When does an AI system count as a discoverer?

You must be able to demonstrate, for each phase of a scientific discovery process, which steps an AI system performed and which steps humans performed. Detection alone does not constitute a discovery; only after independent validation and human interpretation can you attribute a finding to AI assistance.

An analysis of 28 September 2026 of how to separate AI detection from completed scientific discovery argues that attribution becomes verifiable only when you break the discovery process into identifiable phases and record per phase who or what made the contribution. The concrete case is a report that Claude agents identified a previously uncharacterised enzyme system during an AI-supported biology programme. In our assessment, this development matters because research groups and organisations deploying AI in discovery workflows now face a practical obligation to document and defend their attribution claims, and the distinction between detection and discovery has both operational and legal consequences.

At what point does AI assistance become a discoverable contribution?

Searching autonomously through large datasets and producing a plausible candidate is not the same as making a scientific discovery. The reported case involved Claude agents searching through roughly 210 million tokens across 200,000 sequences to identify candidate enzyme systems. The agents detected a combination of a reverse transcriptase, an adjacent partner gene and a repeat array—a pattern the researchers termed array-associated reverse transcriptases (ARTs). That detection is a demonstrable contribution. However, the same announcement states that human scientists carried out the laboratory work and that the function of the system remains under investigation. Both facts are crucial for attribution. You should give the AI system credit for pattern recognition and candidate selection, but not unlimited credit for a result that still lacks independent, human-led validation. The same restraint applies when you assess vendor-reported model performance: a claim from the provider is a starting point for testing, not proof in itself.

What does the discovery process actually require?

A workable attribution framework separates the discovery into distinct phases, each with assigned responsibility:

1. Document research direction and objectives — record which human researchers defined the goals and scope of the search.
2. Record the model's detection work — specify which AI system performed the pattern search, which dataset it searched, and what candidates it identified.
3. Log human review and selection — document which humans reviewed the candidates and which they selected for further investigation.
4. Capture experimental validation — record which humans conducted or supervised laboratory work and what the results showed.
5. Assign interpretation responsibility — identify which humans are responsible for interpreting incomplete or preliminary findings and determining next steps.
6. Maintain a separate evidence layer — keep records in a privacy-preserving format that allows audit of the contribution chain without exposing sensitive research data.

These records make the model's documented contribution easier to audit and keep responsibility with the human decision-maker at each stage.

What does the law say about AI and inventorship?

The United States Patent and Trademark Office revised its guidance on AI-assisted inventions to state that only natural persons can be named as the inventor of a US patent, and that AI systems are tools in the invention process. That legal distinction means acknowledging AI assistance is something other than naming an AI system as a formal human inventor. The reasoning is close to how you handle accounting for AI in professional practice: acknowledge the contribution, keep responsibility with the human. For a research group seeking to publish or patent a finding, this means you must be able to show not only that the AI system detected something, but that humans performed the validation, interpretation and decision-making that turned detection into a discovery.

Why does the distinction matter for your own workflows?

Without explicit separation of detection from validation and interpretation in your own records, a claim about the AI's role remains unverifiable. The consequence is that whoever wants to keep AI contributions demonstrable must build this separation into their documentation from the start. A separate evidence layer—whether a structured database, a timestamped log or an audit trail—allows you to record which steps the AI performed and which steps humans performed, subject to the safeguards and limitations of the system you use. This is not optional if you intend to defend an attribution claim later. The simpler the record at the time, the stronger your position when you need to justify it.

Tooling can help you structure and store these records, but the professional judgement of which findings warrant further investigation, which candidates merit laboratory work, and which preliminary results constitute a genuine discovery remains yours alone. No system can automate that responsibility away.