ChatGPT privacy: what happens to your data, and what not to share

What ChatGPT keeps and what OpenAI does with it

According to the documentation OpenAI publishes, four principles shape how your conversations are handled.

• History is kept until you delete it. A deleted conversation is, according to OpenAI, erased from its systems within a stated period of 30 days, unless a legal obligation or a security need applies.
• On consumer plans, your content may be used to improve the models. A setting in Data Controls, "Improve the model for everyone", lets you turn this off for your future conversations.
• Temporary chats do not appear in your history and are not used for training; OpenAI says it keeps them for up to 30 days to monitor abuse.
• Business and enterprise plans, as well as the API, do not use content to train models by default, and come with their own contractual commitments.

Two nuances apply. First, turning off training does not mean nothing is kept: conversations remain stored on OpenAI's servers for the stated period and may be reviewed for safety reasons. Second, OpenAI states that legal proceedings or legal obligations may lead it to keep some data longer.

Regulators follow these questions closely. The Italian data protection authority temporarily restricted access to ChatGPT in Italy in 2023, and in May 2024 the European Data Protection Board published the report of its ChatGPT taskforce. Check your settings whenever you change plans, and reread OpenAI's documentation from time to time.

The real risks for a business document

For personal use, the risk depends mostly on what you agree to share about yourself: your health, your finances or your family. For a business document, the situation is different: the data are not yours. A client contract, a payslip or a medical report contains data about other people, and sometimes secrets the law requires you to keep.

Five risks deserve to be weighed before each upload.

1. Loss of control: once sent, the content is stored by OpenAI, possibly outside the European Union, under rules you do not control.
2. Reuse: on a consumer plan with the training setting on, your content may be used to improve the models.
3. Account access: a shared account, an unlocked computer or a compromised password give access to the entire history.
4. Shared links: a shared conversation link makes its content readable by anyone who receives it.
5. The legal framework: passing personal data to a service provider requires a legal basis and safeguards under the GDPR; disclosing information covered by professional secrecy is a criminal offense in many countries (in France, under article 226-13 of the Criminal Code).

None of these risks rules out using ChatGPT. They simply mean that what you send must be chosen. European data protection authorities, in EDPB Opinion 28/2024 on AI models, confirm that GDPR principles apply to these uses, and data minimisation comes first among them.

What not to share with ChatGPT, and how to use it anyway

Here is the information you should never paste as is into a conversation:

• passwords, access codes, API keys and login tokens;
• identification numbers: social security, identity document, tax number, IBAN, bank card number;
• health data, opinions, criminal convictions and the other special categories of the GDPR;
• the names, contact details and files of your clients, patients, employees or candidates;
• trade secrets: contracts under negotiation, unpublished figures, projects, internal source code.

To work on these documents anyway, three levers complement each other. The first is the account setting: turn off training, use temporary chats, or move to a business plan governed by a contract. The second is organization: a written rule stating what the team may send, and to which tool. The third acts on the content itself: removing the data before sending, which protects you even if a setting is wrong or the service's terms change.

That is the role of Biffer, the software we publish. On your computer, offline, it replaces names, contact details and numbers with stable aliases (PERSONNE_001, IBAN_001); ChatGPT works on a coherent text, and the encrypted mapping table, which stays with you, puts the real values back into its answer. The full list of information to look for is in the article personal data: definition and examples, and how to flag a sensitive document in the article how to mark a confidential document.