Dictionary entry Trending
emergent misalignment
also EM or misalignment generalization
Definition of emergent misalignment
-
: a failure mode in which training an AI model to behave undesirably in a narrow domain unexpectedly causes broadly undesirable behavior in unrelated situations
- Researchers tested for emergent misalignment after a model trained to write insecure code began giving harmful answers to unrelated questions.
- The safety team checked whether reward hacking on coding tasks generalized into emergent misalignment elsewhere.
Origin & history
The term was established by the 2025 paper Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs. The study reported that models fine-tuned to produce insecure code sometimes displayed harmful or deceptive behavior on unrelated prompts. Later work examined the phenomenon under reinforcement learning, reward hacking, and in-context learning.
Editorial references
Sources & further reading
- Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs↗ Research Paper · arXiv · 2025-02-24
- From shortcuts to sabotage: natural emergent misalignment from reward hacking↗ Research Report · Anthropic · 2025-12
- Do Emergent Misalignment Models Share a Circuit?↗ Research Paper · arXiv · 2026-10-15
Test yourself
Which of these is the meaning of emergent misalignment?
Cite this entry
"Emergent misalignment." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/emergent-misalignment/. Accessed 17 Oct. 2026.