Introduction

Many businesses now hold years of support tickets, chat logs and customer emails, and sooner or later someone suggests using them to train an AI model. It sounds like an obvious idea, but the legal answer is not a simple yes or no. EU data protection law does not ban training AI. The difficulty is that you collected the data for a different reason, usually to deliver your service, and using it for something new has its own rules. Whether you may go ahead depends on decisions you need to make before you start, not after.

This article explains why the real issue is the change of purpose rather than the training itself, how to choose a legal basis, what “we anonymised it” really has to mean, what you owe your customers, and how proposed changes to the rules should affect your plans.

The Real Issue Is the Change of Purpose

The General Data Protection Regulation, usually called the GDPR, has no rule against training a model on personal data. Training is simply one more way of processing data. The problem is that you did not collect the data to train a model. You collected it to answer support requests, deliver orders or run a subscription.

Under the GDPR’s purpose limitation principle, data collected for one purpose may only be used for another if the new purpose is compatible with the original one. So the real question is not “may we train” but “is training compatible with the reason we collected this data”. The GDPR lists the main factors to weigh:

• the link between the original purpose and the new one,

• the context in which you collected the data, including your relationship with the customer,

• the nature of the data, especially whether any of it is sensitive,

• the possible consequences for the people concerned,

• the safeguards you put in place, such as encryption or pseudonymisation.

Carry out this assessment honestly and write it down. If you find that training is compatible, you do not need a new legal basis, but every other GDPR duty still applies.

There is one important trap. Suppose you originally collected the data on the basis of consent. The European Data Protection Board, known as the EDPB, which brings together the national data protection authorities, says that using it for a different purpose then needs fresh consent, unless another legal basis genuinely applies. It also warns that a business cannot quietly switch from consent to a different legal basis after the event. People gave consent for a specific purpose, and it does not stretch to cover a new one.

Choosing a Legal Basis: Consent or Legitimate Interest

Where you need a legal basis for training, there are usually two realistic options, and many businesses reach for the less practical one first.

The first is consent. It feels like the safest choice, because asking permission always feels safe. In practice it often creates the biggest problem. People can withdraw consent at any time, and withdrawing must be as easy as giving it. If thousands of people withdraw after your model has been trained, you face a hard engineering question: what does their withdrawal mean for a model that has already learned from their data? There may be no good answer.

The second is legitimate interest, the basis that most serious AI development relies on. It is not a simple checkbox. In its opinion on AI models, the EDPB made clear that it must be justified case by case through a three step test.

1. Your interest must be real, lawful and clearly described.

2. The processing must actually be necessary to achieve that interest. Ask whether you could reach the same goal with less data, with synthetic data or with a smaller sample.

3. Your interest must be balanced against the rights and interests of the people concerned. AI cases are usually won or lost at this step, because the EDPB gives real weight to what people could reasonably expect. Would a customer sending a support ticket have expected it to end up in a training set? If the answer is clearly no, the balance tends to go against you. The protective measures you add then become decisive, such as filtering or pseudonymising the data, giving clearer information, or offering an easy opt-out from the start.

Be realistic about how settled this is. Legitimate interest for AI training is widely used, but it remains contested and has been challenged before regulators and courts. It is a basis you must be able to justify in writing, not a guarantee.

Why Removing Names Is Not Anonymisation

A very common shortcut is to say the data is fine because the names have been removed. Removing names usually gives you pseudonymised data, not anonymous data. Pseudonymised data still relates to real people, and anyone with the extra information, such as your original records, can identify them again.

The Court of Justice of the EU addressed this in the SRB case. It held that pseudonymised data can fall outside the definition of personal data for a recipient who has no reasonable means of identifying the people behind it. For the company that holds the extra information needed to identify them again, the data remains personal data. If you hold the original customer records, that company is you. Anonymisation is a much higher standard. It means that nobody can realistically identify a person from the data, using any means reasonably likely to be used.

Many teams miss a second question entirely: does the trained model itself still contain personal data? The EDPB says this must be assessed case by case, and it sets two separate tests that a model must pass to count as anonymous.

•  it must be very unlikely both that individuals whose data was used can be identified from it, 

• and that their data can be pulled back out of it through clever prompting.

Two further warnings apply. First, if personal data was processed unlawfully while the model was developed, the EDPB considers that this can affect whether the model may lawfully be used afterwards. That applies to you and to anyone else who uses the model, unless the model has been properly anonymised. Second, check what is actually inside your training data. Support tickets and chat logs often contain sensitive information that customers typed without being asked, such as health details or trade union membership. Under the GDPR these are special categories of personal data with much stricter rules, and the fact that you did not intend to collect them is not an exception.

What You Owe Your Customers

Once you have done the assessment, four duties remain.

• Third-party AI tools. Many businesses do not train their own model but send customer data to another company’s model. Your first task is to read the contract and find out whether the provider trains on your inputs. Business plans often exclude this, while free or consumer versions often allow it. A provider that handles data only on your instructions is your processor. If it uses your customers’ data for its own purposes, it becomes a controller for that processing, meaning a business that decides why and how the data is used. That changes who is responsible for what, and you want to discover it before an incident, not during one.

• Transparency. Your privacy notice must clearly say that you use the data to train AI, and why. If your notice only ever mentioned providing the service, that notice is itself a breach, however strong your assessment is.

• The right to object. If you rely on legitimate interest, people can object to the processing. You must then stop, unless you can show compelling legitimate grounds that override their interests, rights and freedoms. Build an easy way to object into your product, rather than handling each case through individual support emails.

• Erasure requests. Deleting someone from your database is straightforward. Explaining what deletion means for a model that has already been trained is not, and regulators are paying attention to this question. Do not promise anything in your privacy notice that your engineering cannot actually deliver.

Proposed Changes to the Rules

The EU has been working on a package of simplification measures for its digital laws, often called the Digital Omnibus, and it is easy to misunderstand. It has moved in two separate parts. The part dealing with artificial intelligence was adopted separately and mainly changed deadlines under the AI Act. The part that would change the GDPR has followed its own, slower path. Its proposals have included an express rule confirming legitimate interest as a basis for developing AI, with extra transparency and a right to object attached. The wording of that rule has changed more than once during negotiations. The EDPB and the European Data Protection Supervisor have also criticised several parts of the package.

The practical lesson is simple. Do not build your plans on relief that may come later, or may come in a different form than you read about. Until any change is formally adopted and in force, the current GDPR rules described in this article apply. Check whether the rules have changed before relying on this article, and keep your documentation flexible enough to adapt.

Conclusion

In summary, training AI on customer data is not banned, but using data collected for another purpose is a change of purpose that needs a documented compatibility assessment before you start. Consent is often the less practical basis because a trained model cannot easily reflect withdrawals, so most serious projects rely on legitimate interest, supported by a properly documented three step test. Removing names is not anonymisation, the trained model itself may still contain personal data, and your privacy notice, objection handling and provider contracts must all match what you actually do.