Robots Atlas>ROBOTS ATLAS
Ethics & Law

Meta-transparency

2024ActiveUpdated: 17 November 2026Published
Key innovation
Moving the transparency requirement up one level: instead of disclosing every instruction, an AI system openly discloses what kinds of rules govern it and where and why full disclosure is limited.
Category
Ethics & Law
Abstraction level
Pojฤ™cie
Operation level
Law and governanceSystem
Use cases
Public model specs and model constitutionsPolicies for disclosing system and developer promptsFlagging content omitted or added for legal reasonsAI platform transparency reports and governance accountabilityAuditing the gap between stated and observed behaviour

How it works

Meta-transparency is realised through several practices. First, a top-level rules document (model spec, constitution, policy) is published, describing the principles and the hierarchy of instructions governing the model's behaviour. Second, the existence of confidential layers is disclosed โ€” it is acknowledged that beyond the public rules there are system or developer instructions, even if their content is not shown. Third, omissions and modifications are flagged: when content is removed or added (e.g. due to local law), the model openly indicates the type and reason of the intervention without revealing the redacted content itself. Fourth, the rationale for the limits of disclosure is given, so that opacity is predictable and accountable. Thanks to public rules, third parties can audit the gap between stated and observed behaviour.

Problem solved

It resolves the tension between the need for trust and accountability and the unavoidable confidentiality of some instructions (for security, legal, privacy or abuse-prevention reasons). Instead of choosing between full disclosure and complete opacity, it makes the structure and limits of disclosure themselves legible, letting a user evaluate the system without exposing sensitive details.

Components

Public rules documentFoundation โ€” it makes the rules governing the model observable and open to discussion.

A publicly available specification of rules (model spec, constitution, policy) describing how the model's behaviour is shaped and what the hierarchy of instructions is.

Disclosure of confidential instructions' existenceStructural honesty โ€” it removes the illusion of full disclosure.

Acknowledging that alongside the public rules there are non-public system or developer instructions, without revealing their content.

Flagging of omissions and modificationsResponse-level accountability โ€” opacity becomes visible at the point where it occurs.

Openly indicating that content was removed or added (e.g. due to law), with the type and reason of the intervention, but without revealing the redacted content.

Rationale for the limits of disclosurePredictability โ€” it makes opacity understandable and assessable.

An explanation of which categories of information are not disclosed and why (privacy, security, intellectual property, legal requirements).

Implementation

Implementation pitfalls
Meta-transparency as a fig leafHigh

Declaring principles while hiding so many instructions that the public document does not let one realistically predict or audit the model's behaviour.

Fix:Ensure the scope and reasons for confidentiality are narrow, justified and verifiable, and that the stated principles match observed behaviour.
Confusing it with full transparencyMedium

Treating the disclosure of the rules themselves as equivalent to disclosing all instructions; meta-transparency deliberately does not reveal confidential content.

Fix:Clearly communicate that what is disclosed is the structure and limits of disclosure, not every individual instruction.

Evolution

2024
First public OpenAI Model Spec
Inflection point

OpenAI publishes the Model Spec โ€” a document describing the principles that shape model behaviour, institutionalising transparency about the rules themselves.

2025
Model Spec released to the public domain (CC0)

Updates to the Model Spec are released under CC0, together with a commitment to further transparency when model behaviour is adapted in significant ways.