Secrecy vs Accountability Risks
- Analysts say keeping model internals secret can strip away external accountability in security‑focused settings. - Nathan P. Goodman explicitly linked secrecy to accountability problems, and critics flagged light regulatory oversight. - Those arguments suggest governance must weigh safety disclosures against competitive secrecy pressures to avoid unreviewed deployments ( ).
A fight over AI secrecy has shifted from trade secrets to oversight: analysts say hiding how powerful models work can also hide who answers when they fail. (springer.com) Accountability in AI usually means answerability, records, and consequences when a system causes harm. The National Telecommunications and Information Administration said audits, assessments, and other assurance tools depend on giving outside parties enough information to test and challenge a system. (ntia.gov, ntia.gov) That tension is sharper for frontier models, the largest systems built for broad use, because companies often treat model weights, training methods, and internal evaluations as closely held information. The Frontier Model Forum, an industry group launched by Anthropic, Google, Microsoft, and OpenAI in July 2023, says independent, standardized evaluations are part of responsible deployment. (frontiermodelforum.org, anthropic.com) The current argument is not that every lab should publish every detail. It is that security claims are harder to verify when the same company controls the model, the tests, and the public summary of the results. (nathanpgoodman.com, gao.gov) That question has moved from theory to practice. On August 27, 2025, OpenAI and Anthropic published results from a joint safety evaluation in which each company ran internal tests on the other’s public models and released findings in parallel. (openai.com, alignment.anthropic.com) Anthropic said the exercise tested for sycophancy, whistleblowing, self-preservation, support for human misuse, and the ability to undermine safety evaluations and oversight. It also said some model-external safeguards were relaxed so the tests could probe the underlying systems more directly. (alignment.anthropic.com) Governments are building disclosure rules, but many remain limited or voluntary. The European Commission published its General-Purpose AI Code of Practice on July 10, 2025 as a voluntary tool to help companies comply with the European Union AI Act’s obligations for general-purpose models. (digital-strategy.ec.europa.eu) In the United States, the National Institute of Standards and Technology’s AI Risk Management Framework is also voluntary, and NIST’s generative AI profile focuses on risk identification and mitigation rather than mandatory public disclosure. NIST said on April 7, 2026 that it had released a concept note for a new profile on trustworthy AI in critical infrastructure. (nist.gov) Civil-society and policy groups have argued for stronger outside review in higher-stakes settings, including national security. The Brennan Center wrote that national-security AI systems operate in a space with “extremely limited oversight” and called for an oversight authority that can bring more transparency to those uses. (brennancenter.org, justsecurity.org) The thread running through these debates is simple: companies want to protect competitive secrets, while regulators, researchers, and affected users want enough visibility to check safety claims before deployment. As more powerful models move into security-focused and critical systems, that balance is becoming a governance decision, not just a product one. (ntia.gov, metr.org)
