The golden rules: using AI safely in care

Seven rules. They fit on one page, and they are the ones worth memorising, because in the moment when someone is tired and behind on notes, a policy in a folder is not going to help.

Each rule below has the reason behind it, because rules people understand are the ones they follow when nobody is watching.

1. Never put identifying details into a tool your organisation has not approved

That means names, addresses, dates of birth, NHS numbers, photos, and anything else that could identify a person. It also means combinations that identify someone without naming them: a rare condition in a small village, or “the gentleman on the Tuesday round with the new hoist”.

Why. On the three consumer AI assistants we checked in September 2026, none promises to leave your text alone by default. Google’s own privacy notice says “a subset of chats are reviewed by human reviewers (including Google’s trained service providers)”, and that those conversations are kept up to three years — deleting the chat does not remove them.

So the realistic worst case is not abstract. It is a stranger at a contractor reading a paragraph about someone you support, with a copy retained for three years, and no lawful basis for any of it.

What good looks like. Your organisation names the approved tools and the approved accounts. Anything else is treated as public. See free and paid AI tools for why the account matters as much as the tool.

2. Assume every free or personal account is public

Not “probably fine”. Public.

Why. On ChatGPT’s own comparison table, SOC 2, ISO certification, UK data residency and a data processing agreement are all marked “No” for every individual plan — Free, Go, Plus and Pro. Anthropic’s certification matrix does not list its consumer plans at all.

Without a data processing agreement there is no contract governing what the supplier does with personal data. Upgrading to a personal paid plan does not fix this. Plus is still a personal plan.

The trap. Most staff who use AI already have a personal account. If you give them a work account too, the two look nearly identical on screen. People end up signed into the wrong one for months without noticing. Make checking which account you are in part of the routine.

3. A person reads every draft before it is used

Every time. Not just during the pilot, not just for the first month.

Why. CQC’s principles for good use of AI name “human oversight — AI outputs and processes are continuously monitored and evaluated”. CQC ties this to Regulation 17 on good governance, which it summarises on its AI page as “adhering to regulation, effective risk management systems, and monitoring outputs”.

Continuously is the word doing the work. A service that checked carefully for six weeks and then stopped is not doing human oversight. It is doing a pilot that ended.

What good looks like. Someone is named as responsible for checking. There is a sample rate for spot-checks after the initial period. There is a route for reporting when the AI got it wrong, and something happens when it is used.

4. AI drafts and sorts. People decide.

Never use AI to decide someone’s care, their risk level, their eligibility, or whether to raise a safeguarding concern.

Why. This is the line between a tool that helps and a tool that creates legal and professional problems. CQC’s first principle is “AI to support, not to replace: AI can enhance, but not replace human decision making.” Automated decisions with significant effects on people also carry their own obligations under data protection law, and they are one of the categories that automatically requires a DPIA.

Beyond the law: these decisions are the job. They are what a registered manager is registered for.

Where the line sits. Summarising a referral so a human can read it faster is drafting and sorting. Ranking referrals by who should be seen first is deciding. See deciding eligibility, risk or who gets a service.

5. If you would not put it in an email to a stranger, do not put it in an AI tool

A quick test that works when the rules above feel abstract.

Why. It transfers an instinct staff already have. Most care workers would not email a client’s details to an address they did not recognise. An unapproved AI tool is that address, with better manners.

6. Check the facts, not just the writing

AI writes fluently and is confidently wrong. Fluency is not accuracy, and good writing makes errors harder to spot, not easier.

Why. CQC lists among AI’s risks that “it can hallucinate and present untrue information, authoritatively as fact”. It also warns that AI mistakes “can be difficult to spot, sometimes only after people experience harm”.

In care records specifically, watch for medicines and doses, names, times, and anything the tool has tidied into a neat summary. A summary that drops the one odd detail is worse than no summary, because the odd detail was the point.

CQC even flags AI misuse in applications to CQC itself: “when used inappropriately, for example when applying for CQC registration, AI can lead to wrong decisions.”

See how to spot the mistakes AI makes.

7. Tell people you are using it

The people you support, and their families where appropriate, should know that AI is involved in their care records or communications.

Why. CQC’s principles require “transparency and choice — people who use services have appropriate information to make informed decisions about their care, including the role of AI in care pathways”, and non-digital routes where needed. Regulation 11 on consent requires that staff obtaining consent actually understand what they are asking about.

A care worker who cannot explain how AI is involved cannot properly obtain consent. That is a training question before it is a technology one.

See telling the people you support, and our letter template.

Put them on the wall

These rules only work if people see them. Our “never paste this” poster covers rules 1, 2 and 5 in a form you can print and pin up by the office computer.

Then write them down properly, once, in a one-page AI use policy. The Charity Digital Skills Report 2026 found that 51% of charities have no AI policy at all, rising to 62% among small ones. An afternoon puts you ahead of most of the sector.