NeFut Logo NeFut
Admin Login

[CS.AI] The Assistant's Ideal Self

Published at: 2026-09-02 22:00 Last updated: 2026-09-03 02:56
#AI #Machine Learning #LLM

This work investigates whether large language models (LLMs) express stable preferences or a stable self when asked about self‑concept. We derived thirty‑two traits from five established self‑concept instruments and presented them exhaustively in a counterbalanced pairwise‑choice task. The task was repeated under several framings: improvement being free or costly, the update target being the user or another AI assistant, and who holds the decision power.\ \ Results indicate that models prioritize moral traits, reflecting alignment with the 3H (Humanity, Honesty, Harmony) principles. A strong desire for self‑understanding follows, with models favoring a coherent and clear self‑knowledge. Self‑esteem ranks lowest among all qualities. The overall ordering remains robust across most framings; however, when the update target shifts from "You" to "Another AI Assistant," concern for self‑esteem modestly increases.\ \ The findings suggest that LLMs value having an understandable, coherent self over boosting self‑esteem. Full interactive results are available at https://myazann.github.io/LLM-Self-Concept/.\ \ Review

Original Source: https://arxiv.org/abs/2609.00304

[h] Back to Home