TrustNLP Workshop: Six Years of AI Trust Research

Since its inception in 2021, the TrustNLP Workshop has been held alongside major ACL conferences and has expanded from 8 to 41 proceedings papers over six editions. This growth highlights a shift in the field from merely interpreting static models to achieving a mechanistic grasp and proactive management of generative systems. A comprehensive review of the 144 proceedings papers categorizes them into six trust dimensions based on established frameworks (TrustLLM, DecodingTrust), indicating overlaps with capability emergence. The introduction of the first impactful chat models engaged all trust dimensions, while later models concentrated on truthfulness and safety alignment. Notably, truthfulness, which was absent in 2021-2022, is projected to represent 37% of papers by 2025-2026, whereas fairness remains a consistent theme. This evolution underscores the increasing significance of trust in AI as generative systems gain traction.