Explainability in AI refers to the degree to which a human user can understand how and why a model produced a specific output or decision. It is an essential concept in AI assurance because it enables transparency, fosters trust, and supports accountability — especially when AI is used in high-stakes settings like healthcare, law enforcement, or critical infrastructure.

There are two main dimensions of explainability:

• Global explainability: Understanding the overall logic or structure of the model.

• Local explainability: Understanding why the model produced a particular output for a specific input.

Explainability becomes particularly important when users must be able to challenge or rely upon an AI system’s decision — such as when determining medical treatments, approving loans, or assessing threats. Lack of explainability can obscure errors, hide bias, or hinder oversight.

In AI assurance, explainability is assessed by:

• Evaluating whether the system can provide intelligible outputs to different stakeholders (e.g., end users, auditors, regulators).

• Reviewing the use of interpretable models (e.g., decision trees) or model-agnostic techniques (e.g., SHAP, LIME).

• Testing whether explanations are stable, consistent, and contextually meaningful.

Explainability must also consider the audience. A technical explanation suitable for a developer may not suffice for a frontline operator or regulator. Therefore, assurance often involves validating multiple explanation formats and delivery methods.

Explainability is closely linked with other assurance criteria such as fairness (are the reasons for differential outcomes clear?), auditability (can decisions be traced?), and accountability (can someone be held responsible?).

Legal frameworks such as GDPR and the EU AI Act emphasise the right to an explanation, particularly for automated decisions with significant impact. International standards (e.g., ISO/IEC 24029) also define requirements and methods for explainability evaluation.

While complex models like deep neural networks often lack inherent explainability, assurance efforts must bridge that gap. This may involve trade-offs between performance and interpretability or hybrid approaches combining high-performance models with post hoc explanations.

In sum, explainability ensures that AI systems are not black boxes. It enables meaningful human oversight, supports ethical deployment, and provides the foundation for trust and transparency in AI-driven decisions.