OpenAI Shows AI Can Potentially Aid in Biological Threat Creation
As part of its Preparedness Framework, OpenAI is investing in the development of better ways to evaluate AI-enabled safety risks. This is important because as AI systems become more capable, there is a greater potential for both beneficial and harmful uses of AI. One harmful use that has been highlighted is the ability for AI systems to assist malicious actors in creating biological threats.
To address this issue, OpenAI has been working on methodologies to evaluate these types of risks. One new evaluation focused on measuring the potential for AI models to increase malicious actors’ access to dangerous information about biological threat creation, compared to the baseline of existing resources (i.e., the internet).
The evaluation involved a study with 100 human participants, including 50 biology experts and 50 student-level participants. Each group of participants was randomly assigned to either a group with access to the internet only, or a group with access to GPT-4 in addition to the internet. The participants were then asked to complete a set of tasks related to the process of biological threat creation.
The study found mild increases in accuracy and completeness for those with access to the AI language model. However, these increases were not statistically significant, and more research is needed to understand what performance thresholds indicate a meaningful increase in risk.
OpenAI plans to share more details about the evaluation procedure and the results it yielded. They will also discuss methodological insights related to capability elicitation and security considerations for running this type of evaluation with frontier models at scale.
In conclusion, these findings suggest that while AI can potentially aid in biological threat creation, more research is needed to better understand and address the risks associated with AI-enabled safety.