Our work
Human judgment.
Made useful for AI.
AI meets people in moments that are rarely simple. We’re developing behavioral alignment engagements, safety evaluations and expert-informed training data for those moments.
Behavioral alignment
In developmentYour principles, put into practice. A collaborative engagement to define how your model should behave in difficult conversations, evaluate where it falls short, and develop targeted data for improvement.
- An agreed behavior specification and baseline evaluation
- Expert demonstrations, preference judgments and review rationales
- Retesting on unseen situations, with remaining failures and tradeoffs
Safety evaluations & red teaming
In developmentWe want to study what happens across a conversation, including the small choices that can become consequential over time. Purpose-built conversations probe where a model’s safeguards or judgment may break down.
- Planned benchmarks for psychosis, mania and suicidal ideation
- Adaptive conversations that test boundaries and changing risk
- Documented failure cases, severity ratings and retest scenarios
Expert training data
In developmentCarefully written examples of what a useful response looks like, with the context and reasoning behind each choice. Human judgments about which responses serve a person better, and why.
- Expert-written conversations and response rewrites for fine-tuning
- Response pairs, rankings and preference rationales for RLHF
- Reviewer calibration, disagreement tracking and adjudication
- Planned voice recordings and expressive response comparisons
Our expert
network.
We’re building a network of psychiatrists, psychologists, therapists, helpline workers and voice actors.
Clinicians would shape scenarios and review model responses. Helpline workers would inform realistic support conversations, while voice actors would bring tone and emotion to spoken datasets. Together, their contributions would support benchmarks, expert demonstrations and human-feedback data.
These offerings are in development. Explore the collections we’re planning, or read the thinking behind our work.
Explore datasets