What Is GDPR Compliance for AI?
- Compliance means every AI system touching personal data has an identified lawful basis per purpose, a completed DPIA where required, and evidence that both were reasoned.
- There is no certificate. GDPR compliance is self-assessed and demonstrated through documentation.
- Data subject rights apply to AI, including access, objection and erasure, and answering them for a trained model is genuinely difficult.
- Fines reach €20m or 4% of worldwide turnover, and AI has already produced enforcement in Europe.
- Nothing in the GDPR has changed for AI. The proposals that would change it are not adopted.
What compliance actually requires
GDPR compliance for AI means you can show, for every AI system that touches personal data, that you identified the purposes, established a lawful basis for each, assessed the risk to the people affected, put safeguards in place, and can answer their rights requests.
Like the AI Act and unlike ISO standards, it is self-assessed. Nobody certifies you. Supervisory authorities assess you if and when they look, and what they look at is documentation.
A record of processing. Which AI systems process personal data, what data, for what purposes, on what basis, shared with whom, retained how long.
A lawful basis per purpose. Collecting training data, training the model, evaluating it, and running it in production are separate purposes. Each needs its own basis and its own reasoning.
A legitimate interests assessment where you rely on that basis. Purpose, necessity, balancing. Written down, with the safeguards you applied.
A DPIA where the risk is high. Which for most AI affecting individuals means always.
Transparency. Privacy information that actually describes the AI processing, provided to the people concerned. Hardest where data was obtained indirectly.
Safeguards for automated decisions. Where Article 22 applies, human intervention, the ability to express a view, and the right to contest.
A route to exercise rights. Access, rectification, erasure, objection, and restriction, applied to AI processing.
Contracts. Processor terms where a vendor handles personal data on your behalf, and clarity about who is controller for what.
The rights problem
Data subject rights are where AI creates genuine difficulty rather than paperwork.
Access. A person can ask what personal data you hold and how it is used. For training data this is usually answerable. For what a model has learned, it is contested and unsettled.
Objection. Where you rely on legitimate interests, people can object. Regulator decisions on large platforms have treated a workable objection route as central to whether the balancing test succeeds at all, so this is not a peripheral obligation.
Erasure. Removing someone from a training set is feasible. Removing their influence from a trained model is not, in any straightforward sense. The pragmatic position most organisations take is to remove data from training sets, exclude it from future training runs, apply output filtering, and document why full removal from the model is not technically possible. That position needs to be reasoned and recorded, not assumed.
The practical answer is to design for these before deployment. Retrofitting an objection route into a live model is far harder than building one in.
Who enforces it
National supervisory authorities in each member state, coordinated through the European Data Protection Board. In the UK, the Information Commissioner's Office.
Fines reach €20m or 4% of total worldwide annual turnover, whichever is higher, for the most serious breaches. Authorities can also order processing to stop, which for an AI product is more consequential than a fine.
Enforcement in AI has so far concentrated on lawful basis for training, transparency, and the adequacy of objection routes. Several large platforms have been permitted to proceed with model training after agreeing safeguards, which is instructive: the outcome was conditional approval, not prohibition, and the conditions are the useful part.
What has not changed
Nothing in the GDPR has changed for AI.
The Digital Omnibus proposal would add an explicit legitimate interest basis for AI development and operation, with enhanced safeguards and an unconditional right to object, and would create a route for processing special-category data to detect and correct bias. Those provisions are in the data regulation, which remains under negotiation and has not been adopted. The EDPB and EDPS objected to parts of the package in February 2026.
Until adoption, the existing analysis applies: identify your basis, run the three-part balancing test where you rely on legitimate interests, apply safeguards, and document it.
What compliance does not cover
It does not make you compliant with the EU AI Act, which imposes separate per-system obligations. See our EU AI Act service page.
It does not address model quality, safety or bias except where those affect individuals' data rights. ISO/IEC 42001 covers the governance system around those questions. See our ISO/IEC 42001 service page.
It does not cover information security beyond the security of processing obligation, which is why buyers ask for ISO 27001 or SOC 2 alongside.
What to do next
Build the record first: which AI systems touch personal data, for what purposes. Almost every subsequent question depends on it.
Our free AI impact assessment gives a first view, and our readiness and gap assessment produces the detailed position for a fixed fee.
References
FAQ
Is there a GDPR certification for AI?
No. Compliance is self-assessed and demonstrated through documentation. Certification schemes under Article 42 exist in limited form but do not cover AI generally.
Can someone ask us to delete their data from a trained model?
They can ask. Removing data from training sets is feasible; removing its influence from a trained model generally is not. Document the measures you take and why full removal is not technically possible.
What are the fines?
Up to €20m or 4% of total worldwide annual turnover, whichever is higher, plus orders to stop processing.
Do we need a DPIA for every AI system?
For every system likely to result in high risk, which covers most AI making or supporting decisions about people. Where you conclude one is not needed, record why.
Has the legitimate interest basis for AI training been settled?
No. That provision is in the Digital Omnibus data regulation, which is still a proposal.
About Hael
Hael is an advisory firm specialising in AI governance and security compliance, built on fifteen years of regulatory practice advising firms through authorisation, supervision and examination. We take companies through data protection for AI, the EU AI Act, ISO/IEC 42001, SOC 2 and ISO 27001. On engagements involving regulated financial services firms we work alongside Buckingham Capital Consulting, the partner firm that has advised payment and e-money firms on FCA authorisation and compliance since 2013.
This guide is general information and is not legal advice on your particular circumstances.