The word in Article 22(1) is solely, and it is the entire hinge. Put a human in the loop and the article does not apply; put a human in the loop who clicks approve on ninety-eight per cent of what the model proposes and, on the regulators’ reading, it still does.
Where the test comes from
Article 22(1) of Regulation (EU) 2016/679 gives the data subject 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. The Regulation does not define “solely”, and it does not use the phrase “meaningful human involvement” at all.
That phrase comes from the Article 29 Working Party’s Guidelines on Automated individual decision-making and Profiling (WP251 rev.01), adopted in their revised form in February 2018 and endorsed by the EDPB. Their formulation is that to qualify as human involvement, the controller must ensure that any oversight of the decision is meaningful rather than just a token gesture — it should be carried out by someone who has the authority and competence to change the decision, and the reviewer should consider all the relevant data.
Two things follow that are worth being clear about. First, guidelines are not the Regulation: they express the regulators’ interpretation and a court is not bound by them, though supervisory authorities enforce on that basis and national courts have treated them as persuasive. Second, the test is deliberately anti-formalist. It asks what the human does, not whether a human exists in the process diagram.
This is a description of the guidance, not legal advice. Whether a particular review process takes a decision outside Article 22 is fact-specific, and if the answer matters commercially it is a question to put to counsel with the actual workflow in front of them.
Authority, competence, and the whole file
Authority
The reviewer must be able to reach a different outcome. If the reviewer’s only options are approve or escalate, and escalation is rare and slow, the reviewer does not have authority in any operative sense. This is a question about the system’s permission model as much as about the org chart: if the interface has no control that overrides the model’s score, authority does not exist regardless of what the policy says.
Competence
The reviewer must understand what they are looking at well enough to disagree with it. A reviewer who cannot say what the model used, what it weighted, or what a borderline case looks like cannot form an independent view; they can only defer. Competence is why the AI literacy obligation under the AI Act and the Article 22 analysis under the GDPR end up pointing at the same training programme, and why Article 14 of the AI Act on human oversight is the natural companion obligation for the same workflow.
All the relevant data
The reviewer has to see the inputs, not just the output. A screen that shows “risk score: 0.87 — decline” and nothing else makes meaningful review impossible in principle; the reviewer has nothing to reason from but the model’s conclusion. A screen that shows the score, the principal contributing factors, the underlying record and any applicant-supplied context makes it possible — and possible is the most a design can guarantee.
Review designs that fail
- The default-approve queue. Items expire into “approved” if not actioned within a window. The human involvement is literally the absence of an action.
- The throughput target that forbids reading. A reviewer measured on decisions per hour, at a rate implying seconds per case, is being paid to click. The guidance’s “consider all relevant data” condition and a forty-cases-an-hour target are not compatible claims.
- The override that is punished. Where overriding the model counts against the reviewer in quality assurance, the authority is nominal. This one is invisible in a process diagram and obvious in the override statistics.
- Review of the population, not the case. Sampling one per cent of decisions is a valid model-governance control and is not human involvement in the ninety-nine per cent that were not sampled. Each decision either had a human or it did not.
- The human who only sees the output. Covered above, and worth repeating because it is the most common one and the easiest to fix.
What a review that passes looks like
It is not complicated, and it is expensive, which is the real reason these designs are rare. The reviewer sees the model output alongside the record it was computed from and the factors that drove it. They have a control that sets a different outcome, and using it requires a short reason rather than an approval from a manager. They are not measured primarily on volume. Override rates are monitored as a health signal — not to push them down, but because an override rate near zero and an override rate near half both indicate the process is not doing what it claims.
One structural point is easy to miss: a process that genuinely takes a decision outside Article 22 does not thereby escape data protection law. The processing still needs a lawful basis, still needs to satisfy fairness and transparency, and the data subject still has the Article 21 objection right where the basis is legitimate interests, which is covered separately on the right to object to profiling. Sector rules can also impose their own explanation duties regardless of Article 22 — an adverse credit decision is the standard example.
It is also worth deciding whether you want to be outside Article 22 at all. If a decision falls within it but fits one of the Article 22(2) exceptions — necessity for a contract, authorisation by Union or Member State law, or explicit consent — you may lawfully take it on a solely automated basis provided you implement the Article 22(3) safeguards: at a minimum the right to obtain human intervention, to express a point of view, and to contest the decision. Building the 22(3) safeguards for a genuinely automated pipeline is often cheaper and more honest than staffing a review layer whose meaningfulness you will have to defend.
Proving it eighteen months later
The accountability principle in Article 5(2) means the burden is yours, and the evidence you will wish you had is operational rather than documentary. A policy asserting that reviewers may override proves nothing. What proves something is a per-decision record of who reviewed it, what they were shown, how long they had it open, whether they changed the outcome and what reason they gave — kept for as long as the decision can be challenged, and no longer, since Article 5(1)(e) applies to your audit trail too. Aggregate override rates by reviewer and by decision type are the second artefact; they are what turns “reviewers can override” into a checkable claim.
What is still unsettled
The threshold is not fully mapped by case law. The Court of Justice has addressed what counts as a decision, in Case C-634/21 (SCHUFA Holding), decided on 7 December 2023, and what has to be explained about the logic involved, in Case C-203/22 (Dun & Bradstreet Austria), decided on 27 February 2025 — both available from the Court’s own case register. Neither judgment sets out a test for how substantive a human review must be to defeat “solely”. That gap is filled at present by the WP251 guidance and by national supervisory practice, which is not the same thing as settled law, and a controller relying on a thin review layer is relying on an interpretation that has not been tested at the Court. The nearer question — when the article engages at all — is worked through on what counts as a solely automated decision.
Top comments (0)