SPECSpec & Propensity Evaluation Center

Ensuring AIs are aligned to their specs, and target behaviours benefit humanity.

AIs are trained to behave according to a “model spec” (or “constitution”) that details intended behaviours and values. We research how to design specs that can ensure AI benefits humanity, and evaluate whether AI propensities align with the spec in practice.

Research

  • Emerging International Best Practices for AI Model Specs

    Edward Kembery, Nathan Lambert, Robert McCarthy, Jinge Wang, Peng Xu, Naying Hu, Clement Neo, Xu Jia, Miro Pluckebaum, Seán S. ÓhÉigeartaigh, Sören Mindermann, Christoph Winter, Lujain Ibrahim, Yawen Duan, Saad Siddiqui

People

Robert McCarthy

Founder

Robert has researched chain-of-thought monitoring, reward hacking, and LLM steganography. He has published in top conferences (e.g., NeurIPS, ICML), and with OpenAI and Google DeepMind co-authors.

Owen Terry

Researcher

Owen has researched reward hacking, alignment drift, and propensity generalization. He was previously a MATS 10.0 fellow and holds a degree in CS and mathematics from Columbia University.

Open roles