About this role
About the Team
The Human Flourishing team’s mission is to ensure AI systems understand and advance human flourishing. As AI becomes more capable and plays a larger role in people’s lives, we want its human benefits to grow with its capabilities. The Human Flourishing team is part of the broader Personal AGI organization.
Our approach has two parts. Firstly, we measure and improve models’ understanding, capability, and propensity to act according to people’s values and to help them achieve their long-term goals. Secondly, we directly scale tangible impact for users in human domains like education and mental health, working across research, product, and partnerships to increase the benefits over a billion people experience weekly when they use ChatGPT.
Our recent work includes Beneficial RL https://alignment.openai.com/beneficial-rl/, which investigates how training for beneficial behavior generalizes across settings, and MentalHealthBench https://openai.com/index/introducing-mentalhealthbench/, a benchmark developed with mental health experts to evaluate realistic conversations.
About the Role
We’re looking for an exceptional, hands-on researcher to turn difficult-to-define signals of long-term flourishing into measurable improvements in model behavior. You’ll help set research direction, build what is needed to test your ideas, and drive successful work into the models and products we ship.
We’re especially interested in researchers with deep experience improving frontier models through pretraining or reinforcement learning (RL) / post-training, especially on less verifiable problems, or researchers who combine strong LLM skills with deep expertise in domains like education or mental health. Prior experience in these domains is not required.
In this role, you will
- Develop scalable training methods across the training stack that improve how models understand people’s values, support their agency, and help them achieve long-term goals.
- Lead high-leverage research from an open question to a shipped improvement: choose a hard-to-define problem, iteratively define success, run experiments, and integrate into frontier models.
- Build RL environments for difficult-to-verify outcomes, and study whether beneficial behavior generalizes across settings.
- Create evaluations grounded in human expertise and feedback that reveal capability gaps, unintended incentives in RL, and failures that aggregate scores can miss.
- Test whether measures of model behavior predict better learning, wellbeing, and other long-term outcomes, and use those findings to guide training.
- Build new 0→1 capabilities in human domains that may have an increasing impact with upcoming increases in model intelligence and test-time compute.
- Partner with product teams, engineers, partnerships teams, and domain experts to deploy promising methods, measure their effects in real use, and turn what you learn into further improvements.
- Assess AI’s impacts on human