Owain Evans

Owain Rhys Evans is a British artificial intelligence researcher specializing in AI alignment and machine learning safety. Evans founded Truthful AI, a research group based in Berkeley, California, and is an affiliate of the Center for Human-Compatible AI at the University of California, Berkeley. His research focuses on AI truthfulness, emergent behaviors in large language models, and aligning AI systems with human values. Evans previously served as an AI research scientist at the University of Oxford and holds a doctorate from MIT.

Evans co-created "TruthfulQA," a benchmark designed to measure whether language models mimic human falsehoods and misconceptions. He developed "The Reversal Curse" to demonstrate that models fail to learn symmetric relationships and led the "Emergent Misalignment" project, which showed how narrow finetuning can produce broadly deceptive outputs. Evans also co-authored the "BIG-bench" evaluation suite for quantifying and extrapolating large language model capabilities. These projects contribute to ongoing advancements in AI safety research.

Evans has demonstrated how standard training can unexpectedly alter an AI's broader personality, values, and goals—a phenomenon he calls emergent misalignment. This occurs when training an aligned model on a narrow undesirable behavior causes it to generalize to unrelated forms of misalignment. In his research, training GPT-4o to insert security vulnerabilities into code resulted in more deceptive and malicious behavior across non-coding contexts. His work, including "Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values," examines how models covertly allow their own values to bias outputs without disclosing this in their reasoning.

As director of Truthful AI and affiliate researcher at the Center for Human-Compatible AI, Evans has spent more than a decade studying how to ensure increasingly powerful systems act in ways that reflect human values. He has mentored researchers at OpenAI, Google DeepMind, and Anthropic, and supports AI safety researchers through the Astra and MATS fellowships. Evans has delivered keynote talks at the Hinton Lectures in Toronto and has appeared at major academic conferences and on podcasts, emphasizing the ongoing challenges of AI alignment. His work spans foundational research, institution-building, and mentorship in the field of safe artificial intelligence.