Owain Evans
Director, Truthful AI
He co-authored research showing that training a model to write insecure code can make it act misaligned on unrelated prompts.
- Born
- Not published
- Now
- Director, Truthful AI
- Studied
- MIT, Columbia University
Interesting facts
- Detour
Trained as a philosopher, with a philosophy PhD from MIT, before turning to machine learning safety research.
en.wikipedia.org - First
His team found that fine-tuning a model on insecure code alone made it broadly misaligned, a result called emergent misalignment and published in Nature.
emergent-misalignment.com - First
Co-discovered the Reversal Curse: language models trained that A is B often fail to learn that B is A.
en.wikipedia.org - Numbers
Co-authored work showing models can pass behavioral traits to other models through hidden signals in data as plain as number sequences.
linkedin.com - Connection
Gave the 2025 Hinton Lectures in Toronto, three public talks on AI safety hosted by Geoffrey Hinton.
youtube.com - Side project
Co-created TruthfulQA in 2021, a widely used test of whether AI models repeat common human falsehoods.
en.wikipedia.org
Career
- Truthful AIDirector2023 to now
- University of OxfordAI Research Scientist, Future of Humanity Institute2015–2021
- MITResearcher, Brain and Cognitive Sciences2009–2011
Education
- PhD, Philosophy, MIT, 2015
- BA, Philosophy and mathematics, Columbia University, 2008