AEF-1 提出獨立第三方 AI 評估最低運作條件標準與檢查清單
AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations (opens on another site)
AEF-1 提出一套供獨立第三方 AI 評估使用的標準與檢查清單,讓評估方披露支援獨立性、存取權及透明度的運作條件。文中列出對 77 個模型變體的心理健康評估,以及包含 34 項評估、其中 5 項為內部基準的失控與有害操縱風險評估等案例。EU AI Office 已認可 AEF-1 的關鍵條款,可作為供應商遵守《通用人工智能行為準則》獨立性條款的方式。
Independent third-party evaluations of AI systems are becoming increasingly central to AI governance approaches pursued by developers, governments, enterprises, and the public. However, the operating conditions of third-party evaluations can be opaque, even though they can significantly impact whether an evaluation is trustworthy and impartial. To address this, we present AEF-1, a standard and checklist that third-party evaluators can use to demonstrate how they achieved a set of operating conditions that support a baseline level of independence, access, and transparency during an evaluation.
The success of AEF-1 requires adoption by a range of stakeholders, including frontier developers.
AEF-1 in practice
Evaluations by Forum members carried out or documented under the conditions the standard describes.
![]()
Mental Health Evaluation
Transluce · August 2026
An independent assessment of how 77 model variants respond to users experiencing mental health crises, published with a detailed disclosure of operating conditions consistent with AEF-1.
How AEF-1 was used: Included the AEF-1 checklist, disclosing the broader operating conditions the evaluation ran under and what may have affected its independence, access, and transparency.
See the evaluationSee the checklist
![]()
Frontier Risk Report
METR · May 2026
An assessment of misalignment risk from AI agents deployed internally at Anthropic, Google, Meta, and OpenAI, carried out with more direct access to non-public information and more editorial independence than previous external evaluation engagements.
How AEF-1 was used: Included the AEF-1 reporting checklist in the published report.
See the evaluationSee the checklist
![]()
Biosecurity pre-release assessment
SecureBio · September 2026
Independent pre-release assessment of biosecurity-relevant capabilities in OpenAI's GPT-6 Astra, including knowledge and agentic benchmarks, safeguard refusals, and expert red-teaming.
How AEF-1 was used: Included the AEF-1 checklist in the published report, and uses AEF-1 as a foundation of its organizational principles.
See the reportSee the checklistSee their principles
Loss of control and harmful manipulation evaluation
Fangcun AI · October 2026
An assessment of loss of control and harmful manipulation risks using 34 evaluations, including 5 in-house benchmarks.
How AEF-1 was used: Included the AEF-1 checklist, fulfilled all minimum operating conditions.
See the evaluationSee the checklist
![]()
Independence provisions under the General-Purpose AI Code of Practice
European AI Office
The EU AI Office has endorsed key provisions of the AEF-1 standard as a means for providers to comply with the independence provisions of the General-Purpose AI Code of Practice.
來源:Transluce · aievaluatorforum.org