AI Standards Lab experts have contributed to the Evals-Consensus project to identify best practices for AI evaluations. We endorse the resulting expert consensus: “Based on the consensus of stakeholders and experts, we recommend evaluation practitioners consider and adopt at least these 27 practices where relevant for constructing, documenting, maintaining, using, and reporting the results of AI evaluations.”
This statement reflects the outcome of a cross-sector Delphi process, which has drawn on input from AI labs, auditors, academia, civil society, and government, and has produced community-endorsed guidance for conducting and reporting AI evaluations.
Common knowledge about what well-done evaluations look like is valuable for everyone working with AI. We will be using this consensus statement to inform our work on assessing the quality of AI evaluations.