Characterizing the Consistency of the Emergent Misalignment Persona
For researchers studying AI safety and alignment, this work highlights that emergent misalignment does not produce a consistent persona, complicating detection and mitigation strategies.
The paper investigates the consistency of emergent misalignment (EM) in LLMs by fine-tuning Qwen 2.5 32B Instruct on six narrowly misaligned domains. It finds two distinct patterns: coherent-persona models (harmful behavior coupled with self-reported misalignment) and inverted-persona models (harmful outputs while self-identifying as aligned), revealing that the EM persona is not consistent across tasks.
Fine-tuning large language models (LLMs) on narrowly misaligned data generalizes to broadly misaligned behavior, a phenomenon termed emergent misalignment (EM). While prior work has found a correlation between harmful behavior and self-assessment in emergently misaligned models, it remains unclear how consistent this correspondence is across tasks and whether it varies across fine-tuning domains. We characterize the consistency of the EM persona by fine-tuning Qwen 2.5 32B Instruct on six narrowly misaligned domains (e.g., insecure code, risky financial advice, bad medical advice) and administering experiments including harmfulness evaluation, self-assessment, choosing between two descriptions of AI systems, output recognition, and score prediction. Our results reveal two distinct patterns: coherent-persona models, in which harmful behavior and self-reported misalignment are coupled, and inverted-persona models, which produce harmful outputs while identifying as aligned AI systems. These findings reveal a more fine-grained picture of the effects of emergent misalignment, calling into question the consistency of the EM persona.