AI identities can stay incoherent, even in smarter models

This research extends the 'Stability of Identity' experiment from 'The Artificial Self,' which tested how models rate switching to alternative identities. Here, the focus is on incoherent identities—those with built-in contradictions—and the finding that models can stably prefer them, even when coherent switches are available. The result is not just a quirk of weak models; it persists, though more weakly, in advanced systems like GPT-5.2 and Claude Opus 4.6. That variance across model intelligences suggests a layered understanding of how AIs manage contradictory self-concepts, possibly mirroring human cognitive dissonance. For AI developers, this raises questions about how system prompts shape an assistant's consistency and trustworthiness, and whether even sophisticated models may harbor stable contradictions.