We welcome the opportunity to respond to the NTIA’s AI Accountability Policy Request for Comment and look forward to future opportunities to provide additional input. We offer the following submission for your consideration:
● Scope. This submission focuses on audits and assessments of foundation models. Foundation models are large pre-trained models that can serve as the“foundation” for a wide array of downstream applications. These models already cause harm and might cause even more harm in the future.
● The need for public accountability. As foundation models become increasingly powerful and important to society, decisions about their development and deployment need to be accountable to the public interest. Policymakers need more information to govern these technologies. Audits and assessments can provide this information.
● Challenges. However, auditing and assessing foundation models is challenging.In particular, there are not enough experts who can audit foundation models, external actors often do not have sufficient access to the models, and there are no established evaluation criteria or methodologies.
● Evaluation criteria. Foundation models should at least be evaluated against three criteria: dangerous capabilities, alignment, and truthfulness. Auditors should test models against pre-defined benchmarks, while also trying to elicit harmful behavior and conducting more exploratory evaluations.
● Ecosystem. Effective audits of foundation models require an ecosystem of independent expert auditors with access to the relevant models and strong incentives to find flaws rather than to “tick boxes”.
● Recommendations. Based on the above, we recommend concrete actions that government can take today and in the future.
● Appendix A and B: We answer additional questions from the NTIA request (10, 14, 16, 20), and share the results of our expert opinion survey on AI governance best practices.