Comment on Safety Considerations for Chemical and/or Biological AI Models

On the 10th of October, the AI Safety Institute (AISI) housed within the National Institute of Standards and Technology (NIST) released a request for information on Safety Considerations for Chemical and/or Biological AI Models. The request sought information and insights on current and future practices and methodologies for the responsible development and use of chemical and biological (chem-bio) AI models.

CAIP characterizes the safety considerations for chem-bio models as follows:

• Chem-bio models are inherently dual-use: the knowledge that is helpful to develop vaccines may also be used to engineer deadly toxins and biological weapons.
• Malicious actors could purposefully utilize these dual-use capabilities, as could other AI models interacting with chem-bio models.
• The specialized nature of chem-bio models means that safety techniques designed for general-purpose models may not be appropriate.
• Even if existing safety techniques can apply to chem-bio models, they may not be sufficiently robust to malicious actors’ attempts at misuse. 

Given these safety considerations, we provide the following recommendations for the US government:

1. Require screening of customers and orders by cloud labs and gene synthesis providers.
2. Establish a program to develop better evaluations for chem-bio AI models.
3. Make technical and non-technical evaluations on chem-bio AI models mandatory.
4. Invest in research and development of technical risk mitigations for chem-bio models.
5. Consider a licensing regime on certain chem-bio models.

Read the full comment here.