Algorithmic Accountability and the Right to Explanation
4 min read · updated August 3, 2026
“You have a right to an explanation when a machine decides about you” is one of the most repeated claims in this field and one of the most contested among the lawyers who study it. The disagreement is instructive, because it is really about what an explanation is.
The dispute over whether the right exists
The claim traces mainly to European data protection law, which contains provisions on decisions based solely on automated processing that produce legal or similarly significant effects, and language about providing meaningful information about the logic involved.
Whether that adds up to a right to an explanation of a specific decision has been argued in the academic literature from both sides. Wachter, Mittelstadt and Floridi argued in 2017 that it does not: the binding text provides for information about system functionality given in advance, and the stronger formulation appears in a recital, which is interpretive rather than operative. Selbst and Powles replied that this reads the text too narrowly and that “meaningful information” must be assessed by whether it actually enables the data subject to exercise the other rights the same instrument grants — which, for a specific adverse decision, requires something specific.
Both positions are held by serious people and the practical answer has been developing through supervisory guidance and case law rather than through the debate. Note also that the older and more concrete tradition is sectoral: consumer credit regimes in several jurisdictions have required a statement of the principal reasons for an adverse decision for decades, long before anyone framed it as an AI question. That body of practice is the better guide to what a workable explanation duty looks like.
What the duties actually attach to
Reading across the regimes that impose something in this family, the obligations are usually one or more of four distinct things, and conflating them is where compliance projects go wrong:
- Notice — that automated decision-making is in use at all, given in advance, generically.
- Information about the logic — a description of the factors and their general role, at system level.
- Reasons for this decision — the principal factors that drove this outcome for this person.
- A route to contest — human intervention, the ability to express a point of view, and a challenge. This is often the operative right, and it is the one that changes outcomes; an explanation is instrumental to it.
The fourth is worth dwelling on. If the goal is that a wrong decision can be corrected, a mediocre explanation with a real appeal beats an excellent explanation with none. Regimes that focus on interpretability and skip the appeal route are optimising the wrong term.
Four things called an explanation
| Kind | Description |
|---|---|
| global | How the system works in general: inputs, model family, the role of each factor. Satisfies notice duties. Tells an individual almost nothing about their own case. |
| local | Which features drove this particular output. Feature-attribution methods produce these; they are approximations of the model with known instability, and two methods can disagree on the same case. |
| counterfactual | What would have had to be different for the outcome to change. Wachter and colleagues argued in 2018 that this satisfies most of the purposes of an explanation — understanding, contestation, and knowing what to do next — without disclosing the model. It is the most practical form for adverse decisions. |
| process | What data went in, which version decided, what threshold applied, who reviewed it. Not an explanation of the model at all, and often the most useful thing a person can be given. |
Mapping the two lists onto each other resolves most compliance confusion. Notice duties want the global form. Information-about-logic duties want the global form with more detail. Adverse-decision reason requirements want the local or counterfactual form. Contestation rights are served best by the process form, because what a person needs in order to challenge a decision is usually to discover that an input was wrong — and that is a data question, not an interpretability one. Teams that begin by trying to make the model interpretable have usually skipped this mapping and are solving the hardest version of a problem the rule did not pose.
Why faithful explanation is hard
A specific trap for language-model decisions. If you ask a model why it produced an answer, it will produce a fluent account. That account is generated by the same next-token process as the answer, conditioned on the answer — it is a plausible explanation, not a readout of the computation.
This is not speculation. Published work on chain-of-thought faithfulness, including Turpin and colleagues in 2023, has shown that models can be influenced by a feature of the prompt while producing reasoning that never mentions it and instead cites acceptable considerations. The stated reasons and the operative cause came apart, measurably. Presenting such text to a person as the reason for a decision about them is, at best, a misrepresentation of what it is.
Feature attribution methods have their own well-documented problems: sensitivity to baseline choice, disagreement between methods, and vulnerability to models that can be constructed to have innocuous-looking attributions. And there is a genuine tension nobody has resolved — an explanation detailed enough to let someone contest a decision is often detailed enough to let someone game it, which is a real objection raised by fraud and risk teams rather than an excuse.
What to build regardless
Independent of which regime applies, the artefact that makes accountability possible is a record. For each consequential decision: the inputs as received, the model and version, the prompt or feature vector actually sent, the raw output, the threshold or rule applied to it, any human review and what that reviewer saw, and the outcome delivered. Keep it for as long as the decision can be challenged.
That record answers the contestation question without solving interpretability, supports the counterfactual form of explanation, survives a change of model, and is the thing you will be asked for. Its absence is the most common reason an organisation cannot defend a decision it made correctly.
A warning about the other half of these regimes. Many of them soften or switch off when a human is involved, which creates an incentive to add a nominal reviewer rather than a real one. The literature on automation bias is unambiguous that people under time pressure, shown a confident recommendation, mostly confirm it — so a review step with no time budget, no authority to overturn and no record of overturns is a compliance artefact rather than a safeguard. If you rely on human involvement to change your legal position, the thing to be able to show is the overturn rate. A reviewer who never disagrees is evidence that the decision was automated after all.
Legal duties in this area differ by jurisdiction and sector and are actively developing. This page describes the shape of the arguments, not the state of any law, and it is not legal advice.