What would you like to know?

Answer

AI safety testing

Based on the areas officials have flagged publicly, a federal approach would probably centre on reporting requirements for companies training the largest and most capable models, safety-testing obligations before major releases, and some form of incident-reporting regime for serious failures or misuse. It would likely stop well short of a licensing system for AI development itself, reflecting both industry lobbying and a general administration preference for lighter-touch rules that avoid slowing domestic AI investment relative to international competitors, particularly China.

• Mandatory reporting to a federal body once training runs exceed a defined computing-power threshold, mirroring an approach already floated in earlier federal AI guidance.
• Required disclosure of safety-testing results for frontier models before public release, though the scope of what counts as a qualifying model remains contested.
• An incident-reporting mechanism for serious AI failures, modelled loosely on existing frameworks in aviation and pharmaceuticals.
• Possible export-control coordination tying chip and compute access to safety-compliance status, linking the domestic framework to broader national-security policy.