Removing a name does not necessarily anonymise a clinical case. A distinctive combination of circumstances, dates, occupation, location or medical history may still identify someone, particularly when linked with other information. Before entering material into an AI tool, ask whether the task can be answered without transferring a real patient's narrative at all.
This UK-focused article is dated 10 October 2026. The ICO's anonymisation guidance treats identifiability as a contextual assessment, not a search-and-replace exercise. The worked example below is entirely fictional and illustrates a method for reducing unnecessary information, not a certification that any similarly edited real case is anonymous.
Direct identifiers are only the beginning
Names, addresses, NHS numbers and exact dates of birth are obvious starting points. Less obvious details can still narrow the field: an unusual occupation, a publicly reported event, a distinctive family relationship or a precise sequence of appointments.
The question is not simply whether an unfamiliar reader could identify the patient immediately. Consider what a recipient could reasonably link with the text, what other information is available and whether the same account repeatedly supplies related details. A clinician's own identity and workplace can also make an otherwise broad story more specific.
This is why anonymisation should be assessed in context. A case suitable for discussion inside an authorised clinical team is not automatically suitable for transfer to a separate service. The audience, purpose and available information have changed.
Pseudonymisation and anonymisation are not synonyms
Replacing a name with a patient code can reduce exposure, but it may leave a link to the person. The ICO's pseudonymisation guidance, checked on 10 October 2026, explains that pseudonymised information remains personal data for a party able to attribute it using additional information. The recipient's circumstances also matter to the assessment.
Calling someone "Patient A" does not resolve the issue if the prompt contains a highly distinctive history. Nor does changing one demographic detail reliably remove the identifying combination. It can also distort the clinical question, which creates a separate problem.
A safer starting point for an unapproved external learning workflow is often to extract the general question rather than attempt to disguise the whole patient. Where real patient-specific information is necessary, use an appropriately authorised service and process instead of trying to anonymise by intuition.
A worked transformation from narrative to question
Consider this wholly invented draft prompt:
A professional violin restorer who recently featured in a regional television programme attended my city-centre clinic just after returning from an overseas exhibition. A close relative works at the same hospital. The patient has persistent fatigue after an infection and is worried because an important public performance is approaching. Can you suggest what to do?
The prompt contains no name. It nevertheless carries occupation, publicity, location, timing, relationships and plans that could make a real person recognisable. Most of those details do not define the medical information the clinician needs.
A first revision identifies the learning objective: "I need to understand which features change the assessment of persistent fatigue after an infection in an adult." That is a question about clinical reasoning, not a request to manage a recognisable individual.
A more focused reference prompt might be:
In UK primary care, which history and examination findings should change the assessment of persistent fatigue following an infection, and which guidance addresses uncertainty and follow-up?
This does not contain an individual patient's narrative. The clinician must still assess the real patient separately, check applicable sources and recognise that the answer cannot determine care without the necessary clinical information.
Do not remove clinically important information just to make uploading easier
Data minimisation is not clinical distortion. If pregnancy, a relevant treatment, an examination finding or a particular timeline would materially change the question, removing it may produce an answer to the wrong problem.
The response should be to choose an appropriate workflow, not to force the real case into an apparently anonymous prompt. A general educational question can ask how that factor changes assessment without describing the particular person. If the task genuinely requires the detailed record, it belongs in an approved patient-data environment.
The ICO's data-minimisation principle concerns information adequate and relevant to the specified purpose, limited to what is necessary. It does not mean stripping out essential context until a clinical decision becomes unreliable. Define the purpose first, then decide what information and environment it requires.
Free text, images and attachments need separate attention
A document can retain identifying information outside its main paragraph. Headers, filenames, embedded screenshots, comments and visible record interfaces may reveal details that were removed from the body. Images can also contain identifying surroundings or metadata.
These are inspection questions, not a claim that every file has the same hidden fields or that every AI provider receives them all. The practical point is that reviewing only the visible sentence may be insufficient. Check the actual material being transmitted and the tool's processing behaviour.
Avoid uploading a whole letter simply to ask about one general concept. Extracting the question manually may remove the need to transmit the document at all. When an authorised upload is necessary, follow the organisation's process for preparing and checking it rather than improvising a redaction method.
Repeated prompts can rebuild a picture
A single prompt may be broad while a conversation becomes increasingly specific. Follow-up questions can add an occupation, age, investigation date, family circumstance and location in separate messages. The risk assessment should consider the complete exchange and any account-linked history, not only the newest paragraph.
A proposed self-check is to review the conversation as a whole before adding details. Ask whether the new information is necessary for a general answer, whether it changes identifiability and whether the service is authorised for that information. Starting a fresh tab does not, by itself, establish that earlier information is no longer associated with the account.
For teaching, construct a genuinely fictional case rather than preserving a recognisable real narrative with cosmetic changes. Label it as fictional and avoid claiming that its details represent an observed patient or outcome.
Apply the same standard to specialist reference tools
A healthcare-focused brand does not eliminate the need to consider information handling. The same task-first assessment should be applied to iatroX, informed by its current privacy policy. A question that can be answered without a real narrative should not acquire identifying details simply because a specialised tool is available.
A suitable reference workflow is to formulate the general question, inspect the source and apply the answer within the clinical setting where the patient information belongs. This preserves the useful distinction between learning about medicine and transferring a person's record.
The goal is not to make every prompt less informative. It is to ensure that the information sent, the purpose and the processing environment fit together, with uncertainty about identifiability treated as a reason for further assessment rather than a reason to assume anonymity.
Frequently asked questions
Is a clinical case anonymous once names and NHS numbers are removed?
Not necessarily. Indirect details and information available from other sources can still make the person identifiable.
Is replacing a name with a code enough to avoid data-protection requirements?
No; pseudonymised information may remain personal data. The ability to attribute it to an individual and the recipient's context need assessment.
What is a safer alternative for a general learning question?
Extract the clinical question without transferring the real patient's narrative. Where patient-specific information is essential, use an approved service and process rather than relying on informal redaction.