Staying safe

Start with the material, not the tool. Decide its data lane before you paste, upload, record, or photograph anything.

Three data lanes

• Public: Material deliberately published for anyone to read. Check it for information about living people before using a hosted tool.
• Internal: Drafts, routine operations, unpublished research, and partner material. Use only an organisation-approved account and storage route.
• Identifiable: Anything that names a person or could identify one through details, combinations, voice, location, or context. Keep it local, minimise it, or do the work without AI.

Public source does not mean low risk. A published judgment, post, or photograph may still expose a living person.

Pick the route

Use a hosted consumer tool such as ChatGPT only for material allowed by your organisation and the exact account settings. Use an organisation-managed service for approved internal work. Use a verified local route, or no AI, for identifiable material.

These are starting rules, not a promise that any product is safe. Data handling varies by product, plan, workspace, setting, and date. Check the current official guidance for the account you will actually use. OpenAI documents consumer data controls in its Data Controls FAQ. Anthropic explains its current training choices in Claude’s privacy guidance. Google describes Gemini Apps data handling in the Gemini Apps Privacy Hub.

Five checks before every use

1. Data: Which lane applies, and could details combine to identify someone?
2. Input: Is every page, field, image, or recording necessary for the task?
3. Output: What facts, names, numbers, quotations, and omissions must be checked against the source?
4. Human: Who has the authority and knowledge to approve the result?
5. Record: What source, tool, account, date, settings, reviewer, and final decision must be kept?

Reduce identifying detail

Remove what the task does not need before the material reaches a tool.

1. Replace names with stable labels such as “Witness A”.
2. Generalise exact dates and locations when precision is unnecessary.
3. Remove identity, case, phone, vehicle, bank, and contact numbers.
4. Remove rare roles, injuries, events, or relationships that point to one person.
5. Replace very small counts when they reveal who is in the group.
6. Read the result as a whole. Separate harmless details may identify someone when combined.

Keep the re-identification key separate, access-controlled, and outside the AI workflow. If the remaining text can still identify someone, it remains in the identifiable lane.

If material has already left

1. Stop adding more material and report the incident through the organisation’s current process.
2. Record the minimum facts needed to assess it: what entered which product, account, and workspace, when, and by whom.
3. Use the product’s current controls for deletion, retention, sharing, and training where available. Do not assume they recall every copy.
4. Assess possible harm to the people, partners, cases, and operations in the material. Escalate urgent protection needs first.
5. Correct the workflow, access rules, and training. Encourage early reporting so the organisation can respond.

Create the missing rule with write your organisation’s AI rules in one hour, then practise it with public or synthetic material.