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

Key points

• The security principles from the earlier parts carry the most weight here, where data is involved
• Hold anonymise / turn training off / keep it local — plus research ethics
• Responsibility begins the moment personal data is collected; draw the line before anything reaches an AI

Please do not skip this chapter

Chapter 2 of Working with Your Own Files said that what stops therapists is not technical difficulty but vague unease.

Here, working with data, that unease becomes most justified of all. What we handle here is information obtained from living people — patients, clients, survey respondents. A single questionnaire, a single assessment score, a single interview recording: part of someone's life lives in each of them.

So before any technical discussion, we go carefully over where the line sits. Once that settles, you can move forward with real confidence.

The three principles, carried forward and pressed harder

The three security principles from the earlier part do not change. If anything, because what you handle now is raw data, another notch of care is needed.

Principle 1: never hand over anything that identifies a person (anonymise and mask). Delete names, dates of birth, addresses, contact details and IDs. Generalise affiliations and distinctive incidents. Do rigorously, before handing anything to AI, what you already do routinely in a case report.

Principle 2: choose tools and settings where your data is not used for training. Turn on the "do not use my data for training" setting; do important work in temporary chat or on a business plan. And remember, as the earlier part noted, that "not used for training" does not mean "nobody sees it".

Principle 3: the safest thing is to keep it on your own machine. Keep raw data in Obsidian or on your own computer, and when you want to consult an AI, cut out an anonymised extract, or the aggregated figures only. Not "show everything", but "hand over only what may be shown". That idea is again the crux.

The pitfall particular to data — identification by combination

Here is the most important new perspective in this chapter.

Removing the name does not always stop a person being identified.

Take: "lives in ◯◯ city, forties, male, rare condition, former professional baseball player". No name anywhere. And yet an individual is identifiable. In the world of data, a combination of several attributes does the work of a name. Technically this is the risk of re-identification.

Medical and health information in particular is sensitive personal information, a category requiring especially careful handling. Anonymising is not only removing the name; it extends to checking

whether the combination of this information can be traced back to a person

Do a rare condition, a small locality and a distinctive career overlap here? Go over it with that eye before any data reaches an AI.

Research data has another line — ethics

The biggest difference between business data and research data is that research has a second heavy line: ethics.

Research involving human participants requires, as a matter of principle:

• informed consent (explaining the purpose, method and risks, and obtaining freely given agreement)
• ethics review (prior review by the ethics committee of your institution or partner)
• explicit scope of data use (never used beyond what was consented to)

And what matters decisively here is that

"may this be analysed by an AI?" is itself part of the consent and scope-of-use question

Uploading data collected for research to an external AI service the participants never envisaged may go beyond the scope of their consent. The principle — raw data never leaves, only anonymised and aggregated material goes to AI — applies here, and here above all.

⚖️ Data protection law, ethical guidelines for research involving human participants, and each institution's rules are specialist and frequently updated. This guide sets out general practical thinking; final decisions should rest on the rules of your organisation and partners, the relevant law and guidelines, and, where needed, your ethics committee or a specialist.

A safety checklist before working with data

Six things, every time, before data or any part of it 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, age and career could identify someone?
• (For research) Does this sit inside the scope of consent and the ethics approval?
• Is the tool set so that input is not used for training?
• Am I handing over an anonymised or aggregated extract rather than raw data?
• Will I verify the output against my own knowledge, and where needed with a statistician or supervisor?

Once this is a habit, working with data holds no fear.

A prompt to try (anonymisation assistant, data edition)

You are a research assistant who knows data protection well. Below is the list of field names (column headings) for data I intend to use in research. Identify (1) fields that could identify an individual on their own, and (2) fields that could lead to identification in combination. Then propose, for each, whether it should be deleted or generalised (e.g. date of birth → age band).

[Field list] (paste only the list of column names, containing no personal names)

Note that no actual data is handed over at this stage. It is a conversation about design, so it is safe to practise on.

Chapter 2 summary

• The three principles weigh most heavily here, where data is involved
• Removing the name is not enough; combinations identify people, and health data is sensitive
• Research adds the line of ethics — consent, review, scope. Hand over aggregated extracts, never raw data