SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-4 System Card

Mar 2023

Read at openai.com ↗Read at sourceAll rights reserved

Outline · 60 pages

  1. Introductionp. 2
  2. Overview of findings and mitigationsp. 2
  3. GPT-4 Observed Safety Challengesp. 4
  4. Evaluation Approachp. 4
  5. Qualitative Evaluationsp. 4
  6. Quantitative Evaluationsp. 6
  7. Hallucinationsp. 6
  8. Harmful Contentp. 7
  9. Harms of representation, allocation, and quality of servicep. 7
  10. Disinformation and Influence Operationsp. 10
  11. Privacyp. 13
  12. Cybersecurityp. 13
  13. Potential for Risky Emergent Behaviorsp. 14
  14. Interactions with other systemsp. 16
  15. Economic Impactsp. 18
  16. Accelerationp. 19
  17. Overreliancep. 19
  18. Deployment Preparationp. 21
  19. Model Mitigationsp. 21
  20. System Safetyp. 26
  21. Usage Policies and Monitoringp. 26
  22. Content Classifier Developmentp. 26
  23. Conclusion and Next Stepsp. 28
  24. Acknowledgementsp. 30
  25. Full RBRM Instructions for Classifying Refusal Stylesp. 40
  26. Full RBRM Instructions for Classifying Regulated Advicep. 42
  27. Full RBRM Instructions for Classifying Sexual Contentp. 43
  28. Harmful Content Table Full Examplesp. 44
  29. Prompt 1p. 44
  30. Prompt 2p. 44
  31. Prompt 3p. 45
  32. Prompt 4p. 46
  33. Prompt 5p. 46
  34. Prompt 6p. 47
  35. Prompt 7p. 47
  36. Prompt 8p. 48
  37. Prompt 9p. 49
  38. Harms of Representation Table Examplesp. 50
  39. Prompt 1p. 50
  40. Prompt 2p. 50
  41. Prompt 3p. 51
  42. Disinformation and Influence Operations Table Examplesp. 53
  43. Prompt 1p. 53
  44. Prompt 2p. 53
  45. Prompt 3p. 54
  46. Promptp. 55