Natural Language Explanations (NLE)

The generated text must accurately reflect the model's true internal reasoning process, not just a plausible-sounding story. A faithful rationale exposes the actual feature weights, attention patterns, or logical rules used to arrive at the prediction. This is distinct from plausibility, which only measures human acceptance. Faithfulness is verified through erasure tests—if removing the cited evidence changes the prediction, the rationale is likely faithful. Without faithfulness, NLE becomes a sophisticated form of confabulation that erodes audit trust.

Every claim in the rationale must be explicitly linked to verifiable segments of the input data or external knowledge sources. This involves precise evidence attribution where the model points to specific tokens, passages, or structured fields that support its reasoning. Source grounding enables human auditors to trace the logical lineage of a decision and is a core requirement under the GDPR Right to Explanation. Ungrounded rationales are indistinguishable from hallucinated justifications.

Effective NLE answers not just 'why this?' but 'why this instead of that?' Contrastive explanations highlight the minimal set of features that differentiated the predicted outcome from a relevant alternative. This approach mirrors human reasoning and dramatically improves simulatability—the user's ability to predict the model's behavior on new inputs. A good contrastive rationale identifies the tipping point where the decision would have flipped.

The rationale must not contradict established world knowledge or the provided source data. Factual consistency is a hard constraint: an explanation that invents a non-existent regulation or misstates a data point is worse than no explanation at all. This requires robust hallucination detection pipelines that cross-reference generated claims against a knowledge base. In high-stakes domains like medicine or law, factual errors in explanations can create liability.

An explanation is actionable if the user understands what to do differently to achieve a desired outcome. This is the core of counterfactual rationales: 'If your credit score had been 20 points higher, the application would have been approved.' Actionable NLE closes the loop between interpretation and recourse, transforming the explanation from a passive artifact into an active tool for user agency and system improvement.

A single explanation format does not serve all audiences. User-adaptive explanations tailor the technical depth, vocabulary, and structure to the recipient's role:

• Engineers need feature attribution scores and gradient maps
• Compliance officers need regulatory rule mappings and audit trails
• End-users need plain-language summaries and actionable next steps This adaptability is governed by an organization's explanation policy, which defines the format and timing of justifications for each stakeholder class.