Which tools reduce information asymmetry between AI developers and users?

• A. Non-Disclosure Agreements and secrecy about capabilities.
• B. Model Cards, Safety Audits, Third-Party Evaluations, Transparency Reports, and Clear Capability Disclosures.
• C. Hiding performance metrics and evaluation results.
• D. Exclusive internal testing without external review.

Answer: (B)

Reducing information asymmetry comes from transparency and independent verification. Model Cards provide standardized summaries of what the model is intended to do, its capabilities, limitations, and appropriate use cases, helping users quickly assess suitability. Safety Audits bring expertise-based checks that examine safety and alignment issues, offering a rigorous assessment from within an organization or by an external party. Third-Party Evaluations add objective, external validation of claims about performance and behavior, increasing trust beyond internal reports. Transparency Reports disclose how the model behaves over time, including incidents, failure modes, and updates, giving users a clear record of accountability. Clear Capability Disclosures explicitly state what the model can and cannot do, along with known risks, so users can set realistic expectations and avoid misuse.

Options built on secrecy or hiding metrics, such as withholding performance data or relying on exclusive internal testing, keep users in the dark and widen information gaps, undermining safety and trust.