November 21, 2026

GDPR Right to Explanation: What Article 22 and the CJEU Now Require

An automated credit decision letter under a magnifying glass beside a laptop showing a decision-tree diagram and a tabbed volume of EU data protection law

Whether a GDPR right to explanation exists for decisions made by algorithms was, for most of the regulation’s first decade, a question lawyers could argue either way. A bank’s model declines your loan; a credit agency’s number quietly costs you a phone contract. Can you make the organisation tell you why? The General Data Protection Regulation’s binding articles never say “right to explanation”; an explanation of a decision is mentioned only in a recital.

Two judgments of the Court of Justice of the European Union (CJEU) have since moved the ground a long way. They have not finished the job. This article is general information, not legal advice.

The question, stated precisely

Strip away the slogans and three separate questions remain. First, does EU data protection law give a person a right to an individualised, after-the-fact account of how a particular automated decision about them was reached, rather than only a general description of the system? Second, which decisions does that right cover: only those taken “solely” by machine, or also decisions where a human signs off on a machine’s recommendation? Third, what must the account contain when the model is a gradient-boosted ensemble or a neural network whose internal “principles” no person can read off directly?

Credit scoring, insurance pricing and gig-platform deactivation all run on such models. People do look for an answer: a web search returns a Wikipedia entry on the right to explanation and, near the top, the 2017 paper that argued the right does not exist at all.

European Union flags outside a glass government building
EU institutions wrote the GDPR. AI-generated image.

What the text of the GDPR actually says

Four provisions carry the weight. Article 22(1) gives a data subject the right not to be subject to a decision based solely on automated processing, profiling included, that produces legal effects or similarly significantly affects them. Article 22(2) then lists three exceptions: the decision is necessary for a contract, it is authorised by Union or member-state law with suitable safeguards, or the person has given explicit consent. For the contract and consent routes, Article 22(3) requires the controller to provide at least the right to obtain human intervention, to express a point of view, and to contest the decision.

Separately, Articles 13(2)(f), 14(2)(g) and 15(1)(h) require controllers to tell people about the existence of automated decision-making referred to in Article 22 and, at least in those cases, to give meaningful information about the logic involved together with the significance and envisaged consequences of the processing. Article 15 is the right of access, which a person can invoke at any time, including after a decision.

Then there is Recital 71. Among the safeguards it lists for automated decisions is the right “to obtain an explanation of the decision reached after such assessment”. That phrase is the only place the GDPR names an explanation of a decision. Recitals in EU law guide interpretation but do not create free-standing obligations, and Article 22(3) itself omits the word.

The 2017 dispute: does the right exist?

Goodman and Flaxman (2016), in a widely read preprint later published in AI Magazine, said the regulation would effectively create a “right to explanation” for algorithmic decisions. A year later Wachter, Mittelstadt and Floridi (2017) answered with a title that left no room for doubt: such a right does not exist in the GDPR.

Their argument, steelmanned, runs as follows. The binding text of Article 22(3) lists human intervention, a point of view and contestation, but not explanation; the drafters had the Recital 71 wording in front of them and, on the authors’ reading of the legislative history, left it out of the article. Articles 13 to 15 speak of the envisaged consequences of processing, language that looks forward. So what those articles guarantee, in this view, is a right to be informed about system functionality in general, not an ex post justification of one decision. Add the narrow trigger and the trade-secret limits in Recital 63, and any explanation right is thin.

Selbst and Powles (2017) replied that this misreads the text. “Meaningful information” has to be meaningful to the person receiving it, which means information that lets them act on their rights, including the Article 22(3) right to contest. A functional right, under whatever name, was already there.

A third position sat to the side. Edwards and Veale (2017) warned that the triggers were restrictive and unclear, that explanation techniques might not deliver what the law wanted, and that treating explanation as the cure risked a “transparency fallacy”. In a later paper, Wachter, Mittelstadt and Russell proposed counterfactual explanations (the smallest change in inputs that would have flipped the outcome) as a way to give people something useful without opening the model.

Regulators took a middle line. The Article 29 Working Party’s guidelines on automated decision-making and profiling (WP251rev.01), later endorsed by the European Data Protection Board, read “solely” as excluding only decisions with meaningful human involvement, not a token sign-off, and treated the transparency duties as requiring the main criteria and reasons rather than the full algorithm.

What the courts have said since

SCHUFA (C-634/21, 2023)

In OQ v Land Hessen (Case C-634/21, SCHUFA), decided on 7 December 2023, a German credit agency produced a probability score, and a lender refused credit on the back of it. The CJEU held that the automated generation of the score is itself a decision under Article 22(1) where a third party “draws strongly” on it to establish, implement or terminate a contract. Its reasoning was about gaps:

“there would be a risk of circumventing Article 22 of the GDPR and, consequently, a lacuna in legal protection”

CJEU, Case C-634/21, SCHUFA, 7 December 2023, para. 61

The Court also observed that without this reading the person could not get Article 15(1)(h) information from the agency, and the lender would generally be unable to supply it either. SCHUFA did not decide what an explanation must contain. It widened who owes one, and it loosened “solely” in practice: a nominally human decision that leans heavily on a machine score can still sit inside Article 22.

Uber drivers in the Amsterdam Court of Appeal (2023)

In a judgment of 4 April 2023 (ECLI:NL:GHAMS:2023:793), the Amsterdam Court of Appeal held that Uber’s permanent deactivation of three of four drivers’ accounts was based solely on automated processing, because fraud-detection software generated the signals and Uber could not show meaningful human involvement such as a hearing. For the fourth driver, a prior personal conversation was enough. Uber had to disclose at least the factors and their weighting at an aggregate level, enough for the drivers to understand and contest the decision; trade-secret concerns did not justify a complete refusal.

Dun & Bradstreet Austria (C-203/22, 2025)

The decisive step came on 27 February 2025 in CK v Magistrat der Stadt Wien (Case C-203/22, Dun & Bradstreet Austria). A consumer was refused a mobile phone contract after an automated creditworthiness assessment. The CJEU held that Article 15(1)(h) gives the data subject

“a right to an explanation of the procedure and principles actually applied”

CJEU, Case C-203/22, Dun & Bradstreet Austria, 27 February 2025, para. 58

to their personal data, in a concise, transparent, intelligible and easily accessible form. The requirement “cannot be satisfied either by the mere communication of a complex mathematical formula, such as an algorithm” (para. 59). The Court suggested it could be enough to tell the person how far a change in the data taken into account would have led to a different result (para. 62), which is close to a counterfactual explanation. Where trade secrets or third-party data are at stake, the controller cannot simply refuse; it must give the allegedly protected material to the supervisory authority or court, which balances the interests case by case, and a national rule that excludes access whenever a trade secret is involved is precluded (paras. 74 to 76).

Together, these rulings answer the first question. Wachter, Mittelstadt and Floridi were right that Article 22(3) does not say “explanation”, but the Court located an individualised, ex post explanation right in Article 15(1)(h) instead, which is closer to Selbst and Powles’ position. The right is real. Its edges are not.

Competing readings that survive the case law

Scope beyond “solely” automated decisions. Article 15(1)(h) applies “at least” to Article 22 decisions. One reading treats those words as a floor inviting wider application to decision-support systems; the other treats them as the outer limit. The Amsterdam court read the English and French texts narrowly. SCHUFA’s “draws strongly” test narrows the gap, but only for three-party structures; it does not say how much human review takes a two-party decision outside Article 22.

What “actually applied” means for opaque models. For a scorecard with ten weighted factors, “procedure and principles actually applied” is easy to state. For a deep network, the literal procedure is millions of parameters, which the Court has said is not enough. The usual substitutes, feature-attribution methods such as SHAP or LIME and counterfactuals, are post-hoc approximations: they describe a simpler model that mimics the real one near your data point. One reading says the Court’s para. 62 endorses exactly this. The stricter reading says an approximation with unknown fidelity is, by definition, not the procedure “actually” applied. The case law does not choose. (Why such models resist inspection is covered in x9m8’s explainer on how neural networks learn.)

Whether Article 22(3) adds anything. Because the 2025 ruling rests on Article 15, a separate question survives: does the right to contest under Article 22(3), read with Recital 71, require the controller to give reasons at the moment of contest, even without an access request? The answer affects timing, and decisions made under the member-state-law exception.

The EU AI Act now adds a fourth layer. Article 86 gives people affected by decisions based on certain high-risk systems a right to clear and meaningful explanations of the AI system’s role in the decision and of the decision’s main elements, but paragraph 3 makes it apply only where Union law does not otherwise provide the right. After Dun & Bradstreet, that subsidiarity clause may leave Article 86 with a fairly small field. For the wider map, see x9m8’s guides to how AI is regulated in the EU and US and the EU AI Act dates schedule.

Data center aisle with server racks and blue lights
Automated decisions run on infrastructure like this. AI-generated image.

Doctrinal analysis built for this article

To see where the remaining gap sits, the table below breaks each relevant provision into its trigger, its timing, the content it requires, and what the courts have added.

ProvisionTriggerTimingContent requiredJudicial gloss
Art. 13(2)(f) / 14(2)(g)Existence of Art. 22 decision-makingAt collection (ex ante)Meaningful information about the logic, significance, envisaged consequencesLittle direct case law
Art. 15(1)(h)Access request; Art. 22 decisions “at least”Any time, including after a decisionSame words as aboveD&B: explanation of procedure and principles actually applied to the person’s data; formula alone insufficient; counterfactual may suffice
Art. 22(1)Solely automated; legal or similarly significant effectBefore decision (a prohibition with exceptions)None as suchSCHUFA: score is a decision where a third party draws strongly on it; Amsterdam: burden on controller to show meaningful human involvement
Art. 22(3)Decision under contract or consent exceptionAt or after decisionHuman intervention, point of view, contestExplanation not yet held to be part of it
Recital 71Same as Art. 22After assessmentExplanation of the decisionInterpretive aid only
AI Act Art. 86High-risk Annex III system; adverse effectAfter decisionRole of the AI system and main elements of the decisionUntested; subsidiary to other Union law

Three findings drop out. The individualised ex post right now has a firm textual home (Article 15(1)(h)) and a binding interpretation. The two residual gaps are not in the same place: one concerns the trigger (how much human involvement removes a decision from Article 22), and the other concerns content (whether a post-hoc approximation counts as the procedure actually applied). Both judgments concerned credit scoring, and neither engaged with a model whose logic cannot be stated as weighted factors. The hardest version of the content question has not yet reached the Court.

What this analysis does not show: how any national court will rule, how sector rules such as consumer-credit law interact, or how often controllers comply.

A test case that would settle it

One well-chosen preliminary reference could resolve what remains. A hypothetical:

  • Facts. A lender uses a deep neural network to score unsecured loan applications. A credit officer reviews every application but, on the lender’s own records, overrides the model in a small fraction of cases and spends a minute or two on each. An applicant is refused and requests access under Article 15(1)(h).
  • Response. The lender says the decision was not solely automated, so Article 22 does not apply. Alternatively, it supplies a SHAP attribution for the applicant’s file and a counterfactual (“an income higher by a stated amount would have produced approval”). It refuses to disclose the attribution method’s measured fidelity to the real model.
  • Question 1. Does a human review with a documented low override rate and short review time amount to meaningful human involvement, or does the controller “draw strongly” on the score within the meaning of SCHUFA, so that Article 22 applies?
  • Question 2. Does Article 15(1)(h) require that a post-hoc explanation be faithful to the model actually used, and if so, must the controller disclose (at least to the authority or court) evidence of that fidelity?

If the Court answers Question 1 by reference to observable indicators such as override rates and review time, controllers get a workable test and the “solely” loophole narrows. A demand for demonstrable fidelity under Question 2 would push high-stakes uses toward inherently interpretable models.

Wooden gavel resting on law books
Whether a right to explanation exists is still argued in courts and journals. AI-generated image.

Where the law points

What follows is opinion, labelled as such. The 2017 argument that no right to explanation exists has, in substance, lost. The CJEU did not adopt its strongest premise, that Articles 13 to 15 are only forward-looking, and the phrase “actually applied” points squarely at the individual decision. Selbst and Powles’ functional reading is closer to where the law now stands, even though the Court built it from Article 15 rather than Article 22(3).

My reading is that the content question will be answered pragmatically. The Court’s own suggestion in para. 62 signals that counterfactual-style accounts can suffice, and the Court will likely accept post-hoc explanations as long as controllers can show to an authority that the explanation tracks the real model’s behaviour for the person concerned. The trigger question is the bigger worry: rubber-stamp review is cheap and invisible from outside.

Some falsifiable predictions follow:

  • The next CJEU reference touching Article 22 will turn on the human-involvement threshold rather than on explanation content.
  • Courts will not require disclosure of model weights or source code to data subjects in any case where a factor-level or counterfactual account is offered; they will route such material to authorities or courts instead.
  • AI Act Article 86 will rarely be the decisive ground in reported cases, because paragraph 3 defers to the GDPR. A string of decisions resting on Article 86 alone would falsify this.

Researchers could test the empirical side now. A useful study would audit access-request responses from a sample of lenders before and after the 2025 ruling, coding whether each response contains case-specific factors, a counterfactual, or only generic text. Another would measure how faithfully those explanations track the models actually run. For the American contrast, see x9m8’s overview of state AI laws.

Common questions

Can I demand a company’s source code or algorithm under the GDPR?

Generally no. Dun & Bradstreet held that handing over a formula or algorithm would not satisfy the duty anyway; what the person is owed is an intelligible account of how their data were used to reach the result.

Does it apply if a human approved the decision?

It depends on how real the human’s role was. Regulators and the Amsterdam Court of Appeal have treated token sign-offs as insufficient to take a decision outside Article 22, and SCHUFA shows that a decision-maker who draws strongly on a machine score can still bring the process within it. The exact line is unsettled.

Can a company refuse because of trade secrets?

Not outright. Under the 2025 ruling, the controller must give the allegedly protected information to the supervisory authority or court, which weighs the competing interests in the individual case. A blanket national exception for trade secrets is not allowed.

Does the EU AI Act create a separate right to explanation?

Yes, for decisions based on certain high-risk systems with adverse effects. It applies only where Union law does not already provide the right, so in many cases the GDPR route will come first.

References

  1. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 (General Data Protection Regulation), OJ L 119, 4.5.2016, recital 71 and arts. 13, 14, 15, 22. EUR-Lex.
  2. Court of Justice of the European Union, Case C-634/21, OQ v Land Hessen (SCHUFA Holding AG intervening), judgment of 7 December 2023. EUR-Lex.
  3. Court of Justice of the European Union, Case C-203/22, CK v Magistrat der Stadt Wien (Dun & Bradstreet Austria GmbH), judgment of 27 February 2025. EUR-Lex.
  4. Gerechtshof Amsterdam, 4 April 2023, ECLI:NL:GHAMS:2023:793 (drivers v Uber B.V.). Rechtspraak.nl open data.
  5. Wachter, S., Mittelstadt, B., and Floridi, L. (2017). Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation. International Data Privacy Law, 7(2), 76-99. doi:10.1093/idpl/ipx005. Journal page.
  6. Selbst, A. D., and Powles, J. (2017). Meaningful information and the right to explanation. International Data Privacy Law, 7(4), 233-242. doi:10.1093/idpl/ipx022. Journal page.
  7. Goodman, B., and Flaxman, S. (2016). European Union regulations on algorithmic decision-making and a “right to explanation”. arXiv:1606.08813 (preprint; published version doi:10.1609/aimag.v38i3.2741). arXiv.
  8. Edwards, L., and Veale, M. (2017). Slave to the Algorithm? Why a ‘Right to an Explanation’ Is Probably Not the Remedy You Are Looking For. Duke Law & Technology Review, 16(1), 18-84. Duke Law Scholarship Repository.
  9. Wachter, S., Mittelstadt, B., and Russell, C. (2017). Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR. arXiv:1711.00399 (preprint version). arXiv.
  10. Article 29 Data Protection Working Party (2018). Guidelines on Automated individual decision-making and Profiling for the purposes of Regulation 2016/679, WP251rev.01. European Commission.
  11. Regulation (EU) 2024/1689 (Artificial Intelligence Act), art. 86, Right to explanation of individual decision-making. Text via artificialintelligenceact.eu.
  12. Wikipedia contributors. Right to explanation. Wikipedia (cited only as evidence of public interest in the question).

429 Posts

Peter Kalafatis writes and edits X9M8. He holds a bachelor's degree from LIU Post and a master's in Media Studies from The New School, with academic work in the philosophy of science, human behavior, AI, and memetics.

View All Posts

Leave a Reply

Your email address will not be published. Required fields are marked *