Publisher: Anthropic Published: 2025-07-07 URL: https://www.anthropic.com/news/the-need-for-transparency-in-frontier-ai
"A Framework for AI Development Transparency" is a policy proposal published by Anthropic on 7 July 2025. It proposes a targeted transparency regime for the largest frontier AI developers, built around mandatory disclosure of secure-development practices and system documentation, backed by whistleblower protections. Anthropic frames the proposal as preferring "regulatory flexibility as the technology evolves" over prescriptive rules.
Summary
The framework applies only to the largest frontier developers rather than to the industry as a whole, and is presented as an affirmative regulatory model that an industry could adopt while standards continue to develop. Anthropic describes the approach as anti-prescriptive, favoring "regulatory flexibility as the technology evolves" so that industry consensus can form without rigid mandates.
Key provisions
The proposal sets out four elements:
- Secure Development Framework (SDF). Covered developers would publicly disclose how they assess and mitigate risks, with particular attention to chemical, biological, radiological, and nuclear (CBRN) risks and to misaligned-autonomy risks. The requirement functions as a public-disclosure mandate analogous to Anthropic's own Responsible Scaling Policy (Anthropic's Responsible Scaling Policy (Version 3.1)).
- System cards. Developers would document testing procedures, evaluation results, and safety mitigations at deployment, and update that documentation for substantial revisions.
- Whistleblower protection. Making a false statement about compliance would become a legal violation, which in turn would activate existing whistleblower protections. This element supplies the proposal's enforcement mechanism.
- Flexible standards. The framework avoids rigid prescriptions so that industry consensus can evolve over time.
Covered-entity thresholds
The framework targets only the largest developers, applying to a developer with either annual revenue of approximately $100 million or research-and-development or capital expenditures of approximately $1 billion per year. Smaller developers and startups fall below these thresholds and would be exempt. Anthropic relates this scoping to the combined compute-threshold and revenue-test approach used in California SB 53 — Transparency in Frontier AI Act and New York RAISE Act (S. 8828).
Reception and context
The proposal is one of the first frontier-lab policy proposals advancing binding transparency rules. Its structure tracks that of California SB 53 and the New York RAISE Act, legislation in which Anthropic was visibly involved. Anthropic presents the framework as its affirmative case for regulation, in its own words: "we're not just lobbying against regulation — here's what we're lobbying for." The stance is described as contrasting with OpenAI's more preemption-first and liability-shield-first posture (see Illinois SB 3444 — Artificial Intelligence Safety Act).
Relationships
- supports: California SB 53 — Transparency in Frontier AI Act, New York RAISE Act (S. 8828), AI Whistleblowing.
- depends-on: Anthropic's Responsible Scaling Policy (Version 3.1), Frontier Compliance Framework (February 2026).
- related: Managing Advanced Cyber Risks in Frontier AI Frameworks, Anthropic, OpenAI (for contrast).
- contradicts: Heavy prescriptive regulatory approaches like EU AI Act (Regulation 2024/1689) (Anthropic advocates flexibility in implementation).