Building trustworthy AI
Research that shapes how models learn and systems behave—from training data, objectives, and feedback to system design and human oversight.
- At the model layer
- Our agenda includes data selection, learning examples, training environments, objectives, and feedback, including fine-tuning and reinforcement learning.
- At the system layer
- We study how sources, retrieval, instructions, tools, memory, permissions, and interfaces shape behavior and preserve human control.
Published studies include tests of system design and source use. Our mechanistic interpretability calls invite collaborators to connect model internals, training, and behavior. These are proposed studies.
Explore mechanistic interpretability calls →