Guidelines are how a field turns good intentions into practice that others can check. With AI, the stakes are unusual: the same tool can draft a discharge leaflet, screen ten thousand abstracts or flag a deteriorating patient, and its errors are often invisible to the person using it. Guidelines tell us what to check, what to report and who stays responsible.
Five reasons to use them
- Accountability stays with people. The frameworks agree that responsibility cannot be handed to a machine. Authors remain accountable for AI-assisted work, and AI cannot be listed as an author (ICMJE, COPE). Health professionals remain responsible for decisions that AI informs (WHO).
- Transparency makes work checkable. Stating which model and version was used, for what, and how its output was checked lets editors, readers and patients judge the result (GAMER, CHART, ICMJE).
- AI is evaluated like any other intervention. A tool that changes care needs evidence in stages, from early clinical evaluation to randomised trials, with harms reported (DECIDE-AI, SPIRIT-AI, CONSORT-AI).
- Fairness is measured, not assumed. A model can perform worse for groups that were thin in its training data, including people who do not speak English. Good guidance asks for results by subgroup (FUTURE-AI, TRIPOD+AI).
- Confidentiality comes first. Pasting identifiable patient data or an unpublished manuscript into a public AI tool can breach confidentiality. ICMJE notes that confidentiality may rule out uploading a manuscript to AI tools that cannot assure it.
Why nursing needs its own voice
Nurses are often the people who act on AI output at the bedside, and the last human check before it reaches a patient. Where the law requires human oversight of high-risk AI, as the EU AI Act does, that overseer will often be a nurse. Using these guidelines well in nursing research and education, and citing them, helps make that last check a real one.
Where things are heading
Three shifts are under way. Voluntary principles written around 2019 to 2021 are turning into binding rules: the EU AI Act, and journal policies that now ask authors, reviewers and editors to disclose their use of AI. General-purpose large language models are being used for tasks no guideline anticipated, so guideline groups publish faster and more of them keep their documents “living”. And evaluation is moving beyond accuracy on a test set toward effects in real care, the interaction between people and AI, and equity. Expect this library to change every few months.
What changes for education and research
Writing, editing and translating text are becoming almost free. That does not make education or publication meaningless; it moves their value. What becomes scarce is what AI cannot supply on its own: verified data, sound methods, honest reporting, clinical judgement, and a named person who takes responsibility. Health-sciences education will have to teach these directly, and assess them in ways that AI cannot fake. Research and publishing will lean more on checks of provenance and integrity, some of them automated, but human accountability for what is claimed remains the anchor in every guideline we have read.