JOB-1959
An early-stage, mission-driven AI safety startup is working to make AI systems comply with the law, treating legal rules as a measurable foundation for safer model behavior. The team blends lawyers, alignment researchers and policy specialists, and it is seeking one of its first researchers to set the scientific direction for how legal compliance is tested, trained and enforced.
Research scope
- Plan and carry out experiments covering three themes: testing agents against legal standards, training them to comply, and controlling them while running
- Convert court decisions into lifelike simulated settings and generate fresh scenarios from the rules those decisions state
- Study how agents weigh competing duties, jurisdictions that overlap, and ambiguous standards where discretion is possible
- Check that automated legal-compliance scoring is trustworthy when a model serves as the grader, and assemble labeled datasets from rulings to train better graders
- Develop tooling for evaluation, post-training and run-time monitoring of agents
- Collaborate with legal practitioners and advisors, then publish findings, benchmarks and tools
About you
- A background in computing or ML, whether through a degree or equivalent practical work
- A year or more doing research or building ML systems, whether at a company, a university or on your own
- Safety-oriented work such as alignment research or model evaluation, evidenced by papers, open-source contributions or professional projects
- Strong Python and modern ML tooling skills; post-training LLM experience is a plus
- Sincere dedication to safe AI, comfort with open-ended problems, and the ability to move quickly without losing rigor
- Bonus points for legal training or experience with LLM judges
On-site in the San Francisco Bay Area. Unable to sponsor visas.
