The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Trends

Emerging hypothesis · Active

Evaluation access is growing faster than evidence of independent oversight

Latest assessment: · Evidence updated

Standing score
Unscored
Direction of power
Undetermined
Window
Unclear
Confidence
Low

What we think is happening

Testing relationships and shared reporting may expand access while developers or governments retain control of timing, methods and publication. The mechanism is conditional oversight: evaluators can inspect some capability without gaining autonomous accountability authority.

Scope: International/public evaluation arrangements and reporting frameworks; no completed independence audit.

Potential direction of power

Government testing coordination and shared reporting can improve access without giving evaluators control over timing, methods or publication. The reported U.S.-first access restriction remains unconfirmed. Internal release decisions alone cannot establish whether independent public oversight exists.

Supporting cases

Australia explains its new AI review committee's advisory role

Australia's Digital Transformation Agency published an account October 7 of the AI Review Committee's first meeting, held June 26. The committee advises agencies on selected high-risk, sensitive or novel AI uses before deployment. Its first discussions included document-analysis work and an internal document-search proof of concept.

The committee's terms make its advice non-binding. It does not endorse deployments, grant immunity or replace agencies' responsibility for legality and implementation. It can also review serious incidents after an agency's response. This is added government review capacity, rather than an independent regulator with a veto over agency AI.

For The Race, the significance is an institution that supplies expertise across agency boundaries. Its advisory character and accountability arrangements limit what can be inferred about public control.

Distinct connection: For The Race, the significance is an institution that supplies expertise across agency boundaries. Its advisory character and accountability arrangements limit what can be inferred about public control. Overlapping evidence is not independent confirmation.

Read the case and sources →

A study finds little published safety evidence across Chinese model releases

SemiAnalysis’s October 8 study examined 857 model releases from nine leading Chinese developers between 2021 and September 15, 2026. It found published developer safety results for 31 releases, or 3.6%; just nine had results available at or before launch.

Researchers matched model cards, technical reports and release notes to specific releases. General claims that a model was evaluated did not qualify. Tests could cover harmful outputs, privacy, jailbreak resistance or dangerous capabilities. “No disclosure found” does not mean no private testing occurred. Variant-counting practices differ between companies, and the study does not provide a comparable US denominator. Primary study ↗; Corroborating reporting ↗.

This helps distinguish published regulatory architecture from what outsiders can inspect about particular models. It strengthens the need to examine disclosure and independent oversight separately, without turning a documentation study into a finding that the models are unsafe.

Read the case and sources →

Window of contestability

Unclear for public evaluators: testing access alone does not establish control of methods, timing, publication or remedies.

What would change this assessment

Confirmed access, independent funding and publication rights, participation by affected institutions and an effective route to require correction would demonstrate stronger accountability.

Related cases

← All 24 trends