When does an AI system count as a discoverer? Separating detection from validation in discoveries

Only attribute a scientific discovery to an AI system when you can show, per phase, what the system did and what humans did. Detection by Claude, for example, is a demonstrable contribution, but only after independent validation and human interpretation can you speak of a completed discovery.

On 23 September 2026 Anthropic reported that Claude agents helped identify a previously uncharacterized enzyme system during an AI-supported biology programme. That is a concrete occasion to sharpen an old question: at what moment does AI support turn into a contribution to a scientific discovery, and when does it not? OpenAI is often cited as a point of comparison, but in this case Claude is the AI system under examination.

What exactly did Anthropic report about Claude and the enzyme system?

According to Anthropic, Claude agents searched through roughly 210 million tokens spread across 200,000 sequences of reverse transcriptases. From these, Anthropic says the agents identified candidate systems and detected a combination of a reverse transcriptase, an adjacent partner gene and a repeat array. The researchers called that combination array-associated reverse transcriptases (ARTs).

Anthropic adds two things that bound the story. First: human scientists carried out the laboratory work. Second: the function of the system is still under investigation. Both facts are stated literally in the announcement and are crucial for attribution in discoveries (see the Anthropic source).

Into which phases should you split an AI system's contribution?

Attribution becomes verifiable as soon as you break the discovery process into identifiable phases and record, per phase, who or what made the contribution. In our assessment these are the auditable steps:

• Framing the question: who set the research direction and the search task?
• Detection: who or what found the pattern or anomaly in the data?
• Hypothesis: who formulated the testable assumption?
• Novelty assessment: who established that the candidate had not yet been described?
• Experimental design and execution: who designed and carried out the experiments?
• Validation and interpretation: who independently confirmed the result and interpreted its meaning?

In Claude's case, Anthropic's own account places the large-scale search and detection of the ART pattern with the model, while human researchers reviewed the candidate and performed the experiments. The supplied account indicates that humans provided the research direction, reviewed the candidate, conducted the experiments and are responsible for interpreting the incomplete findings. That distinction is the same logic behind verifiable AI verification per step: an outcome only counts when you can inspect the steps leading to it.

Why is Claude's detection not yet a completed discovery?

A discovery is more than a plausible candidate. This article's contention is that the ART find cannot yet be attributed to Claude as a completed discovery, for two reasons that follow directly from Anthropic's text.

• The biological function of the enzyme system remains under investigation. A detection without a validated function is an indication, not a conclusion.
• Human scientists performed the laboratory work, while the system’s function remains under investigation.

Searching autonomously and producing a plausible output is not the same as making a scientific discovery. In our assessment it is correct to give Claude credit for pattern recognition and candidate selection, but no unlimited credit for a result that still lacks independent, human-led validation. The same restraint fits testing vendor-reported model performance: a claim from the provider is a starting point for testing, not proof in itself.

What does Nature's Co-Scientist study add to this picture?

The Nature publication describes Co-Scientist as a multi-agent system based on Gemini that generates novel, testable hypotheses and reports experimental validation in areas including drug repurposing, target discovery and antimicrobial-resistance research. (see the Nature publication).

The study presents the system explicitly as scientist-in-the-loop: according to the paper, human experts define goals, review and select hypotheses, and conduct or supervise validation. Nature also states that the validations are preliminary and do not replace further preclinical or clinical research. That confirms the phased picture: AI can propose and prioritise hypotheses, while credit must also weigh human goal-setting, assessment and laboratory work.

There is one more separate boundary, from the law. The United States Patent and Trademark Office states in its revised guidance on AI-assisted inventions that only natural persons can be named as the inventor of a US patent, and that AI systems are tools in the invention process (USPTO guidance). That legal distinction means that acknowledging AI assistance is something other than naming an AI system as a formal human inventor. That reasoning is close to how you handle accounting for AI in professional practice: acknowledge the contribution, keep responsibility with the human.

How do you record AI contribution and human validation in practice?

The consequence not named in the sources themselves is, in our assessment, concrete for research groups: whoever wants to keep AI contributions demonstrable must explicitly separate detection from validation and interpretation in their own records. A workable set of fields looks like this:

1. the model version and the exact prompts or search tasks;
2. the size and provenance of the data searched;
3. the steps in which candidates were selected;
4. the human reviews and the choices made;
5. the experimental protocol and the raw results;
6. the final attribution decision per phase.

Such records can make the model’s documented contribution easier to audit. A separate privacy-preserving evidence layer could help research groups record these steps while limiting exposure of sensitive research data, subject to the safeguards and limitations of the system used. The same need for a separate evidence layer applies to demonstrating human oversight of AI decisions: without recorded steps a claim about the human's role remains unverifiable. For the ART case the simple conclusion holds: credit for detection, no credit for a discovery that still has to be confirmed.

Sources: The article draws on Anthropic's announcement about Claude, the Co-Scientist study in Nature and the revised inventorship guidance of the United States Patent and Trademark Office (USPTO).