AI privacy best practices: a routine for work documents and text
Most privacy mistakes with AI happen in a hurry. You need a reply drafted or a report summarised, and pasting the whole thing is the fastest way to ask. This routine takes a few minutes and works with any service, with or without Anonymizator.
Before: ask whether it belongs there
Start with permission. France's data protection authority, CNIL, advises users to submit only information they are authorised to share and warns against putting confidential information into consumer services. If the document came to you under a contract, a professional duty or your organisation's rules, check those first. No preparation step gives you a permission you didn't have. CNIL's guidance on generative AI
Then use the account your organisation approves. Terms depend on the plan as well as the provider. OpenAI and Anthropic both say they don't train on inputs from their covered business and commercial products by default, with exceptions such as feedback you choose to send. A paid personal plan isn't the same as an organisational one. OpenAI enterprise privacy, Anthropic commercial policy
Before: give the task less to work with
The European Commission describes data minimisation as limiting personal data to what the purpose requires. In daily work that means naming the task and asking which details help with it. To tighten one paragraph, send that paragraph. To summarise a contract, the parties' bank details probably add nothing. European Commission on the GDPR principles
Files need the same treatment as text, and more of it, because you see less of what's inside them.
Give AI the task, not the whole file
Before: replace what has to stay
Sometimes the people and numbers are part of the task. A summary of a meeting has to say who agreed to what. Replacing each name with a placeholder keeps the sentence working: the AI can still tell who did what without learning who they are.
Then read what's left. The Commission explains that details which identify a person when combined are personal data, and that pseudonymised information which can be used to re-identify someone stays personal data. A job title, a date and a rare event can do the work of a name. European Commission on personal data
Researchers have tested this. In one study, language models worked out who was behind pseudonymous online accounts from what those accounts had written. In another, 240 US participants struggled to anticipate what a text gave away, and their rewrites often left those risks in place. Those results belong to the models and tasks studied, but they are a good reason to read for context as well as names. Lermen et al., USENIX Security 2026, Wang et al., CHI 2026
During: four questions a setting doesn't answer together
People often check one setting and consider it done. These are separate questions, and each service answers them differently:
- Is it used for training? In ChatGPT, turning off "Improve the model for everyone" covers new conversations under that setting. Anthropic's consumer policy depends on your permission and on specific circumstances such as safety review. OpenAI data controls, Anthropic consumer policy
- How long is it kept? A temporary chat in ChatGPT isn't used to improve models, but OpenAI may keep a copy for up to 30 days for safety. OpenAI temporary chat
- Who can access it? A no-training default says nothing about provider access or your workspace's administrators. Read the terms for your plan.
- Does it go anywhere else? Data sent through a GPT action falls under the recipient's policy and may be kept longer. Connected apps deserve the same question.
Check your AI privacy settings
After: close what you opened
Anyone with the link can view a conversation shared from a personal ChatGPT account, and deleting the link doesn't remove copies recipients have already saved. Check what a conversation contains before you share it, and go through your old links now and then. OpenAI on shared conversations
Archiving keeps a conversation. Deleting one in ChatGPT removes it from your account view and schedules permanent deletion within 30 days, with documented exceptions. Files you saved separately need their own clean-up. OpenAI on deleting chats
Clean up when the task is done
The routine as a checklist
- Am I allowed to share this, and with this service?
- Am I using the approved account?
- What does the task need? Have I cut the rest?
- Are the names, numbers and references that must stay replaced?
- Have I read what's left for details that identify someone?
- Do I know the answers on training, retention, access and onward sharing?
- Is the list that links placeholders to people outside the conversation?
- Afterwards: are the chat, shared links, uploads and connected apps cleaned up?
Where Anonymizator fits
Anonymizator handles the replacing step. The web app prepares Word documents, Excel workbooks and text files, one file or the whole folder, and saves protected copies you can review. The Chrome and Edge extension does this for text you copy or select. Reveal brings the original details back into the answer.
It doesn't choose your service, find every private detail or decide whether you may share something. By default the result is reversible pseudonymization, and it isn't anonymous.