OpenAI's Response to NIST's AI Executive Order

OpenAI responded to NIST's Request for Information (RFI) under the AI Executive Order, presenting its views on three key topics: AI capability evaluation, red-teaming, and synthetic media and provenance.

OpenAI currently tracks four major risk areas—cybersecurity, CBRN (chemical, biological, radiological, and nuclear), persuasion, and model autonomy—through its Preparedness Framework to mitigate potential risks.

In a recent large-scale evaluation related to CBRN, GPT-4 showed only a mild uplift in its ability to provide information on generating biological threats compared to existing internet search results. This suggests the importance of measuring how much greater a risk AI poses compared to existing resources.

For effective AI risk evaluation, OpenAI emphasized the following three principles:

• Measuring change relative to a baseline: The degree to which AI poses greater risk than existing resources (e.g., internet search) should be measured.
• Collaboration with domain experts: External experts in the relevant field should be involved in study design and scoring to ensure objective evaluation.
• Collaboration with AI experts: Evaluations should be designed based on a deep understanding of AI systems in order to fully elicit a model's potential capabilities.