SafetyPage

Government · Government · UK Government

Frontier AI: capabilities and risks – discussion paper

28 Apr 2025

ReadSource at gov.uk ↗Read hereOGL v3.0

Outline · 757 paragraphs

  1. Frontier AI: capabilities and risks – discussion paper¶1
  2. Introduction¶6
  3. What is the current state of frontier AI capabilities?¶17
  4. How frontier AI works¶19
  5. Frontier AI can perform many economically useful tasks¶29
  6. Frontier AI models can be augmented with tools to make them more autonomous¶50
  7. Frontier AI could be more capable than evaluations indicate¶60
  8. Limitations of frontier AI¶69
  9. How might frontier AI capabilities improve in the future?¶81
  10. Recent AI progress has been rapid¶83
  11. Recent progress was driven by systematic trends in compute, data and algorithms¶92
  12. Scaling laws: performance improves predictably with increased compute and data¶97
  13. Rapid AI progress is likely to continue for several years¶110
  14. Advanced general-purpose AI agents might be developed in the future¶119
  15. What risks do frontier AI present?¶123
  16. Cross cutting risk factors¶130
  17. It is difficult to design safe frontier models in open-ended domains¶133
  18. Evaluating the safety of frontier AI systems is an open challenge¶137
  19. It may be difficult to track how frontier AI systems are deployed or used¶143
  20. AI safety standards have not yet been established¶151
  21. Insufficient incentives for AI developers to invest into risk mitigation measures¶154
  22. There may be significant concentration of market power in AI¶157
  23. Societal harms¶160
  24. Degradation of the information environment¶162
  25. Labour market disruption¶179
  26. Bias, Fairness and Representational Harms¶184
  27. Misuse risks¶191
  28. Dual Use Science risks¶193
  29. Current capabilities¶195
  30. Projected capabilities¶199
  31. Potential Risks and Impacts¶201
  32. Cyber¶203
  33. Current Cyber Capabilities of Frontier AI¶205
  34. Projected Capabilities¶210
  35. Potential Risks and Impact¶216
  36. Disinformation and Influence Operations¶220
  37. Loss of control¶226
  38. Humans might increasingly hand over control to misaligned AI systems¶233
  39. Future AI systems might actively reduce human control¶237
  40. Future AI systems may have the disposition to reduce human control¶240
  41. Frontier AI shows early signs of capabilities that could be used to reduce human control¶246
  42. Conclusion¶258
  43. Glossary¶263
  44. Annexes¶293
  45. Descriptions of figures¶296
  46. References¶355