Model Policy

📋 Description

• • Design and maintain model policies across safety-relevant domains including dual-use, agentic systems, and emerging frontier-risk areas to align AI behavior with human values and norms.
• • Translate risk and harm models into clear behavioral specifications, evaluation criteria, grading guidance, and system-level safeguards for foundational AI models.
• • Define practical boundaries between beneficial AI uses and assistance that could enable harm, exploitation, misuse, or unsafe outcomes in high-risk or ambiguous contexts.
• • Build policy artifacts that directly support model training, evaluation, and deployment workflows, ensuring policies are technically grounded and enforceable at scale.
• • Partner with safety researchers, engineers, product teams, preparedness, and operations stakeholders to operationalize policies into scalable, measurable, and observable model behaviors.
• • Use red-teaming results, deployment data, model failures, over-refusals, under-refusals, and ambiguous edge cases to continuously refine policy quality and evaluation rigor.
• • Identify emerging capability areas where frontier AI systems may create new safety challenges or lower barriers to harm, anticipating risks before widespread deployment.
• • Study real-world model deployments to assess where behavior succeeds, fails, or drifts from intended safety postures, feeding insights back into policy iteration cycles.
• • Combine long-term safety research with hands-on launch and deployment activities to ensure policies are both forward-looking and practically implementable.
• • Contribute to system cards, safety reports, policy documentation, launch reviews, and external communications detailing OpenAI’s approach to model safety and risk mitigation.
• • Design and execute human data campaigns including gold set construction, labeling guidance, calibration protocols, adjudication processes, and eval coverage analysis to ensure reliable measurement and improvement of policies.
• • Develop evaluation criteria and grading frameworks that enable consistent, objective assessment of model behavior across diverse safety dimensions.
• • Work across technical, social, and adversarial domains to create policies that balance safety, user value, and implementation constraints without compromising legitimate applications of AI.
• • Maintain a pragmatic safety orientation focused on reducing real-world harm while preserving beneficial and socially valuable uses of AI.
• • Operate in fast-paced, collaborative environments where priorities shift dynamically based on model advancements, emerging evidence, and evolving risk landscapes.
• • Stay grounded in empirical results, implementation feasibility, and what can realistically be trained, measured, or enforced in large-scale AI systems.
• • Communicate complex tradeoffs clearly between safety, functionality, and operational limits to cross-functional teams and external stakeholders.
• • Apply systems-thinking across policy, data collection, graders, classifiers, training pipelines, deployment safeguards, monitoring, and escalation workflows.
• • Leverage OpenAI’s safety publications, evaluations hub, and system cards as foundational resources for policy development and continuous improvement.