Anthropic is testing whether AI models can do a job usually reserved for senior scientists: judging the quality of AI safety research proposals. The experiment, detailed on Anthropic’s Alignment Science Blog, is called TASTE, and it asks a pointed question. If AI is going to help accelerate alignment research, can it also tell good research ideas from bad ones?
This matters more than it might sound. As AI systems start contributing to research, the bottleneck shifts from generating ideas to evaluating them. A model that can write a hundred research proposals is only useful if something can sort the promising ones from the dead ends. Human reviewers are slow, expensive, and limited in number. If models can share that load reliably, safety research speeds up. If they can’t, we get a flood of proposals nobody can properly vet.
📋 What the researchers did
The core setup is straightforward. Anthropic put AI models in the role of reviewer and had them assess AI safety research proposals, then checked how well those judgments held up. The goal is to measure “taste,” the harder-to-pin-down skill of recognizing which research directions are worth pursuing. It’s the difference between a model that can follow instructions and one that can exercise judgment.
That distinction is the whole point. Scoring an essay against a rubric is one thing. Deciding whether a research idea is novel, feasible, and actually worth the effort is a much fuzzier task, and it’s exactly the kind of work that separates a junior contributor from an experienced one.
💡 Why it matters for practitioners
If you build with or evaluate AI systems, the TASTE line of work points to a few practical takeaways:
- Model-as-judge is moving upstream. We’ve seen AI used to grade outputs for a while. Using it to evaluate research direction is a bigger ask, and it tests whether models have real judgment or just pattern-matching.
- Evaluation is becoming the constraint. As generation gets cheap, the ability to filter quality becomes the scarce resource. Whoever solves reliable AI evaluation unlocks the next stage of AI-assisted work.
- Trust needs measurement. You can’t just assume a model’s review is sound. Anthropic’s approach is to benchmark the judgments against a standard, which is the right instinct before anyone leans on model reviewers in production.
What stands out here is the honesty of the framing. Anthropic isn’t claiming models have great taste. It’s building a way to find out, which is the correct order of operations for a lab that keeps publishing on its own systems’ limits.
⚠️ The limits
Judging research is one of the hardest evaluation tasks out there, and results from a single benchmark shouldn’t be treated as the final word. A model that scores well on a curated set of proposals may still miss the truly unconventional idea, the one that looks wrong until it works. There’s also the deeper risk baked into the whole exercise: if we outsource judgment about safety research to the same class of systems we’re trying to make safe, the evaluation has to be watertight. Anthropic’s own framing treats this as an open question, not a solved one.
The broader trend is clear. Labs are racing to figure out how much of the research loop AI can handle, from generating hypotheses to running experiments to reviewing the work. TASTE targets the review step, which may be the most important one to get right. Bad generation wastes compute. Bad judgment sends the whole field in the wrong direction.
Expect more work like this as AI moves deeper into the research process. The question of whether models can be trusted to evaluate, and not just produce, is going to define how fast and how safely that shift happens. Full methodology and results are available in Anthropic’s original post on the Alignment Science Blog.