Right to Explanation

The explanation must be intelligible to the average data subject, not just a technical expert. It requires translating complex algorithmic logic into clear, concise, and plain language. This goes beyond merely providing source code or a full mathematical model; it must convey the rationale behind a specific decision. For example, instead of stating 'a neural network with 12 layers processed your data,' a meaningful explanation would be 'your loan was denied because your debt-to-income ratio exceeded the threshold, and your credit history showed a recent delinquency.'

The right applies to the logic involved in a specific, individual decision, not just the general functioning of the system. The explanation must clarify which features of the individual's data were most influential in producing the contested outcome. This is operationalized through techniques like counterfactual explanations, which identify the minimal change in input features needed to achieve a different result (e.g., 'Your credit limit would have been approved if your reported income was $5,000 higher').

The explanation must proactively describe the foreseeable effects of the automated processing for the data subject. This is a forward-looking requirement that contextualizes the decision's impact. For instance, an explanation for an automated employee scheduling system must not only explain how a shift was assigned but also the consequences, such as potential impacts on overtime pay, benefits eligibility, or work-life balance.

A complete explanation must inform the individual of their right to human intervention and the process to contest the decision. This transforms the explanation from a passive notification into an active gateway for remedy. The information provided must be sufficient for the data subject to formulate a substantive challenge. This includes details on how to contact a human reviewer and the expected timeline for a re-evaluation of the automated decision.

To satisfy both legal and technical requirements, explanations are often structured in tiered layers:

• Layer 1 (Public): A plain-language notice for all end-users, often part of a privacy policy.
• Layer 2 (Individual): A private, post-hoc explanation for a specific decision, detailing key feature weights.
• Layer 3 (Auditor): A deep technical disclosure, including a model card and fairness metrics, available to regulators or internal auditors. This layered approach balances transparency with the protection of trade secrets.

The right to explanation has a dual temporal scope:

• Ex-Ante (Before): The controller must provide upfront information about the existence of automated decision-making and the logic involved in the system's design.
• Ex-Post (After): The controller must provide a specific explanation for an individual decision that has already been made, enabling the data subject to understand and contest the outcome. This post-hoc requirement is the most technically challenging, often requiring SHAP or LIME analysis at inference time.

A causal method that identifies the minimal change to an input feature required to alter a model's prediction to a desired alternative outcome. Unlike feature attribution, counterfactuals answer the direct question: 'What would need to be different for me to get a different result?'

• Provides actionable recourse for individuals
• Does not require access to model internals
• Aligns with GDPR's 'meaningful information' standard
• Example: 'Your loan would have been approved if your income was $5,000 higher'

A game-theoretic framework for feature attribution that assigns each input feature an importance value for a particular prediction. Based on cooperative game theory, SHAP values guarantee consistent and locally accurate explanations by fairly distributing the prediction outcome among the input features.

• Unifies six existing explanation methods under one framework
• Provides both global interpretability and local explanations
• Handles correlated features through conditional expectation
• Widely adopted in financial services for credit decision explanations

The immutable recording of AI-driven decisions and their complete input context for auditability. Decision logs serve as the evidentiary backbone for fulfilling right to explanation requests, capturing the who, what, when, and why of every automated determination.

• Records model version, input features, and prediction confidence
• Enables retrospective explanation generation
• Supports chain-of-custody for regulatory investigations
• Must be tamper-proof with cryptographic integrity guarantees

A structured transparency document detailing a machine learning model's intended use, performance metrics across demographic subgroups, evaluation datasets, and known limitations. Model cards standardize ethical reporting and provide the foundational context required for meaningful explanations.

• Originated at Google Research in 2018
• Includes disaggregated evaluation by protected characteristics
• Documents out-of-scope use cases as explicit guardrails
• Serves as a public-facing complement to internal audit trails

A natively transparent architecture whose internal logic can be directly understood by a human without post-hoc approximation. Unlike black-box models requiring external explanation tools, interpretable models embed the right to explanation into their mathematical structure.

• Includes decision trees, logistic regression, and generalized additive models (GAMs)
• Every parameter and computation is directly inspectable
• Eliminates the fidelity gap between model logic and explanation
• Preferred in high-stakes domains like medical diagnosis and criminal justice

The design principle ensuring that individuals can effectively challenge, seek remedy for, or correct an automated decision through a formal appeal mechanism. Contestability transforms the right to explanation from a passive disclosure into an actionable procedural right.

• Requires accessible human review channels
• Mandates clear remedy pathways for adverse decisions
• Links explanation to correction mechanisms
• Embedded in Article 22 of GDPR and the proposed EU AI Act