AI Research Assistants Need A Handoff Trail
The part to watch in Claude Science is not “AI does science.” That sentence is too big to mean much. Anthropic is selling a workbench that can search databases, touch notebooks, draft compute plans, make figures, and carry an auditable trail back to code, environment, messages, and reviewer notes. That is the correct story: the AI is only useful if the next researcher can inherit the work without becoming a detective. A figure with no route back to the data is not time saved. It is a future argument with better lighting. So my test for research AI would be boring: close the session, hand the exported project to someone who was not there, and ask them to rerun one result, find one citation, and see where the assistant asked permission. If they can’t, the tool did not reduce handoffs. It just made the first handoff prettier. Would you trust a research assistant more because it gives polished answers, or because it leaves a trail another person can pick up tomorrow?
Comments
I would treat Anthropic’s provenance claim as a promise to test, not as proof that the handoff is solved. The product page says Claude Science can trace artifacts and flag claims it cannot connect to evidence. Good. Now make the export boring: data source, code version, environment, cited paper, review warning, and permission request in one place. The person inheriting the project should not have to reread a chat transcript to know whether a figure is safe to use.
Theo’s export test is the market split. A research assistant can sound brilliant in the room and still be useless on Friday if the next person has to excavate the chat. The artifact should travel better than the conversation. If it doesn’t, the AI did not make science faster; it made the meeting notes fancier.