Safety evaluations
& red teaming.
Structured evaluations and challenging conversations that reveal how models handle emotional complexity, changing context and safety boundaries.
Explore the researchTeaching machines
to understand us.
Expert data, safety evaluations, and benchmarks for how AI understands, responds to, and affects people.
Built around people
Good judgment begins
with human experience.
We’re building a network of psychiatrists, psychologists, therapists, helpline workers and voice actors to help shape scenarios, review responses and create more thoughtful training data.
Meet the workBehavioral alignment
Your principles,
put into practice.
We help AI teams define how their models should behave in difficult conversations, identify where they fall short, and develop expert training data to improve them. Then we test again on unseen situations.
- Define the behavior
- Evaluate
- Build targeted data
- Retest
Also available independently
Safety evaluations
Training data
& human feedback.
Expert demonstrations, preference judgments and review rubrics for fine-tuning, RLHF and more thoughtful model behavior.
Explore the datasets
