Why We’re Doing This
We at the Arcadia Impact Alignment Team want to understand the state of automated research in AI safety compared with capabilities research. Right now, there's no existing way to measure this.
Frontier labs have started publishing how much of their AI R&D is now done by AI [1][2], but they do not separate safety from capabilities. We want to take the first step towards measuring this for AI safety, and learn which tasks AIs help speed up and what the bottlenecks are. We acknowledge that measuring this accurately seems hard, but we think a scrappy point estimate will be valuable. Part of this process is figuring out how to measure this better.
We hope these results provide AI safety researchers, strategists, and policymakers with evidence to guide research, funding, and policy. Over time, we want this to become a public baseline that tracks the capabilities–safety gap and pushes labs to report the two separately.
Dual Use & Privacy
We acknowledge that insights gained from this survey have the potential to speed up capabilities work. We will not release scaffolds, advice, or artifacts that we think could speed up capabilities research. All individual responses will be anonymised and kept private, and we will only publish aggregated results. If you'd rather not answer a question, especially if you think answering could cause harm, just write ‘prefer not to say’.