Frontier Model Disclosure

Definition

Frontier model disclosure is a requirement that developers of high-capability AI systems publish enough information and evidence for qualified outsiders to evaluate safety and deployment claims without necessarily releasing model weights.

Current Synthesis

The source positions disclosure between company assertion and compulsory open-weight release. Its purpose is external contestability: independent researchers need sufficient access to methods, evaluations, limitations, incidents, or other evidence to test what developers say. Disclosure is therefore useful only when its scope, recipients, confidentiality rules, and consequences are concrete enough to support real evaluation.

Key Claims

• High-capability systems warrant stronger disclosure because private claims can shape public deployment decisions.
• Disclosure and open weights are distinct; evaluation access need not require unrestricted model distribution.
• External researchers need evidence sufficient to reproduce, challenge, or qualify developer claims.
• Shared disclosure and evaluation formats can support international coordination before a binding treaty exists.
• Disclosure without access rights, standards, or consequences can become transparency theater.

Evidence

• Middle path - EP 35: Who Actually Controls AI? The Governance Gap Explained calls for mandatory disclosure while explicitly saying that the requirement need not mean open weights.
• External evaluation - EP 35: Who Actually Controls AI? The Governance Gap Explained says outsiders need enough transparency to assess company claims.
• Coordination link - EP 35: Who Actually Controls AI? The Governance Gap Explained pairs information sharing with common evaluation standards as a practical early form of international governance.

Counterevidence & Qualifications

The episode does not define which systems qualify, which artifacts must be disclosed, who receives access, how trade secrets or security-sensitive details are protected, or what follows from an adverse finding. Excessive public disclosure can itself increase misuse or erode legitimate confidentiality. The concept therefore records an evidence-access principle, not a complete statutory design.

What Changed

• Established disclosure as a distinct governance mechanism between secrecy and mandatory open weights.
• Connected external contestability to shared international evaluation practices.

Related Concepts

• Frontier Model Release Governance - deployment gate that disclosure evidence can inform.
• AI Safety Coordination - information-sharing context in which common disclosure formats may operate.
• Open Source AI Models - broader access model that disclosure does not necessarily require.
• Mandatory AI Incident Investigation - post-incident evidence-access mechanism with a narrower trigger.
• AI Regulatory Capture Risk - risk that disclosure obligations are designed around incumbent capabilities.