Part 1 | Foundations — Why local files, and how to use them safely

Chapter 2. Where the line sits on privacy and security [most important]

3 / 16

Key points

What actually stops therapists is not difficulty, but vague unease
There are only three principles to hold — anonymise, turn training off, keep it local
The aim is not to remove the unease but to convert it into rules

Please do not skip this chapter

The biggest reason therapists stall with generative AI is not technical difficulty. It is

vague unease

"Is it all right to put patient information in?"
"Could what I enter be extracted, or used for training?"
"Am I breaching my facility's rules?"

Healthcare, long-term care and welfare carry a high duty of confidentiality, so this unease is correct. Which is exactly why it should not be left vague, but converted into explicit rules.

The unease does not vanish by magic. But once you can see the line — how far is safe — you can step forward with confidence.

Three principles cover ninety per cent of it

Principle 1: never hand over anything that identifies a person (anonymise and mask)

The most certain and most important rule. Before agonising over what may be handed over, delete the identifying parts before you hand anything over. That alone removes most of the risk.

Say you want a patient record summarised:

name → replace with "Mr A", "man in his sixties"
date of birth, address, phone number, ID → delete
employer or a distinctive anecdote → generalise ("works at ◯◯ Construction in Tokyo" → "does physical labour")

In other words:

get it into a state where a stranger reading the text could not tell who it is about — then hand it over.

Therapists already do this every time they write a case report. You are simply carrying that instinct across. You can even have the AI help with the anonymising itself (see the assistant prompt below).

Principle 2: choose tools and settings where your data is not used for training

The worry that "what I typed will be trained on and surface in someone else's answer" is controllable if you choose the tool and the setting correctly.

The state of the main free tools, as of 2026:

ChatGPT (free): under Settings → Data Controls, you can turn off use of your input for model improvement. Check and change this setting first. Doing important work in temporary chat mode is also effective.
Claude (free): on consumer plans, input is by default not used for model training (terms can change, so check the current terms when you use it).
Gemini Notebook (free): uploaded material and chats are stated not to be used for training Gemini Notebook's models. There is a caveat, though: if you submit feedback, Google staff may review that content, including your input and uploads. "Not used for training" does not mean "no human ever sees it".

The practical rule that follows is simple.

Use free tools with training turned off. But genuinely confidential material (raw patient data and the like) should not go to the cloud at all. Hand over anonymised extracts only.

Principle 3: the safest thing is to keep it on your own machine

If it never goes to the cloud, it never leaves. That is the strongest security there is.

Obsidian, covered later in this guide, keeps every file only on your own computer. Which lets you split things:

grow the knowledge base itself locally, safely
when you want to consult an AI, copy across an anonymised portion only

Not "show the AI everything", but "cut out the part that may be shown, and hand over that". This is where safety and usefulness meet for a therapist.

Concrete rules for the workplace

Translating the principles into the language of practice.

1. Check your organisation's rules first. Hospitals, clinics and facilities almost always have rules on handling information. Generative AI at work may be explicitly permitted, or explicitly forbidden. Your organisation's rules come before your own judgement. Check them first.

2. Hold "sensitive personal information" in mind. Information about health and medical care sits in a category requiring particular care. Even with the name removed, a rare condition, a distinctive incident and a small locality can combine to identify someone. Anonymising is not only removing the name; it extends to breaking up the combinations that lead to identification.

3. Do not put raw patient information in the cloud. Templating an assessment report, or improving an explanation, works perfectly well with a fictional case or an anonymised extract. There is almost never a need to upload real raw data.

4. Do not take the output at face value (clinical judgement stays with the human). AI will write things that are not true, and write them plausibly (hallucination). It is excellent as a first draft for summaries and drafting, but the accuracy of the final clinical judgement and of any medical statement must be confirmed by you, the professional. This is the same distance Foundations kept returning to.

⚖️ The detail of data protection law and of handling medical information belongs to law and guidelines. This guide sets out general practical thinking; final decisions should rest on your organisation's rules, the relevant law, and, where needed, advice from a specialist.

Never using it is also a risk

Worth repeating from Foundations.

"It looks risky, so I will never use it."

carries its own risk over the medium term, because the productivity gap with colleagues and other facilities who do use it may widen considerably over a few years.

What matters is not zero-or-everything, but

starting where there is no personal information at all

Papers, published guidelines, invented cases, your own notes — there is no shortage of material that would be fine outside. Start there and there is nothing to fear.

The breath before you press send — a safety checklist

Five things, every time, before a work file goes to an AI.

Have I removed anything that identifies a person — names, IDs, contact details?
Have I checked whether a combination of condition, locality and occupation could identify someone?
Is the tool set so that input is not used for training?
Does this breach my organisation's rules?
Am I going to verify the output against my own professional knowledge?

Once this is a habit, there is nothing left to be afraid of.

A prompt to try (anonymisation assistant)

In the case note below, replace every piece of personal information, and anything that could identify an individual, with a mask (◯◯) or a generalised expression. Keep the medical content, the course of events and the meaning of the assessments intact as far as possible. At the end, list every substitution you made.

[Case note] (paste the text you want anonymised here)

Save this as a GPT or Gem once, and you have a safe first pass available every time.

Chapter 2 summary

What stops people is vague unease; convert it into rules and you can move forward
Three principles — anonymise, turn training off, keep it local
Your organisation's rules come first; raw patient data never goes to the cloud, only anonymised extracts

← Back to contents