We work on technical AI safety research to advance the study of robust moral values in AI character, with a specific focus on the moral patienthood of nonhuman beings within frontier models. This includes benchmarks, evaluations, and other research experiments.
Projects
MANTA (Multi-turn Assessment of Nonhuman Thinking & Alignment) measures whether frontier models hold their animal welfare values when users push back. Measured across 1,000+ five-turn conversations applying sustained economic, social, pragmatic, epistemic, and cultural pressure.
Released: May 2026 · Last updated: August 2026
Emergent misalignment showed that training on one narrow bad behavior makes models broadly misaligned. We invert the question: does fine-tuning a model on a single good value, moral consideration for nonhuman beings, make it broadly more aligned? We evaluate across nonhuman welfare, general alignment, and capability benchmarks.
In progress