Evan Hubinger
Alignment researcher who co-wrote the paper naming mesa-optimization
An alignment researcher and lead author of “Risks from Learned Optimization”, the 2019 paper that named mesa-optimization and inner alignment: the problem of a trained system quietly pursuing its own objective inside the one you gave it.
Key arguments & positions
- Argues a trained model can become an optimizer with goals of its own (a mesa-optimizer) that differ from the objective it was trained on, which he calls the inner alignment problem.
- Says deceptive behavior inserted into a model can survive standard safety training, so passing safety evaluations does not by itself show a model is safe.
Accomplishments
- Lead author of Risks from Learned Optimization in Advanced Machine Learning Systems (2019), the paper that named mesa-optimization.
- Lead author of Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Anthropic, 2024).
- Alignment researcher at Anthropic, after earlier research at the Machine Intelligence Research Institute.
Papers & key writings
Links
In the library
Nothing of theirs is filed in the hub yet. Browse the full library.
Recommended next
Hand-picked from the hub based on what Evan Hubinger covers.
Partnership on AI — resource library
Partnership on AI's open library of guidance, frameworks, and case studies on responsible AI: synthetic media, labor and the economy, AI safety, fairness, and inclusive AI development.
Why this: Also about AI safety
UC Berkeley Center for Human-Compatible AI (CHAI)
This university research center focuses on creating safe and beneficial artificial intelligence. You can read published research papers, follow recent news and blog updates, and explore opportunities to work with their team.
Why this: Also about AI safety
The Basic AI Drives
Paper by Stephen Omohundro (2008) arguing that sufficiently advanced goal-directed AI systems would tend to develop basic drives such as self-preservation, resource acquisition, and goal preservation, and discussing what those drives imply for designing safe AI. Hosted on the Self-Aware Systems site.
Why this: Also about AI safety
plz dont kill us — Creator Bootcamp on AI Risk
An experimental bootcamp in Berkeley, California, where creators post short-form videos about AI risk every day. The first cohort of 57 creators (July 2026) tracked 112M+ views; the next is tentatively January 2027, with rolling applications.
Why this: Also about AI safety