My last recommendation is to support the development of evaluations for AI capabilities and risks. The AI Action Plan already includes this, but it should go one step further and consider restricting models that fail to meet industry-standard performance levels of safety. (7/7)

My second recommendation is to push AI companies to share safety-relevant knowledge with each other and other relevant stakeholders. This would involve mandating reporting requirements, disclosure of unexpected capabilities in new models, and sharing threat intelligence. (6/7)

My first recommendation is to require AI companies to adhere to their own risk management plans. Companies like OpenAI and Anthropic have already published frameworks describing their planned risk mitigations, but these need to be made legally binding to have any effect. (5/7)

At the same time, AI is advancing too rapidly for government to keep up with traditional regulation. The solution is to promote industry self-regulation – make AI companies figure out the best way to keep their products safe and then make sure they actually follow through. (4/7)

It's true that we need to increase AI adoption, but quite simply – people don't want to use things that aren't guaranteed to work! There's still too many hallucinations, security concerns, and other liabilities for companies to feel confident relying on AI for important tasks. (3/7)

Check it out here: (2/7)

Check out the full response to the RFI here: [4/4]
cset.georgetown.edu/publication/...

These standards, if implemented, would go a long way towards mitigating the potential risks of AI and increasing public trust and confidence in using it, allowing us to realize its benefits sooner than we could otherwise. [3/4]

Specifically, AISI should develop standards on topics such as model training, pre-release internal & external security testing, cybersecurity practices, if-then commitments, AI risk assessments, and processes for testing and re-testing systems as they change over time. [2/4]