Where to draw the line

How cheaply results can be checked tells us what could be handed to agents. It does not tell us what should be. Two kinds of decisions stay with people: judgment, about people and about science, and anything with consequences outside the lab.

Do not delegate

• Decisions about people. Grading, admissions, hiring, contract extensions, promotions, reference letters and career advice. Agents may draft text or collect material. People decide and sign.
• Judgments in peer review. The opinion in a review is the reviewer's. Do not upload confidential manuscripts or proposals unless the funder's rules and the tool's terms allow it.
• Scientific claims and authorship. An agent is never an author. For every claim in a paper, a person has checked it and is responsible for it.
• Clinical decisions. No diagnosis or treatment output without validation and accreditation. One group built a diagnostic sequencing pipeline with agents in one summer. It still needs validation and accreditation before it is used for patients.
• Actions that cannot be undone, without a person's approval. Deleting data, sending emails, and submitting forms, papers or payments.

Take care

• Personal and patient data. Data protection law, data-use agreements and ethics approvals apply to agents in the same way as to people. Use enterprise terms, keep data local where required, and log where data was sent.
• Students' learning. Students need certain skills to judge agent output later, and agents can let them skip learning those skills. Some tasks should be done by hand first.
• Sameness. Slides, reviews and possibly research directions are becoming more alike.
• Costs. Tokens, compute, energy, and dependence on a few vendors. Budget for these costs. Use the strongest tool available, and review that choice often. Keep a working fallback on open-weight models that run on local or institutional hardware. The same fallback serves data that must not leave the institute.
• E-mail and calendars. Some people now give agents access to them. No message goes out without a person approving it. See Establish guardrails.
• The evidence behind these recipes. The first recipes come from one research centre, the Helmholtz Munich Computational Health Center. They are based on what people reported, and people who already use agents a lot are overrepresented. The centre works mostly computationally, and its wet-lab groups use agents less. Test them before relying on them.

The related practice on aiforscience.eu: State human-in-the-loop requirements.