AI and personal data: what the model actually receives
« Does ChatGPT keep what I type to it? » The question comes up as soon as we start using an assistant for real: medical appointments, invoices, contracts, official mail. This guide explains what is actually sent to a language model, what remains, what the GDPR requires, and which habits really make a difference.
What a model receives when you write to it
A language model doesn't browse your computer: it receives text sent by the application you are using. This text contains more than just your last sentence.
- Your message, word for word, including punctuation and typos.
- A portion of the conversation history, to keep track of the context.
- The application's internal instructions (the "system prompt").
- The content of attached files: text extracted from a PDF, an image sent to a vision model.
Everything in this text leaves your device. What isn't in it cannot leak.
That's the key point: protection isn't handled "at the AI provider's end", it happens before sending, within the application building the request.
Training, retention, logging: three different things
| Question | What it means |
|---|
| Training | Is your content used to improve the model? On professional APIs, no by default; on consumer plans, it depends on the settings. |
| Retention | How long is the request stored? Usually a few days for abuse detection, then deleted. |
| Logging | Does the application save the content in its technical logs? This is the publisher's choice, independent of the AI provider. |
An application can thus be flawless on the provider side yet leak through its own logs. Conversely, an app that masks secrets before sending reduces the risk regardless of the provider.
What the GDPR says, in practice
- Minimization: only transmit data necessary for the expected response.
- Transparency: know which sub-processors are involved (AI provider, host, payment).
- Right of access and erasure: be able to view and delete what is stored about you.
- Storage limitation: defined retention periods, not indefinite storage "just in case".
- Security: encryption of credentials, user-restricted access to data.
For a personal assistant, minimization is the most effective lever: a credit card number never helps the model answer "reschedule my Tuesday appointment".
The habits that actually make a difference
- Never paste a full identifier when a part is enough: "the card ending in 42" rather than the sixteen digits.
- Describe a document instead of photographing it whenever possible.
- Check that the application allows you to view and delete what it has remembered about you.
- Prefer an application that masks sensitive data before sending over a direct access to the model.
- Read the privacy policy to identify sub-processors and retention periods.
This is exactly the model chosen by Happy Assistant: automatic filtering of secrets before calling the model, warning the user before sending, memory stored within the application and erasable memory by memory.
Frequently asked questions
Can an AI assistant be GDPR compliant?
Yes, provided it documents sub-processors, limits transmitted data, defines retention periods, and allows access and erasure.
Should I completely avoid AI assistants for personal matters?
No. The risk is greatly reduced as soon as the app filters sensitive data before sending and keeps memory on its side rather than with the model provider.
Is a photo riskier than text?
Yes. Text can be masked automatically, an image cannot: everything visible in the photo is transmitted, including anything lingering in the background.