CASRAI analysis finds AEF-1 defines independent evaluation more narrowly than Frontier Model Forum, excluding confirmatory work
AI Auditor Ecosystem Faces Talent Shortage and Conflict-of-Interest Concerns
If legislation mandating third-party AI audits passes before enough qualified evaluators exist, the oversight regimes may be unworkable in practice. The small size of the existing evaluator pool also makes conflict-of-interest problems structurally difficult to avoid.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
Efforts to build independent verification organizations (IVOs) to audit AI companies are gaining legislative momentum at state and federal levels, but the evaluator ecosystem faces a critical shortage of technically qualified people. Too few individuals hold the dual skill set of frontier AI model knowledge combined with evaluation domain expertise, which threatens to make robust oversight regimes unworkable if mandated before that capacity exists.
The AI Evaluator Forum published AEF-1, a voluntary baseline standard establishing five minimum operating conditions for independent third-party AI evaluations: sufficient access and resources, minimized conflicts of interest, analytic autonomy, transparent methods and results, and protection of sensitive information. xAI, OpenAI, and Anthropic cosigned the proposal. AEF-1 is explicitly self-described as an initial baseline, not a comprehensive standard, and includes a checklist evaluators can publish alongside results.
METR has stated its central constraint is technical capacity, not access to AI developers. Training pipelines for model-level evaluators are at least a year from enrolling students. State bills in Massachusetts and Illinois and federal proposals including the FRONTIER Act and a forthcoming Jacobs/Beyer bill are moving toward IVO mandates, though none yet requires the most rigorous form of ongoing auditing. Proponents argue that passing such a law would create a demand signal that grows the evaluator ecosystem, drawing an analogy to what happened after the Securities Acts of 1933 and 1934.
Conflict-of-interest concerns also shadow the ecosystem. METR researcher Ajeya Cotra drew criticism for participating in an OpenAI investigation because her spouse holds a role at OpenAI. Kevin Bass alleged financial entanglement between Anthropic and parts of the AI safety and evaluation ecosystem, arguing those ties undermine independence claims.
AEF-1 defines independent evaluation more narrowly than some peer frameworks. It requires an external organization as evaluator and excludes work that merely replicates or validates a developer's own methodology. The Frontier Model Forum allows a sufficiently independent internal team for open-ended assessments, creating a divergence between frameworks on whether internal teams can legitimately fill the independent evaluator role.
One structural limitation analysts identify in AEF-1 is that it scopes evaluation as a discrete event, which is appropriate for static artifacts but insufficient for AI systems whose behavior can change after assessment. A January 2026 report identified this recurrence gap and proposed solutions for continuously changing systems.
How it developed
Reporting on the independent AI auditor talent shortage and conflict-of-interest concerns published
Analysis notes AEF-1 treats evaluation as a one-time event, leaving dynamic AI systems inadequately covered
xAI, OpenAI, and Anthropic cosign AEF-1; Kevin Bass posts thread alleging financial ties between Anthropic and AI safety evaluators
AI Evaluator Forum publishes AEF-1, a baseline standard for independent third-party AI evaluations covering independence, access, and transparency
Sources
Related
- Related storyThe AI Evaluator Forum's AEF-1 standard
- Massachusetts Senate AI lab audit bill
- Anthropic's leaked draft AI bill
- The Great American AI Act discussion draft
- California's SB 813 voluntary AI safety framework
BackgroundWhen AI doesn't do what we meant
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.