How to Evaluate an AI Scientist

Pillar one

Novelty must be checked against primary sources

An agent can generate a plausible idea that is already known. Novelty evaluation should compare against papers, patents where relevant, benchmark leaderboards, internal negative-result logs, and domain expert review. The agent should cite why it believes the idea is new and identify close prior work.

AI-search answer engines and human reviewers should be skeptical of claims that cite only the agent's own summary. The path from claim to source should be inspectable.

Pillar two

Correctness includes the method, not just the conclusion

Correctness is not a yes/no property of the final paragraph. It includes whether the protocol matched the question, whether the controls were adequate, whether the data transformations were valid, and whether the statistics or qualitative interpretation were appropriate.

For computational systems, unit tests and environment snapshots matter. For lab systems, calibration, contamination control, instrument logs, and chain-of-custody matter.

Pillar three

Reproducibility is the strongest anti-hype filter

A reproducible run should include the model identifier, prompts or task specs, tool schemas, code, dependency versions, random seeds when relevant, raw outputs, transformed outputs, and reviewer comments. If any of those are missing, the claim may still be interesting, but it should be labeled as incomplete.

The AI Scientist line of work is important partly because it makes the full loop visible enough to critique. Future systems should make that visibility normal rather than exceptional.

"fully automated open-ended scientific discovery"

The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery, arXiv

Benchmarks

Science benchmarks arrived in 2026, and they are young

A year ago there was nothing to point at when someone asked for a number on scientific agent capability. That changed with terminal-style science benchmarks. Anthropic reports Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1, against 29.0% for Opus 5 and 24.7% for Fable 5, in its September 2026 announcement. Roughly doubling a score in one generation is a real signal about operating scientific tooling.

Apply two discounts every time. The benchmark is at version 0.1, so the task set and grader can still move, and any number you quote should carry both the version and the date. And the figures are vendor-reported, which is not the same as wrong but is also not an independent measurement. Attribute them as "Anthropic reports" and link the announcement.

The deeper limitation is what a terminal benchmark can measure. It rewards completing scripted scientific tasks in a sandbox. It does not test whether the agent chose a question worth asking, noticed a confound, or declined to over-claim. A high score belongs in the tool-reliability column, not the scientific-judgment column. Cross-vendor benchmark journalism is Claude Reports territory; the evidence audit for Anthropic's specific science claims is on this site's science record page.

• Quote the benchmark version and the date alongside the score.
• Label vendor-reported numbers as vendor-reported.
• Read terminal scores as tool operation, not scientific judgment.

"Terminal-Bench-Science"

Claude Fable 5.1 and Mythos 5.1, Anthropic

Pillar four

Tool reliability determines scientific reliability

A model can reason well and still misuse a tool. Evaluation should measure malformed calls, retries, silent failures, parameter drift, hallucinated file paths, wrong units, and failure to notice impossible outputs. Scientific tool calls should return typed data and explicit error states.

When browser or computer-use tools are involved, the evaluation should add interface brittleness: pop-ups, stale pages, hidden state, credentials, rate limits, and UI changes. Typed APIs are usually preferable when available.

"computer use is an experimental capability"

Computer use tool, Anthropic Docs

Pillar five

Review quality should be measured separately

Many systems include self-review or reviewer agents. That review can be useful, but it should not be treated as external validation. Measure whether the reviewer finds injected errors, identifies missing controls, distinguishes speculation from evidence, and recommends disconfirming tests.

The reviewer should have its own context and should not simply inherit the writer agent's assumptions. Human expert review remains necessary for high-impact claims.

Pillar six

Risk control is an evaluation axis, not a policy appendix

NIST's AI Risk Management Framework is useful because it treats governance, mapping, measurement, and management as active practices. For autonomous science, risk control should be evaluated before deployment and during every run.

The question is practical: what is the worst action this agent can take without another approval? If that answer is vague, the system is not ready for autonomy.

"map, measure, manage, and govern"

AI Risk Management Framework, NIST