Training an AI model: legitimate interest must be proven

On 21 October 2026, Member States (EU · Council of the EU) endorsed, in their position on the digital omnibus, a new Article 88 bis of the GDPR: developing and operating an AI system or model “may” rely on legitimate interest. The same text states, in a recital, that AI does not “in itself constitute a legitimate interest”. Between 2021 and 2025, the Irish regulator (Ireland · DPC) reviewed some 180 AI products and services, making this legal basis one of its main areas of focus.

The omnibus does not create a right to train a model on personal data: European regulators (EU · EDPB) have already accepted it, case by case, since December 2024. What it does not remove is the burden of proof. The question is no longer whether you may train a model, but whether you can prove you did it properly.

What is already settled

• The opinion on AI models (EU · EDPB): adopted on 17 December 2024 (Opinion 28/2024), it accepts legitimate interest case by case, under a three-step test. A model is anonymous only if it is very unlikely that individuals can be identified or their data extracted through queries. Unlawful development can make deployment unlawful, unless the model is effectively anonymised.
• The link with the AI Act (EU · Regulation (EU) 2026/1744): since 27 July 2026, the amended AI Act (Article 4a) has extended to all AI systems and models the exceptional, strictly framed option of processing sensitive data to detect and correct bias, previously reserved for high-risk systems. The GDPR still applies to any training on personal data.
• The regulators’ joint opinion (EU · EDPB and EDPS): on 10 February 2026, they found a dedicated AI article unnecessary, since legitimate interest was already accepted in the 2024 opinion.
• Meta (Ireland · DPC): on 21 May 2025, the regulator took note of the plan to train on public content from adult Facebook and Instagram users, from 27 May 2025, after several measures: user notices, objection forms, filtering, a legitimate interest assessment.
• LinkedIn (Ireland · DPC): on 7 November 2025, the DPC stated that it had neither approved nor found compliant LinkedIn’s model training, while deciding not to intervene after changes: narrower scope, filters on sensitive content, protection of minors, a report required within five months.
• Web scraping (EU · EDPB): Guidelines 03/2026, adopted on 7 July 2026 for consultation: minimisation at collection, exclusion of sites aimed at minors, avoiding sites that block automated collection.

What is still pending

• Article 88 bis (EU · Council of the EU): processing for the development and operation of AI “may” rely on legitimate interest under Article 6; a recital states that individuals’ rights may prevail, in particular where the data subject is a child. The right to object remains the one in Article 21, not the “unconditional” right proposed by the Commission.
• Residual sensitive data (EU · Council of the EU, Article 9): a derogation would cover sensitive data incidentally present in training sets, provided the controller has sought to avoid collecting them, then erases them without undue delay or, where that is impossible or disproportionate, shields them from any further use.
• Parliament’s position (EU · European Parliament): no vote as of 29 October 2026; MEPs can still amend or delete these provisions.
• Final guidelines (EU · EDPB): public consultation on anonymisation (02/2026) and web scraping (03/2026) closed on 30 October 2026. No date had been announced for the final versions as of 29 October 2026.

The real issue: the evidence file

In the Meta and LinkedIn cases, the Irish regulator asked for the same documents: a legitimate interest assessment, a data protection impact assessment, a compatibility assessment with the original purpose, then an evaluation report after launch. None of these requirements would disappear with Article 88 bis. The Council position recalls that applying legitimate interest to AI is assessed case by case. European regulators (EU · EDPB and EDPS) had already noted that the proposed article brought “no legal clarification”.

The Irish report of 25 September 2026 confirms it: out of some 180 products and services supervised, most engagements ended with recommendations, the regulator saying it stands ready to intervene urgently where risks remain poorly mitigated. The regulator does not approve: it lets those who have documented proceed, under supervision.

Three points to watch

1. Model anonymity is not presumed. The EDPB requires a demonstration that individuals can be neither identified nor have their data extracted through queries. A model declared anonymous without testing is still a model processing personal data.

2. The supply chain. Under the 2024 opinion, data processed unlawfully during development can taint deployment. Under Guidelines 03/2026, anyone reusing a dataset already scraped by someone else is responsible for their own processing. Buying a model or a dataset does not exempt you from checking its origin.

3. Objection must come before training. European regulators (EU · EDPB and EDPS) ask for people to be informed early enough, because removing data from a trained model can prove technically difficult. A right to object opened after the fact offers little protection.

What data teams and DPOs should do

1. Map training sources: internal, scraped, purchased data, with a legal basis for each.
2. Write the legitimate interest assessment before training, together with the impact assessment and the compatibility assessment.
3. Minimise at collection: filters on sensitive data, exclusion of sites aimed at minors, respect for technical blocks such as robots.txt.
4. Open objection before training, with a reasonable lead time and a form that actually works.
5. Test data extraction on the model before declaring it anonymous, and keep the results.
6. Require model and dataset suppliers to give a written warranty on the lawful origin of the data.
7. Set up an erasure procedure for sensitive data discovered afterwards in training sets.

Training a model is allowed; training it without a file is not.

The question to ask on Monday morning

To your head of data science and your DPO: “For each training dataset, can we show its source, its legal basis and the date on which people were able to object?”