🕑 Estimated reading time: 4 minutes  |  Last updated: May 29, 2026

Using personal data to train or fine-tune AI is not automatically unlawful, but it must meet ordinary data protection rules: a lawful basis (often legitimate interests, after a careful balancing test), purpose compatibility, transparency, data minimisation, security and respect for objections. European regulators, in the EDPB’s December 2024 opinion on AI models, confirmed legitimate interests can work in some cases but stressed safeguards. Regulators have fined AI developers, including Italy’s €15 million fine against OpenAI in 2024. Nigeria’s Data Protection Act applies the same core principles.

Graphs of performance analytics on a laptop screen

Every business has a data treasure chest: customer support chats, emails, call recordings, purchase histories, loan applications. Generative AI makes that data look newly valuable. Why not fine-tune a model on five years of support tickets so it answers like your best agent? Or train a credit model on every repayment record?

Those ideas may be perfectly workable, but only if you respect data protection law. Here is how regulators are approaching the question.

The Starting Point: AI Is Just Another Processing Purpose

There is no special exemption for AI in the GDPR or the Nigeria Data Protection Act 2023 (NDPA). Training a model on personal data is “processing,” so the usual principles apply: lawfulness, fairness and transparency, purpose limitation, data minimisation, accuracy, storage limitation, security and accountability.

Choosing a Lawful Basis

Consent

Consent can work for small, opt-in projects, but it is fragile at scale: people can withdraw it at any time, and it is hard to remove one person’s data from a trained model.

Legitimate interests

Most organisations look to legitimate interests. In December 2024, the European Data Protection Board (EDPB) published an opinion on AI models confirming that legitimate interests can be a valid basis for developing and deploying AI, provided organisations pass a three-step test: a genuine, clearly articulated interest; necessity, meaning the processing is needed and less intrusive options would not work; and a balancing test showing individuals’ rights do not override the interest. Factors include people’s reasonable expectations, the source of the data, safeguards such as pseudonymisation, and whether people can easily object.

Purpose compatibility

If data was collected for one purpose, such as delivering orders, reusing it to train AI must be compatible with that original purpose or have its own lawful basis. Transparency about the new use is essential.

What Regulators Have Done

Is a Trained Model “Personal Data”?

The EDPB said AI models trained on personal data cannot automatically be considered anonymous. Whether a model is anonymous depends on how likely it is that personal data could be extracted from it, directly or through queries. If it is not anonymous, the model itself may be subject to data protection rules, which matters when you share or sell it.

Sensitive Data

Health, biometric, financial, religious and other sensitive data need extra justification under both the GDPR and NDPA. Training on such data, for example medical records or voice recordings used for identification, should only proceed with a strong legal basis and a data protection impact assessment.

The Nigerian Perspective

The NDPA requires a lawful basis, transparency and respect for data subject rights, including the right not to be subject to solely automated decisions with significant effects. The NDPC’s GAID 2025 expects DPIAs for high-risk processing and emerging technologies. Nigerian banks, fintechs and telcos building AI models on customer data should document their lawful basis, update privacy notices to mention AI training, and give customers a clear way to object.

Changes Being Discussed in Europe

The European Commission’s Digital Omnibus package, proposed in November 2025, included ideas to clarify when legitimate interests can be relied on for AI development. Those proposals are going through the EU legislative process separately from the AI Act timing changes agreed in May 2026, so organisations should follow current rules and guidance for now.

A Practical Checklist Before Training AI on Personal Data

  1. Define the purpose precisely: what will the model do?
  2. Minimise: remove fields you do not need, strip names and identifiers, and pseudonymise where possible.
  3. Exclude sensitive data unless you have a clear legal basis.
  4. Choose and document your lawful basis, including a legitimate interests assessment.
  5. Run a DPIA, considering risks such as memorisation and data leakage.
  6. Update privacy notices and offer an easy objection route.
  7. Check vendor terms: does your AI provider use your data to train its own models?
  8. Test the model for leakage of personal data in outputs.
  9. Plan for rights requests, including how you will handle objections and deletion.

Decide which customer data your AI actually needs

You can often use customer data to build useful AI, but not by quietly feeding everything into a model and hoping for the best. Treat AI training as a new purpose that needs its own justification, safeguards and transparency. Done properly, it can be both lawful and good for customers.

Frequently Asked Questions

Can I paste customer data into a public chatbot?

Generally not without careful assessment; it may involve disclosure to a third party and a cross-border transfer. Use approved enterprise tools with appropriate contracts.

Do I need consent to fine-tune a model on support tickets?

Not necessarily, but you need a lawful basis, transparency and safeguards. Legitimate interests may work if your assessment supports it.

Can people ask to be removed from a trained model?

They can object or request erasure; how you honour that technically depends on the model, which is why minimisation and planning matter.

Leave a Reply

Your email address will not be published. Required fields are marked *