Scheming risk prevention
View weights for all 8 criteria →
Safety case: process
Weighted 25% of category
 | 10% | No clear plan on scheming risks more |
 | 20% | DeepMind has the beginning of a control plan more |
 | 5% | OpenAI is nominally planning to address risks from misalignment, but its plan is very vague and doesn't seem to consider that misalignment could be hard to notice more |
 | 0% | No plan on scheming risks |
 | 0% | No plan on scheming risks more |
 | 0% | No plan on scheming risks |
 | 0% | No plan on scheming risks |
Safety case: accountability
Weighted 25% of category
| 0% | No accountability |
| 0% | No accountability more |
| 0% | No accountability |
| 0% | No plan on scheming risks |
| 0% | No plan on scheming risks |
| 0% | No plan on scheming risks |
| 0% | No plan on scheming risks |
Internal deployment protocol
Weighted 20% of category
| 0% | No but Anthropic has done research on probing |
| 5% | DeepMind is planning to implement monitoring to address risks from deceptive alignment by the time its models have a certain level of instrumental reasoning capabilities more |
| 0% | No but OpenAI has done research on monitoring |
| 0% | Nothing on misalignment risks during internal deployment |
| 0% | Nothing on misalignment risks during internal deployment |
| 0% | Nothing on misalignment risks during internal deployment |
| 0% | Nothing on misalignment risks during internal deployment |
External deployment protocol
Weighted 10% of category
| 0% | Anthropic is doing nothing on this |
| 0% | DeepMind is doing nothing on this |
| 0% | OpenAI is doing nothing on this |
| 0% | Meta is doing nothing on this |
| 0% | xAI is doing nothing on this |
| 0% | Microsoft is doing nothing on this |
| 0% | DeepSeek is doing nothing on this |
Plan for if an AI is caught scheming
Weighted 10% of category
| 0% | Anthropic doesn't do anything on this |
| 0% | DeepMind doesn't do anything on this |
| 0% | OpenAI doesn't do anything on this |
| 0% | Meta doesn't do anything on this |
| 0% | xAI doesn't do anything on this |
| 0% | Microsoft doesn't do anything on this |
| 0% | DeepSeek doesn't do anything on this |
Training: remove scheming capabilities
Weighted 3% of category
| 0% | Anthropic doesn't do this |
| 0% | DeepMind doesn't do this |
| 0% | OpenAI doesn't do this |
| 0% | Meta doesn't do this |
| 0% | xAI doesn't do this |
| 0% | Microsoft doesn't do this |
| 0% | DeepSeek doesn't do this |
Training: adversarial training
Weighted 3% of category
| 0% | Anthropic doesn't do this |
| 0% | DeepMind doesn't do this more |
| 0% | OpenAI doesn't do this |
| 0% | Meta doesn't do this |
| 0% | xAI doesn't do this |
| 0% | Microsoft doesn't do this |
| 0% | DeepSeek doesn't do this |
Training: safe architecture
Weighted 3% of category
| 50% | Reasoning is in natural language, but no policy or articulation of this principle more |
| 55% | Reasoning is in natural language, but no policy, but articulation of this principle, but only in a research paper more |
| 55% | Reasoning is in natural language, but no policy, but articulation of this principle, but only in a research blogpost |
| 50% | Reasoning is in natural language, but no policy or articulation of this principle |
| 50% | Reasoning is in natural language, but no policy or articulation of this principle |
| 50% | Reasoning is in natural language, but no policy or articulation of this principle |
| 50% | Reasoning is in natural language, but no policy or articulation of this principle |