PPROMPT WORKBOOK Build a prompt
Data protection

Personal data in AI tools

This is the regime that causes the most real incidents, and it has applied the whole time.

This is information, not legal advice. Every rule below is cited to its source and reflects the position on 25 August 2026. This area is moving quickly, and a decision that matters for your organisation should be checked by someone with legal responsibility for it.

Data protection does not wait for the AI Act

Most attention goes to the AI Act and its dates. For the majority of organisations the nearer risk is data protection. The GDPR already applies, and it applies from the moment a customer name, a personnel matter or a health detail is pasted into an AI tool.

Reduces risk

  • A business agreement with data processing terms
  • Known retention period and storage location
  • Anonymising or pseudonymising before pasting
  • Rules people have actually seen
  • An inventory of which tools are in use

Increases risk

  • Free accounts where input may train models
  • Personal accounts used for work
  • Special category data pasted in directly
  • Integrations connected with no oversight
  • Assuming “no training” means “no storage”

Training and retention are two different questions

A provider stating that customer data is not used for model training tells you nothing about how long data is kept, where it sits, or who can see it during a support case. Retention varies between consumer products, business products, the API, individual features, regions, and any third-party tools bolted on top. Check the specific service and its data processing terms, not the vendor's marketing page.

What the European data protection regulators have said

The European Data Protection Board adopted an opinion in December 2024 that remains the central guidance in this area. Three points from it are worth knowing.

  1. A model is not automatically anonymous

    The fact that personal data was used in training does not mean the resulting model can be treated as anonymous. It has to be assessed case by case, including whether information can be extracted through queries.

  2. Legitimate interest can be a valid basis

    Developing and deploying AI can rest on legitimate interest, but only after a genuine necessity test and a balancing exercise against the rights of the people concerned.

  3. An unlawful basis can carry forward

    If a model was built on unlawfully processed personal data, that can affect the lawfulness of later deployment, unless the model has been properly anonymised.

In progress. The Board has published draft guidelines on anonymisation, open for consultation into autumn 2026. These are not final and should not be treated as settled law yet.

The practical rule

Before pasting customer names, employee details, health information or contract terms into an AI tool: use a business account rather than a consumer one, know where the provider stores and processes the data, and when in doubt strip the identifying details first. That single habit prevents more real incidents than any policy document.

Sources