What an explanation can actually be

The right, stated carefully

Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing which produces legal effects or similarly significantly affects them, with exceptions — contract necessity, legal authorisation, explicit consent — and in the exception cases requires safeguards including the right to obtain human intervention, to express a point of view and to contest the decision.

Whether it contains a "right to explanation" is genuinely contested among lawyers. The operative articles require meaningful information about the logic involved; the phrase "explanation of the decision reached" appears in Recital 71, which is not binding in the same way. The Court of Justice's 2023 SCHUFA ruling took a broad view of what counts as an automated decision, holding that producing a credit score on which a third party draws heavily can itself be the decision. The direction of travel is towards more explanation rather than less, and the precise scope is still being litigated.

India's DPDP Act does not contain an equivalent automated-decision provision. That absence is one of the sharpest practical differences between the two regimes.

What the technical field can produce

Suppose an organisation wants to comply. What can it actually generate?

Feature importance. Which inputs mattered most, globally or for this case. Methods include SHAP, based on cooperative game theory, and LIME, which fits a simple model locally around the case. Both are free, open-source Python libraries that run on any laptop.

Counterfactuals. The most useful form for a person: had your income been £2,000 higher, or had you not had the missed payment in March, the decision would have been different. This is actionable in a way that a ranked list is not, and the leading academic argument is that counterfactuals satisfy most of what people actually want from an explanation without exposing the model.

Rule extraction and surrogate models. Fit an interpretable model that approximates the complex one, and explain that. Cheap, and it explains the surrogate rather than the system.

An inherently interpretable model. Sometimes the answer is not to explain a black box but not to use one. For many tabular decision problems, a well-built scorecard or small decision tree performs close to a complex model, and Cynthia Rudin's argument that high-stakes decisions should use interpretable models by default is the strongest version of this position.

Why a feature list is not a reason

Here is the crux, and it survives every technical advance.

"The three factors that most influenced this decision were your postcode, your account age and your device type" is not a reason. It is a description of a computation. A reason is something a person can evaluate as fair or unfair, contest as factually wrong, or act upon. A feature list gives you none of the three, because it does not say what value you had, what value would have sufficed, or why that factor is legitimate.

There is a further problem: the explanation methods themselves are approximations, and different methods can produce different attributions for the same decision. An institution that wanted to could select the explanation that looked most defensible. Researchers have demonstrated this vulnerability directly, constructing models whose behaviour is discriminatory while their generated explanations appear benign.

So an explanation is not a technical artefact you can trust in the way you trust an audit. It is a claim, made by the same organisation whose decision is in question.

What to ask for

If a decision has gone against you, four requests are worth making in writing, and they are more effective than "explain your algorithm".

1. Was this decision made solely by automated means? A yes engages specific rights in several jurisdictions; a no invites the next question, which is who the human was and what they saw.
2. What personal data about me was used? An access request, with a statutory deadline.
3. What would have had to be different for the decision to go the other way? The counterfactual request. It is answerable, and refusal to answer it is itself informative.
4. How do I obtain human review, and by when?

Those four fit in a short letter, cost nothing, and create a record. The last lesson in this module is how to write it.