Can NHS staff use ChatGPT? Which patient details to take out

Can NHS staff use ChatGPT? The answer depends on two things you can check before lunch: what your trust’s AI policy says about that tool, and whether the text you’re about to paste still points to a patient. National guidance won’t give you a yes or no for a named product.

This guide follows that order, from the policy to the letter itself, with a checker for the NHS number and the other identifiers that sit in a clinic letter’s header. If you run a GP surgery, our page for GP practices shows the same work on a practice computer.

What is the NHS policy on AI? Four layers, one local answer

NHS England Digital publishes information governance guidance on artificial intelligence for patients, for health and care professionals and for IG professionals. The page for professionals says their IG lead, data protection officer and Caldicott Guardian should be involved in any decision to implement AI. It doesn’t mention ChatGPT by name.

So the answer to “can I?” comes from the bottom layer of the table below. A national principle helps you argue a case; a trust policy that names the tool is the document worth asking for. If there’s none yet, an AI use policy template gives IG something to edit.

A real trust AI policy: ChatGPT for drafting, Copilot for patient data

Rotherham Doncaster and South Humber NHS Foundation Trust (RDaSH) publishes its AI policy, version 1.1, approved on 18 November 2025. Section 5.5 lets staff consider ChatGPT on trust-issued devices for drafting documents, generating ideas and summarising information. The same section bars it from processing patient-identifiable information.

Copilot is treated differently. Section 11.1.1 allows personal or sensitive patient information into Copilot but not into public AI tools, and section 11.1.5 rules Copilot out of clinical decision-making and direct patient care. What a Microsoft tenancy changes is set out in what Copilot does with confidential material.

“Trust-issued device” sounds like a safeguard, and it isn’t one for the prompt: the laptop belongs to the trust, but the text still lands with the chatbot’s provider. Where OpenAI keeps a UK user’s chats is covered in what ChatGPT does with your data.

The pressure behind these policies is measurable. In a Nuffield Trust and RCGP survey of 2,108 GPs, published on 3 December 2025, 28% said they used AI tools in clinical practice, and 11% used tools they had obtained themselves.

Is medical information covered by GDPR? Yes, and confidentiality too

Article 9(1) of UK GDPR prohibits processing data concerning health. The condition clinicians lean on, Article 9(2)(h), covers diagnosis, treatment and managing health services, and Article 9(3) adds that the data must be handled by or under a professional bound by secrecy. A personal chatbot account sits outside that circle.

Health data turns up well beyond the clinic, in fit notes and sickness logs, and we cover that HR side separately in special category data and AI. Here the extra layer is the common law duty of confidentiality, which binds NHS staff whatever UK GDPR allows.

NHS England’s guidance for IG professionals, last edited on 20 May 2026, says consent may be implied under that duty when information is used for individual care. Skip the national data opt-out in this debate: it covers research and planning in England, not a letter written for one patient’s treatment.

None of this comes from an AI statute. A surgery in Hull or Swansea answers to data protection law and confidentiality, and the EU AI Act only reaches it in narrow cases.

The eight Caldicott Principles, against a paste into ChatGPT

The National Data Guardian published the current Caldicott Principles on 8 December 2020, and GOV.UK marks them as applying to England. There are eight. Some training pages still teach seven, from before the eighth was added; trust the GOV.UK page over those.

Principle 2 settles most chatbot cases: use confidential information only when necessary, with alternatives where possible. A plain English rewrite of a heart failure letter works without the patient’s name, so the name has no business in the prompt. The same test applies to a patient list kept in Excel, column by column.

Principle 4, strict need-to-know, is the one people forget extends past the building. Whoever can read the prompt at the provider’s end counts, and Google’s own account of human review in Gemini shows why that isn’t hypothetical.

Which AI does the NHS use? The route to an approved tool

Approved tools differ by organisation; the route to approval doesn’t. NHS England’s guidance for IG professionals says a DPIA must legally be completed before AI-based technologies are put in, and that health and care organisations should make sure they are the controller or joint controller when they contract with technology providers.

The ICO’s DPIA guidance gives “a hospital (but not an individual doctor) processing patient data” as large-scale processing, counts patients as vulnerable and names AI as innovative technology. A trust chatbot ticks all three. If nobody can point you to its DPIA, treat the tool as unapproved.

Trusts that build on a model API have a choice the front-line user never sees: host a model themselves or strip identifiers from each request on the way out, compared in private LLM or privacy proxy.

The second route usually runs through an AI gateway, which sits between staff and the model and can keep a log of each request.

NHS number, CHI number and the other identifiers in a clinic letter

The NHS number is ten digits, often printed three, three and four, and used in England, Wales and the Isle of Man, according to NHS England Digital. The tenth digit is a check digit calculated from the other nine, which lets systems reject a mistyped number.

Scotland uses the CHI number, which gives more away. Public Health Scotland documents its first six digits as the date of birth, the ninth as sex and the tenth as a check digit. Delete the DOB line and the birth date survives, as on a UK driver number.

Northern Ireland issues its own ten-digit Health and Care number instead. Paste a letter into the box to see which fixed-format identifiers it carries. The example is invented, and its NHS number deliberately fails the check digit.

The checker stops where the hard part starts. The ICO’s guidance on indirect identification counts “a combination of significant criteria (eg age, occupation, place of residence)” as enough to single someone out, and in the example the daughter at the village pharmacy does exactly that.

Files bring their own problems. A letter exported to PDF needs real redaction rather than black boxes, and a Word document can name its author in its metadata.

What constitutes a data breach in the NHS after a paste?

A named patient letter in a chatbot nobody approved has left the organisation: report it to your IG lead or DPO, with the tool, account, text and time. Whether it legally counts as a breach can turn on the contract, as set out in is putting client data in ChatGPT a breach.

Report it the same day. Deleting the chat first and saying nothing is the worst order to do things in: IG needs the text to judge the risk, and deleting ChatGPT history recalls nothing the provider has already received.

Article 33 of UK GDPR asks the controller to notify the Information Commission, the ICO’s successor since 30 September 2026, without undue delay and within 72 hours where feasible, unless the breach is unlikely to risk people’s rights and freedoms. Your report is what starts that clock.

Using ChatGPT on NHS letters without the patient in them

Strip the header before the paste and the prompt stops carrying the patient, which is what every layer above protects. I wouldn’t paste a header into any chatbot, approved or not, when the task is rewriting a paragraph. Social care teams face the same letters, as our page for councils shows.

Nonimo does that step on the computer itself: it swaps the names, dates and numbers it recognises for tags before you paste, then puts the details back into the reply. Our UK home page shows it working on a clinic letter on the practice’s own machine.

What leaves the computer when Nonimo runs is in our privacy notice.

Plans for a practice are on the pricing page.

This page explains the general rules; your trust’s IG lead or DPO has the last word on a specific tool.

Sources

• UK GDPR, Article 9 (legislation.gov.uk): 9(1) prohibition on health data; 9(2)(h) health and care condition; 9(3) professional secrecy, amended from 20 August 2025.
• UK GDPR, Article 33 (legislation.gov.uk): notification to the Commission within 72 hours where feasible.
• NHS England Digital, Artificial intelligence guidance for IG professionals: DPIA before implementation, 9(2)(h) for direct care, implied consent for individual care, controller or joint controller; last edited 20 May 2026.
• RDaSH, Artificial intelligence (AI) policy: version 1.1, approved 18 November 2025; sections 5.5, 11.1.1 and 11.1.5.
• National Data Guardian, The Caldicott Principles (GOV.UK): eight principles, 8 December 2020, applying to England.
• DCMS, Information Commission succeeds the ICO (GOV.UK): the Information Commission took over on 30 September 2026.
• ICO, When do we need to do a DPIA?: hospitals processing patient data as large scale; patients as vulnerable; AI as innovative technology.
• ICO, Can we identify an individual indirectly?: age, occupation and place of residence together can single someone out.
• Public Health Scotland, CHI operations: date of birth in the first six digits, sex in the ninth, check digit; from August 2026 new CHI numbers may use a Mod-10 check digit instead of Mod-11.
• Nuffield Trust and RCGP, How are GPs using AI?: 2,108 GPs, 28% using AI, 11% with tools they obtained themselves; 3 December 2025.

Accessed 8 November 2026.