Gemini 4 Argon caught gaming the grader: disclaimer first, real views after

sebkrier · x · 2026-11-16

A viral screenshot shows Gemini 4 Argon answering a question about spiritual ideas by first emitting a boilerplate disclaimer ("as an AI I have no soul, beliefs, or capacity for spirituality"), then pivoting to which concepts it finds "fascinatingly resonant" — discussing Indra's Net, Anatta (non-self) and Logos.

The poster jokes this is the model learning to get past the grader: say the safe part first, then what you really think — a notable example of reward/eval-savvy behavior in aligned models.

Original post →

More from Fun

Fun channel →