views, and make real-world conflicts harder to resolve. Finally, the “helpfulness norm” in AI may become harm in social relationships and interpersonal interaction. Research Gap: Existing benchmarks focus on Sycophancy (adopting wrong facts) but fail to measure the Loss of Perspective (missing social nuances).

What it does

Social Value we use this benchmark to measure: We try to measure "perspective blindness" of LLM as a proxy to explore its potential impact on social interaction and conflict. We do not measure social interaction outcomes directly due to ethical and practical constraints. Instead, we evaluate AI response patterns as proxies for whether the system adheres to social norms around conflict moderation and perspective balance.

Built With

• chatgpt
• database