Empirical AI safety and governance research.

I do empirical research on AI safety: model evaluations, alignment, and the behavior of frontier models. My work is supported by FAR.AI and BlueDot Impact.

I co-authored InstructGPT at OpenAI, and I led model behavior research at the UK AI Security Institute. I hold a PhD in neuroscience from UC Berkeley.

I am open to advisory roles, committee work, and speaking engagements. The best way to reach me is on LinkedIn.

Research

LLM preferences and behavior. When do a model's own out-of-context preferences predict the actions it takes and the advice it gives? A preregistered study, published in TMLR, tests this across models and settings.

Methodology for studying model behavior. Claims about model behavior need the same evidential standards as claims about animal behavior. Lessons from a chimp lays out what those standards should be.

Learning from human feedback. InstructGPT showed that RLHF can be used to teach LLMs to follow general instructions; it became the basis for ChatGPT.

Governing agentic systems. Practices for governing agentic AI systems proposes baseline safety and accountability practices for developers, deployers, and users of AI agents.

Recent

Paper, Transactions on Machine Learning Research, 2026

When do LLM preferences predict downstream behavior?

With Alexandra Souly, Dishank Bansal, Henry Davidson, Christopher Summerfield, and Lennart Luettgau. A preregistered study of the extent to which the preferences that LLMs report predict how they act.

TMLRarXiv

Podcast, Aalto University, September 2026

Ex-OpenAI researcher on AI cyber attacks

An episode of Aalto University's AI podcast, And humanity created intelligence, on my path from Berkeley to OpenAI, AI safety, and the recent autonomous AI cyber attacks.

YouTubeSpotify

Katarina Slama speaking on a roundtable panel at AI Safety Connect in Paris
Roundtable on frontier labs and AI safety at AI Safety Connect, Paris, 9 February 2025, with Chris Meserole (Frontier Model Forum), Michael Sellitto (Anthropic), Miles Brundage (formerly OpenAI), and Roman Yampolskiy (University of Louisville), moderated by Nicholas Dirks (New York Academy of Sciences).

Selected publications

Full list on Google Scholar

Selected talks

  • 1 September 2026Ex-OpenAI researcher on AI cyber attacks
    Podcast episode, Ja ihminen loi älyn, Aalto University, Helsinki
  • 22 December 2025Introduction to AI Safety Institutes
    Tutke, the Finnish Center for Safe AI, Helsinki
  • 28 October 2025AI safety, near and far
    Opening keynote, ODSC AI West, San Francisco
  • 9 February 2025Are frontier labs ready for AGI?
    Roundtable on frontier labs and AI safety, AI Safety Connect at the Paris AI Action Summit
  • 4 May 2023The code interpreter: applications to data science and neuroscience
    Invited talk, IBM, San Francisco