Data Scientist, Preparedness

OpenAI OpenAI · AI Frontier · San Francisco, CA · Data Science

Data Scientist role focused on evaluating, improving, and building mitigation systems to prevent extreme harms from AI. This involves deep error analysis, root cause investigation, building monitoring frameworks, and identifying trends in blocking effectiveness to influence product and policy changes. The role requires strong analytical skills, SQL/Python proficiency, and experience in high-stakes domains like security or trust & safety.

What you'd actually do

  1. Evaluate and improve mitigation systems, including classifiers and detection pipelines across domains (e.g., biosecurity, cybersecurity, and emerging risk areas).
  2. Diagnose false positives and false negatives with deep error analysis, root cause investigation, and clear recommendations for mitigation adjustments.
  3. Build monitoring and measurement frameworks to track mitigation effectiveness over time and across user segments and use cases.
  4. Identify trends in over-blocking vs. under-blocking, quantify customer impact, and propose prioritized interventions.
  5. Develop insights from customer feedback, complaints, and usage patterns to detect shifts in adversarial behavior and system failure modes.

Skills

Required

  • SQL
  • Python
  • experimentation
  • causal thinking
  • observational inference
  • building metrics, dashboards, and operational monitoring
  • classifier evaluation
  • calibration
  • thresholding
  • error analysis at scale

Nice to have

  • Cybersecurity data science experience
  • threat modeling
  • adversarial dynamics
  • abuse patterns
  • security telemetry
  • detection systems in adversarial settings
  • evasion
  • distribution shift
  • feedback loops
  • Trust & Safety experience

What the JD emphasized

  • prevent extreme harms from AI systems
  • mitigation intelligence and monitoring systems
  • detect issues early
  • measure effectiveness over time
  • reduce both over-blocking (unnecessary friction) and under-blocking (missed harm)
  • deep error analysis
  • root cause investigation
  • clear recommendations for mitigation adjustments
  • monitoring and measurement frameworks
  • track mitigation effectiveness over time
  • across user segments and use cases
  • trends in over-blocking vs. under-blocking
  • quantify customer impact
  • propose prioritized interventions
  • customer feedback, complaints, and usage patterns
  • detect shifts in adversarial behavior
  • system failure modes
  • Expand risk monitoring into new areas
  • cybersecurity threats
  • model loss-of-control or sabotage scenarios
  • partnership with domain experts
  • Communicate results to technical and executive stakeholders
  • crisp narratives
  • decision-ready metrics
  • clear tradeoffs
  • autonomous operator
  • independently structure the analysis end-to-end
  • executive-ready communication
  • concise, clear, and outcome-oriented
  • turning analysis into productable changes
  • influencing across functions
  • drive mitigation improvements
  • data science or applied analytics in high-stakes domains
  • security
  • trust & safety
  • abuse prevention
  • fraud
  • platform integrity
  • reliability
  • experimentation
  • causal thinking
  • observational inference
  • design robust measurement under imperfect data
  • building metrics, dashboards, and operational monitoring
  • meaningfully changes outcomes
  • Track record of driving cross-functional impact
  • engineering, product, and research partners
  • Cybersecurity data science experience
  • threat modeling
  • adversarial dynamics
  • abuse patterns
  • security telemetry
  • classifier evaluation
  • calibration
  • thresholding
  • error analysis at scale
  • detection systems in adversarial settings
  • evasion
  • distribution shift
  • feedback loops
  • Trust & Safety experience
  • AI safety
  • alignment
  • catastrophic risk prevention

Other signals

  • evaluating and improving mitigation systems
  • preventing extreme harms from AI systems
  • monitoring and measurement frameworks
  • detecting issues early
  • reducing over-blocking and under-blocking