AI Safety Fund Rejects 'Slop' Applications, Reveals What Gets Funded
If you care about AI safety, here's what insiders look for—and what they reject.
A grant reviewer for the Corrigibility Research Fund—a project that funds research on making AI systems safe and controllable—has written about the five types of applications they routinely reject. The fund's goal is to support work that ensures AI does what humans want, even as it becomes more capable. But many applicants miss the mark, often because they don't understand what 'corrigibility' really means or are just trying to get funding for unrelated projects.
The most common problem is AI-generated 'slop': applications written by AI that are full of errors, bad formatting, or generic content. The reviewer notes that while AI can help non-native English speakers contribute, a high AI-detection score usually means the application is junk. Another major issue is proposals to build 'governance layers' for AI agents—basically, control mechanisms to keep AI in check. The reviewer rejects these because they're often written by AI and don't address the deeper research questions. Similarly, many applicants try to rebrand their existing projects (like climate research or journalism funding) as corrigibility work, which the reviewer sees as a stretch.
Other rejected types include projects that just test whether AIs misbehave (without advancing understanding) and those that are too theoretical or vague. The reviewer emphasizes that good applications are clear, short, and directly address open problems in corrigibility. They also warn against defining basic terms in the introduction, assuming the reviewer doesn't know their own field.
This matters because effective AI safety research is crucial as AI systems become more autonomous. If funding goes to poorly conceived projects, real progress on keeping AI safe could be delayed. The reviewer's insights also reveal the challenges of evaluating AI safety work and the importance of clear communication in research proposals.
- Most grant applications for AI safety research are rejected because they are AI-generated 'slop' or off-topic, showing the difficulty of funding effective work.
- A common but flawed approach is building 'governance layers' to control AI agents, which the reviewer says doesn't address core research problems.
- Good applications are clear, concise, and directly tackle open questions in making AI safe and controllable.
Why It Matters
Funding the right AI safety research is critical to ensuring AI systems remain under human control as they grow more powerful.