TASTE Benchmark: Frontier Models Score Near Chance on AI Safety Research Judgment
Anthropic published TASTE (The AI Safety Taste Evaluation) on August 28, a benchmark designed to measure whether frontier AI models can judge pairs of AI safety research proposals the same way experienced human researchers do.
The finding is striking: the best-performing model, Claude Fable 5, agrees with expert human researchers only 60% of the time. Claude Opus 5 and GPT-5.6 Sol — both at or near the top of every agentic coding leaderboard — score close to the 50% chance baseline, meaning their judgment on safety research quality is effectively random relative to what expert researchers prefer.
Human researchers, by contrast, agree with the benchmark’s labels 77% of the time after a structured discussion process.
Why This Benchmark Exists
The motivation is practical. If AI systems are to accelerate AI safety research — a scenario that may become necessary if AI development outpaces human oversight capacity — those systems need to evaluate research quality on dimensions that resist automated verification. Judging safety research proposals is precisely the kind of task with no clear reward signal: unlike writing code, there is no test suite to run.
TASTE was built from 93 human-written research proposals developed in Anthropic’s Fellows Program. Researchers used Claude Opus 4.6 to reverse-engineer motivating prompts for each proposal, then generated new proposals from those prompts using a diversified scaffold. Four AI safety researchers rated each set of proposals, discussed disagreements in pairs, and revised their scores.
The key design insight: a discussion stage combined with filtering for self-reported “strong” confidence raised estimated human agreement by 15 percentage points, from 53% pre-discussion to 68%. The final benchmark contains 92 pairwise comparisons with 77% estimated human agreement — high enough to treat it as a reliable signal.
Key Results
| Model | TASTE Agreement |
|---|---|
| Human researchers | 77% |
| Claude Fable 5 | 60% |
| Claude Opus 5 | ~50% (near chance) |
| GPT-5.6 Sol | ~50% (near chance) |
The benchmark report notes that nearly all models perform within two standard deviations of chance, with per-model confidence intervals spanning roughly ±10 percentage points given the 92-pair dataset size.
The Fable 5 result is meaningful: 10 percentage points above chance on a hard-to-verify judgment task. But it also illustrates the gap. Fable 5 sits atop SWE-bench Verified at 95.0%, can write production code autonomously, and reasons through multi-step agent tasks at near-human speed. On the question of which safety research direction is more important or tractable, it performs no better than a coin flip at the frontier.
What This Means
Two interpretations are worth holding simultaneously.
First, the capability gap is real and consequential. The models that are most capable at verifiable agentic tasks — SWE-bench, Terminal-Bench, tau-bench — show no generalisation to qualitative research judgment. That matters because automated AI safety research requires exactly this generalisation, and TASTE suggests it is not happening at current capability levels.
Second, the benchmark is small and hard. 92 pairs, each requiring genuine domain expertise in AI safety. Human agreement itself is only 77%, meaning even expert researchers often disagree. The near-chance scores may partly reflect the benchmark’s genuine difficulty rather than a categorical model limitation. Fable 5’s 60% suggests some signal is being picked up.
The paper, co-authored by Hasan Baig (Anthropic Fellows Program), Hailey Joren, and Joe Benton, does not frame the result as a failure. It frames it as a baseline. The benchmark is open; the paper is public. The implicit message: researchers who want to automate safety research need this signal.
For now, the task of judging which AI safety research directions are worth pursuing remains a human job — one that frontier models cannot reliably assist with.