Why protect personal data before AI?
A client email can become an AI prompt in one copy and paste. The names, account details and private history go with it.
Before you ask for a clearer reply, decide which of those details the task actually needs. Often, the person's name contributes nothing to the answer.
Why this matters now
Everyday AI tasks can involve client correspondence, employee information or financial records. A repeatable preparation step helps people decide what they intend to share before sending it.
Make the privacy decision before the text leaves your workflow. A useful habit needs to work on the ordinary Tuesday when someone is in a hurry.
What do OpenAI, Anthropic and DeepSeek say?
The rules depend on the service, account and settings. Paying for a personal subscription does not necessarily give you the terms of a business service.
OpenAI: Consumer content may be used to improve models, subject to the relevant controls. OpenAI says business offerings and its API are not used for training by default. Check which service your account actually uses. Consumer data use, business data policy
Anthropic: Its consumer policy describes model improvement when users allow it, with additional provisions for safety review and explicit opt-ins. Commercial inputs and outputs are not used for training by default, although submitted feedback and other opt-ins can be used. Consumer policy, commercial policy
DeepSeek: Its hosted-service policy says it collects prompts, may use personal data to improve its technology, and processes and stores personal data in China. It also describes training opt-out rights and tells users not to submit sensitive personal data. This policy concerns the services it covers, not every deployment of a DeepSeek model. DeepSeek privacy policy
Read the actual terms. A provider's nationality alone does not answer who receives your text, where it is processed or what controls apply.
A training setting answers only one question
You also need to know what is retained, who can access it and whether the service is approved for your work. For example, OpenAI says it may retain a copy of a Temporary Chat for up to 30 days for safety purposes. Temporary Chat guidance
French privacy regulator CNIL advises users to submit only information they are allowed to share, and warns against providing confidential information to consumer services. CNIL's guidance for users
A no-training setting is useful. You still decide whether the information belongs there.
Check your AI privacy settings
Remove what the task does not need
Before using work text with AI:
-
Check the service. Use the account and workflow your organisation permits.
-
Reduce the text. Send the relevant passage. Remove unrelated personal and confidential details.
-
Protect private IDs. Replace names, account details and private references that the task does not need.
-
Read the context. A rare diagnosis, exact event or distinctive job title can make someone recognisable even without a name.
-
Keep the list separate. If you use reversible replacements, do not put the list of placeholders and original details, or a file that holds it, into the AI conversation.
Asking an external AI to remove private data after you have pasted it there means it has already received that text.
Where Anonymizator fits
Anonymizator offers a web app for text and supported files, including batches from a folder, and a Chrome and Edge extension for copied text and page selections. Review the protected copies or text, add any missed IDs and choose what to share.
Hosted Protect sends text encrypted to Anonymizator, where it is decrypted for detection. The service does not retain the text or the detection results. Current IDs, placeholders and Reveal stay in your browser.
By default, this is reversible pseudonymization. It does not guarantee anonymity, find every private detail or establish that a particular disclosure is permitted. Pseudonymization reduces the link to a person while retaining a way to reconnect it. EDPB explanation