CDT and Coalition Urge NIST to Cover Discovery-Oriented AI Evaluation

What the framework does

The National Institute of Standards and Technology (NIST) has published a draft framework that gives organizations structured guidance on how to test, evaluate, verify and validate (TEVV) AI systems. According to CDT, the aim is to enable AI products that are more fit for purpose and to support more effective risk management by AI developers and deployers.

CDT says its submission praises NIST's effort to improve best practices in this area. In its view, the guidance will provide a helpful concrete framework and standardized terminology. Stakeholders can use these to build their own practices for judging whether an AI system will be effective in its intended context and for identifying risks and impacts that deserve attention. CDT argues such frameworks can lead to AI systems that better serve consumers' needs, protect individual rights and are less likely to produce unexpected harms.

The case for discovery-oriented evaluation

The central recommendation is that NIST better account for what the comments call "discovery-oriented" evaluations. In these, an organization does not only measure specific properties known in advance. It also works to understand what is important to measure in the first place.

CDT says this approach matters especially for general-purpose AI tools, where evaluators often prioritize surfacing unknown risks, impacts and failure modes. To support it, the comments recommend that NIST:

• adjust the definition of TEVV that it adopts;
• include more guidance on how users can define the constructs to be measured;
• include mechanisms for refinement and feedback;
• add an example of discovery-oriented evaluation in a more complex scenario.

Roles, system types and real-world conditions

The coalition also asks for more concrete guidance on which actors are best placed to intervene at different development stages. This would help organizations understand the multi-disciplinary nature of a TEVV-Athlon. The comments further call for specific considerations for different types of AI systems, with agentic applications given as one example.

On realism, the comments suggest NIST expand its treatment of how to make sure test scenarios and environments reflect real-world AI uses. They also ask the agency to consider settings where AI systems are used in other languages and in multi-cultural contexts, and to describe methods evaluators can use to broaden public participation. Finally, the groups want NIST to integrate more domain-specific resources from the NIST AI Metrology Center into its guidance on how evaluators can move from concept to practice.

What it means

The framework is still a draft, and the coalition's comments are input to NIST's process rather than final changes. For organizations that build or deploy AI, the exchange signals where civil society groups want evaluation guidance to go: toward finding unforeseen problems in general-purpose systems, and toward testing that reflects different languages, cultures and real-world use.