Executive summary: An evidence-based comparative analysis of Anthropic’s Responsible Scaling Policy, Google DeepMind’s Frontier Safety Framework, and OpenAI’s Preparedness Framework (all updated in 2025) finds broadly similar, misuse-focused approaches to monitoring dangerous capabilities (bio/chem, cyber, and AI self-improvement) but highlights weakening commitments, governance differences, and persistent vagueness about concrete “if-then” actions—leaving substantial uncertainty about whether these policies would prevent catastrophic outcomes.
Key points:
Common architecture, different labels: All three frameworks commit to testing for dangerous capabilities and gating deployment behind safeguards; they track broadly the same areas (CBRN/bio-chem, cyber, and AI self-improvement), emphasize misuse over misalignment, and use threshold concepts (Anthropic “Capability Thresholds”/ASLs, DeepMind CCLs, OpenAI high/critical risk tiers).
How risks are evaluated: Anthropic triggers comprehensive assessments after step-change indicators and tests “safety-off” variants; DeepMind runs Early Warning Evaluations with alert thresholds and brings in external experts; OpenAI relies on scalable automated proxies validated by deep-dive red-teaming and domain tests.
What happens at the thresholds: Anthropic pairs thresholds with ASL-3/4 deployment and security safeguards plus executive/board signoffs; DeepMind requires a governance-approved “safety case” and RAND-style security levels but is explicit that some measures need field-wide coordination; OpenAI allows deployment of “high-risk” models only with safeguards and pledges to pause training for “critical-risk” models.
Governance and posture differences: Anthropic foregrounds internal roles, whistleblowing, and public capability reports; DeepMind spreads authority across multiple councils and stresses industry co-adoption; OpenAI routes decisions through a Safety Advisory Group and board committee, with a notable (but high-level) training-pause commitment.
2025 regressions and recalibrations: Labs added process detail but also softened parts of earlier commitments—e.g., conditional adoption tied to competitors, reduced safeguards for some CBRN/cyber cases, OpenAI removing “persuasion” from its tracked categories, and Anthropic stepping back from pre-defining ASL-N+1 evaluations—raising doubts about robustness under competitive pressure.
Unresolved crux: will this avert catastrophe? The documents remain more specification-plus-tests than operational plans with hard triggers; senior leaders’ stated P(doom) still diverge markedly (e.g., ~25% vs. ~2%), underscoring real uncertainty about whether these frameworks, even if followed, meaningfully reduce existential risk and suggesting a need for stronger, coordinated standards and regulation.
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, andcontact us if you have feedback.
Executive summary: An evidence-based comparative analysis of Anthropic’s Responsible Scaling Policy, Google DeepMind’s Frontier Safety Framework, and OpenAI’s Preparedness Framework (all updated in 2025) finds broadly similar, misuse-focused approaches to monitoring dangerous capabilities (bio/chem, cyber, and AI self-improvement) but highlights weakening commitments, governance differences, and persistent vagueness about concrete “if-then” actions—leaving substantial uncertainty about whether these policies would prevent catastrophic outcomes.
Key points:
Common architecture, different labels: All three frameworks commit to testing for dangerous capabilities and gating deployment behind safeguards; they track broadly the same areas (CBRN/bio-chem, cyber, and AI self-improvement), emphasize misuse over misalignment, and use threshold concepts (Anthropic “Capability Thresholds”/ASLs, DeepMind CCLs, OpenAI high/critical risk tiers).
How risks are evaluated: Anthropic triggers comprehensive assessments after step-change indicators and tests “safety-off” variants; DeepMind runs Early Warning Evaluations with alert thresholds and brings in external experts; OpenAI relies on scalable automated proxies validated by deep-dive red-teaming and domain tests.
What happens at the thresholds: Anthropic pairs thresholds with ASL-3/4 deployment and security safeguards plus executive/board signoffs; DeepMind requires a governance-approved “safety case” and RAND-style security levels but is explicit that some measures need field-wide coordination; OpenAI allows deployment of “high-risk” models only with safeguards and pledges to pause training for “critical-risk” models.
Governance and posture differences: Anthropic foregrounds internal roles, whistleblowing, and public capability reports; DeepMind spreads authority across multiple councils and stresses industry co-adoption; OpenAI routes decisions through a Safety Advisory Group and board committee, with a notable (but high-level) training-pause commitment.
2025 regressions and recalibrations: Labs added process detail but also softened parts of earlier commitments—e.g., conditional adoption tied to competitors, reduced safeguards for some CBRN/cyber cases, OpenAI removing “persuasion” from its tracked categories, and Anthropic stepping back from pre-defining ASL-N+1 evaluations—raising doubts about robustness under competitive pressure.
Unresolved crux: will this avert catastrophe? The documents remain more specification-plus-tests than operational plans with hard triggers; senior leaders’ stated P(doom) still diverge markedly (e.g., ~25% vs. ~2%), underscoring real uncertainty about whether these frameworks, even if followed, meaningfully reduce existential risk and suggesting a need for stronger, coordinated standards and regulation.
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.