What information must an explanation of an AI-assisted decision…

Direct answer

An effective explanation first states the exact decision and its effect on the person. It then shows the roles of AI and people, the material data used and its source, and the rules, thresholds and uncertainties that affected the result. Missing or correctable information and the route for challenge must also be clear. 'The algorithm calculated it' does not explain why this result was reached.

In plain language

If an application is refused, 'the system found you ineligible' is not useful. You need to know which information was considered wrong or missing, which rule was applied and what you can correct or contest.

Why this matters

Without the material reason, a person cannot correct inaccurate data, provide counter-evidence or challenge the decision meaningfully. A polished but irrelevant account creates only the appearance of transparency.

Do not confuse

A general system explanation describes how the model works broadly; a decision-specific explanation describes the particular result.
A data explanation identifies the information used; a factor explanation shows how it affected the result.
An explanation gives the grounds for a decision; it does not automatically make the decision correct or lawful.
Disclosure supplies information; an effective challenge explanation must be sufficiently actionable.
Confidential or security-sensitive information may be protected, but the material reason cannot be hidden completely.

What should you do?

State the decision identifier, date, scope, effect and whether it remains in force.
Separate the roles of the AI system, human reviewer and final decision-maker.
Identify the material data used, its source and date, and how it was linked to the person.
Explain the factors, rules, thresholds and material counter-evidence that affected this result.
Make uncertainty, missing information and the limits of the explanation visible.
State what data can be corrected, what new evidence can be submitted, the deadline and the secure challenge route.
Test the explanation against the actual decision record and for intelligibility to the intended reader.

How do you audit it?

Can the person understand which decision and effect apply to them?
Are the boundaries between AI and human roles described accurately?
Can the material data, source, date and entity match be seen?
Are the factors and rules that actually affected the result explained?
Was a general model description presented as a decision-specific explanation?
Does the person know what to correct and where to submit a challenge?
Were privacy or trade-secret grounds used to conceal the material reason completely?

Limit

The required scope of an explanation varies by jurisdiction, decision type and risk. Security, third-party rights and legal privilege may limit some detail. The existence of an explanation is not, by itself, proof of accuracy, fairness or legal compliance.