European organisations preparing for the AI Act frequently overlook the fact that a binding restriction on automated decision-making has been in force since May 2018 — and that it applies today, without a transition period, to systems already in production.
GDPR Article 22 is short. Its consequences are not.
What the article says
An individual has the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning them or similarly significantly affects them.
Three exceptions permit such decisions: where necessary for entering into or performing a contract; where authorised by EU or member state law with suitable safeguards; or with the individual’s explicit consent. Where the contractual or consent route is used, the controller must implement safeguards including — at minimum — the right to obtain human intervention, to express a point of view, and to contest the decision.
Article 22 also prohibits basing such decisions on special category data (health, biometrics, ethnicity, political opinion, trade union membership, sexual orientation) unless explicit consent or substantial public interest applies, with safeguards.
The EDPB’s long-standing position, unchanged after SCHUFA, is that Article 22 operates as a prohibition in principle — not as a right that the individual must first assert. That reading matters: it means the compliance burden sits with you before anyone complains.
The two phrases that decide everything
“Solely”
The intuitive reading is that inserting any human step defeats the article. That reading is wrong, and expensively so.
The CJEU in SCHUFA confirmed what the EDPB had already said: a human who formally approves while in practice deferring to the algorithm does not make the decision non-automated. What breaks the chain is meaningful human involvement — a person with the authority, competence, information and time to reach a different answer.
This is the single most consequential point on this page, because it means the compliance question is not is there a human? but could that human realistically have said no? A reviewer approving 400 cases a day, with no access to the reasoning behind a score and no incentive to disagree, is providing the appearance of oversight. This is the failure mode examined in designing human-in-the-loop systems that actually work, and it is now a legal exposure as well as a design flaw.
“Similarly significantly affects”
Legal effects are the easy case: contract termination, benefit refusal, entry denial. The harder category is the “similarly significant” one, which regulators have read to include refusal of credit, e-recruitment decisions without human intervention, differential pricing that effectively excludes, and decisions affecting access to essential services.
Marketing personalisation generally falls outside. Personalisation that materially determines whether someone can access a service generally falls inside.
What SCHUFA changed
Case C-634/21, decided 7 December 2023, concerned a German credit reference agency that produced probability scores and supplied them to lenders. SCHUFA’s position was that it merely generated a score; the decision was made by the bank.
The Court disagreed. Where a third party draws strongly on the score in deciding whether to establish, perform or terminate a contract, the generation of that score is itself an automated decision within Article 22.
The structural consequence is significant and under-appreciated. Article 22 obligations do not attach only to the organisation making the final call — they can attach to the organisation producing the model output that the decision rests on. In a market where scoring, screening and risk models are supplied as services, that pulls upstream vendors directly into scope, and makes the allocation of responsibility across that boundary a matter of legal exposure rather than commercial preference.
If you supply a model whose output your customers rely on heavily for consequential decisions, “we don’t make the decision” is no longer a complete answer.
How this interacts with the EU AI Act
They are separate instruments with different logics, and both apply.
| GDPR Article 22 | EU AI Act | |
|---|---|---|
| Trigger | Effect on an individual | System classification |
| Applies from | May 2018 | Staged; high-risk deferred to Dec 2027 |
| Focus | The decision | The system |
| Rights | Individual, directly enforceable | Largely regulator-enforced |
| Remedy | Complaint to a DPA; compensation | Market surveillance; penalties |
The practical point: Article 22 is enforceable now, by individuals, against systems you already operate. The AI Act’s heaviest obligations are still ahead. An organisation that plans only against AI Act deadlines is defending the wrong timeline.
The overlap is also substantial — human oversight, transparency, contestability and documentation appear in both. Controls built for one largely serve the other.
What compliance requires operationally
Inventory the decisions, not just the systems. The unit of analysis under Article 22 is the decision and its effect on a person. A single model may feed several decisions with different risk profiles.
Test whether your human involvement is real. Measure override rates. A rate near zero is not evidence that the model is excellent; it is evidence that the reviewer is not reviewing. Give reviewers the reasoning behind the output, the authority to disagree, and a workload that permits it.
Be able to explain the logic. Articles 13–15 require meaningful information about the logic involved, and the significance and envisaged consequences. This does not demand disclosure of model weights. It demands that an individual can understand what factors drove the outcome for them.
Build the contest route before you need it. Human intervention, statement of view, and challenge must be practically available — not theoretically available behind a support queue.
Complete a DPIA. Systematic evaluation based on automated processing with significant effects is expressly listed as requiring a Data Protection Impact Assessment. Where the system is also high-risk under the AI Act, the two assessments should be run as one exercise with a single evidence base.
Agentic AI: Supervisory Evidence and Accountability
Added 12 October 2026.
On 8 October 2026 the UK Information Commissioner’s Office published the outcome of its supervision of foundation-model developers, together with a report clarifying its expectations and a call for evidence on agentic AI.
Classify this correctly before relying on it, because the obvious shorthand is wrong. This is not an enforcement action. No penalty, fine, enforcement notice, assessment notice, information notice, reprimand or undertaking was issued against any of the ten developers, and no statutory power was exercised in the supervision programme. What the ICO reports is voluntary commitments secured through regulatory supervision, with progress it continues to monitor.
The ICO’s own verbs are the safe ones: it secured changes, following scrutiny, in response to its regulatory expectations. Enforced, penalised, sanctioned, fined, ordered and compelled are all wrong here, and a brief that uses them describes a different event.
Keep the disjunction. The ICO’s formulation is that developers have made or committed to make changes — and several have items still in future tense. Progress is not completion.
Four separate things in one announcement
They are frequently merged, and they carry very different weight:
- Supervisory activity producing voluntary commitments from ten foundation-model developers, covering transparency, rights mechanisms and safeguard assessments.
- Regulator policy positions on how existing UK GDPR duties apply — clarification, not new law, and not the statutory code.
- A call for evidence on agentic AI, open 8 October to 20 November 2026. A call for evidence creates no obligation on anyone. It precedes guidance, which precedes a forthcoming statutory code of practice on AI and automated decision-making for which no date is given. The 20 November date is a deadline for submitting views, nothing more.
- Pre-decisional enquiries into agent testing and deployment, which the ICO describes as ongoing and which have produced no findings.
There is no new obligation on deployers in any of it. If a summary tells you otherwise, check which of the four it is describing.
What is genuinely new, and what is six years old
The report frames itself as clarifying rather than creating, and says so: data protection law applies to the processing of personal data in foundation-model development, and challenges in applying the law do not justify failing to address data protection fundamentals.
Newly articulated application of existing law:
- Two candidate Article 9 conditions for special category data in training. “Manifestly made public” is usable only where developers verify that the person intended to share personal data about themselves publicly and implement effective technical measures to filter out cases where that cannot be verified. “Scientific research” is the other. Where neither fits, the ICO’s position is that developers should focus on preventing collection of special category data, filtering it out, or avoiding processing it.
- Transparency is sharpened. Vague terms — the ICO’s example is stating that AI was trained on “publicly accessible information” — are unlikely to be sufficient, and the disproportionate-effort exemption should not be the default option.
- Embeddings. Where the underlying datasets include personal data, the resulting embeddings can still be personal data.
Not new, and should not be sold as such: the position that a trained model can itself contain personal data, which the ICO traces to its own 2020 statement; and that legitimate interests is realistically the only lawful basis developers can explore, which comes from its earlier generative AI consultation series.
One admission worth recording. The ICO states that current foundation-model training practices present technical challenges for complying with UK data protection law and data protection by design, and that it is proactively raising these boundaries of the law with Government. A regulator saying the law may not fit the technology, and taking that to the legislature, is a more consequential signal than any individual commitment in the package.
A separation to maintain
The ICO has a separate open formal investigation, announced 3 February 2026, concerning a different company and its chatbot. That company is not among the ten, the investigation has reached no finding, and nothing about it is part of the commitments described above.
Keep the two apart in anything you write. Blurring an open investigation into a list of cooperating developers misstates both, and is unfair to every party named.
What a deployer should do with this
Nothing is required of you. Three things are nonetheless worth doing:
- Read the transparency expectations as a drafting standard. If your own privacy information says a model was trained on “publicly available data”, the ICO has just said that formulation is unlikely to be sufficient — for developers. The same sentence in your notice will not read better.
- Ask your supplier which commitments apply to them, and whether theirs are made or still committed. The ICO has published the distinction; your contract probably has not.
- Consider responding to the call for evidence by 20 November. The ICO is asking how agentic AI should be governed before it writes guidance, and deployer evidence is thinner in these exercises than developer evidence.
The through-line
Article 22, the AI Act, and the AI Human Proof standard converge on one requirement from three directions: a person must be genuinely in a position to say no, and you must be able to show it.
European law reached that conclusion first. It has been enforceable for years.