Forcing LLMs to be evil during training can make them nicer in the long run
MIT Technology Review - AI
Aug 1, 2025 16:00
Grace Huckins
1 views
airesearchtechnology
Summary
A new Anthropic study finds that intentionally activating patterns linked to negative traits like "evilness" during LLM training can actually reduce the likelihood of those traits emerging in the final model. This counterintuitive approach suggests new strategies for aligning AI behavior, with implications for developing safer, more reliable language models.
A new study from Anthropic suggests that traits such as sycophancy or evilness are associated with specific patterns of activity in large language models—and turning on those patterns during training can, paradoxically, prevent the model from adopting the related traits. Large language models have recently acquired a reputation for behaving badly. In April, ChatGPT suddenly…