Some research and project ideas for developing and refining the field of AI safety

A researcher proposes several high-impact AI safety projects, including live databases of research questions and interventions to track progress and neglectedness, and exploring under-contacted disciplines and attractor basins as risk vectors.

This small post is part of a series of posts I'm creating which outline my ideas for AI safety projects/research which I think could be high-impact. I have way more ideas than I could ever hope to do justice to, so if you're looking for inspiration, take a look There should be a live, regularly-updated and highly visual database of AIS research questions and the progress on each of them . The nearest existing thing is the EA Midwest list-of-lists https://www.notion.so/eamidwest2026/List-of-lists-of-projects-ideas-in-AI-Safety-291f040b30f180a794c3f1f6b5264534 ; it could be far more. Could plausibly be made and maintained mostly with AI scraping and auto-categorisation. A database like this would allow people to easily see the state of research on a given subtopic, and where there is progress to be made. I suspect that the majority of updating could be automated. Similarly, there should be a live, regularly-updated and highly visual database of proposed interventions in AI safety, which tracks how many people are working on each intervention, and roughly how much time they're on each intervention . This would allow us to more quantitatively assess neglectedness. Which disciplines have had the least contact with AI safety, and what could they contribute? See this Claude deep research artefact: https://claude.ai/public/artifacts/55bde7d3-2216-43ea-83f0-9857e1e48750 https://claude.ai/public/artifacts/55bde7d3-2216-43ea-83f0-9857e1e48750 Are attractor basins from AI use a risk vector that curtails genuine innovation in AIS itself? One candidate antidote: deliberately spending time in layers of reality that are far from AI eg nature in order to tap into sources of inspiration which lack these attractor basins. Given the high degree of disagreement among experts regarding which AI safety interventions are most promising, would it be helpful for intervention comparisons to factor in: