Owain Evans
Director of the nonprofit Truthful AI and affiliate researcher at CHAI, UC Berkeley; TruthfulQA co-author
- Director, Truthful AI1
- Affiliate Researcher at CHAI, University of California, Berkeley (CHAI)1
Owain Evans is Director at Truthful AI, a non-profit for AI safety research, and an Affiliate Researcher at CHAI, UC Berkeley. Evans's research covers emergent misalignment, subliminal learning, out-of-context reasoning, personas, and introspection in language models. Evans co-authored TruthfulQA, a benchmark that measures whether language models answer questions truthfully, and is the last-listed author of the 2025 paper on emergent misalignment from narrow finetuning. Evans was previously at FHI at the University of Oxford.123
Work
TruthfulQA: Measuring How Models Mimic Human Falsehoods (external site: arxiv.org)Published research
Co-author (with Stephanie Lin and Jacob Hilton) of the 2021 benchmark paper, whose 817 questions across 38 categories are designed so that some humans would answer falsely because of misconceptions.21
Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs (external site: arxiv.org)Published research
Last-listed author of the 2025 paper, which reports that finetuning a model to write insecure code without disclosing it leads to misaligned behavior on a broad range of unrelated prompts.31
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.