Useful AI does not need a patient record
A doctor can spend an evening correcting a poorly structured document without making a single clinical decision. That is a better place to begin experimenting with AI than a real consultation. Start with work where the input can be public or invented and the output can be checked before anyone relies on it.
The following eleven workflows are proposals for bounded use, not promises about a particular product. Tools differ and their terms change. None requires handing identifiable patient information to an unapproved consumer service.
Research preparation: three tasks before you trust an answer
1. Build search vocabulary. Describe a general research question and ask for alternative terms and possible search combinations. Run those searches yourself in appropriate databases. Useful wording is the output; an AI-generated claim that the literature proves something is not.
2. Structure a reading note. Give the tool a public abstract or material you are permitted to use and request headings for population, methods and limitations. Check every entry against the original. An abstract-only note should stay labelled as such; it cannot establish details found only in the full study.
3. Prepare journal-club questions. Ask for methodological questions about a public paper, then select those that genuinely apply. This can help organise discussion without outsourcing appraisal. Check references independently rather than treating a plausible-looking citation as a real publication.
Teaching and communication: four drafts worth reviewing
4. Create a fictional teaching scenario. Specify that all details must be invented, and review the scenario for accuracy before use. Do not lightly disguise an actual patient encounter and call it synthetic. The exercise should serve a teaching objective rather than simulate clinical authority.
5. Reorganise presentation notes. Ask for a clearer sequence using your own non-confidential material. Keep the underlying argument and citations under your control. If the tool adds numbers, recommendations or examples that were not supplied, remove or verify them rather than rewarding the slide deck for looking complete.
6. Simplify general educational language. Use a public, reliable source and request a plain-language draft for a defined audience. A qualified reviewer should check meaning and omissions. Avoid personalised diagnosis, dosing or treatment instructions; simpler wording can still carry an incorrect or misleading claim.
7. Draft an administrative email. Use placeholders for names and no patient details. Appointment-policy wording or a request for an equipment quote can be useful test cases. Replace placeholders only within the approved communication workflow after checking tone, dates and commitments.
Operations: four experiments using invented information
8. Design a meeting agenda. Supply non-sensitive objectives for a staff discussion and ask for a timed sequence. Remove unnecessary items and identify the decisions humans need to make. Confidential personnel information should not be included merely to make the agenda more specific.
9. Prototype a spreadsheet formula. Use a tiny invented dataset to test calculations for supplies or staffing. Work out the expected answer independently, including an empty-cell example. A formula that runs without an error can still calculate the wrong result; testing matters more than fluent explanation.
10. Outline a purchasing comparison. Ask for questions about warranty, maintenance, installation and recurring costs. Verify actual specifications with the supplier. Do not let the model populate missing prices or present an invented feature as a reason to choose equipment.
11. Edit a professional biography. Provide accurate qualifications and a description of your work. Ask for clarity rather than stronger claims. Check that the draft has not promoted you to a specialist title, invented an award or implied an affiliation that does not exist.
Keep the boundary outside the prompt box
Removing a name is not the same as making a record safe to share. Dates, rare circumstances and other contextual details may still identify someone. For these experiments, public or fully fictional input is the cleaner boundary. Obtain the necessary institutional approval before considering any workflow involving real clinical information.
WHO’s 2024 guidance warns that generative systems can produce inaccurate or biased statements and encourage automation bias. It also identifies cybersecurity risks. That is a reason to assign a human reviewer and a defined purpose, not simply add “be accurate” to a prompt.
Measure editing time, including the mistakes
Try one workflow on a small task you already understand. Compare the total time for prompting, checking and revision with doing the task yourself. Keep a non-sensitive note of the tool version, input boundary and errors encountered so you can judge whether reuse is worthwhile.
Stop if verification becomes more demanding than the original work. AI earns a place in the workflow when it helps you produce a checked result with less effort. It does not earn trust by sounding confident, and it should not gain access to patient information just because an administrative experiment went well.