Risk scoring and fairness in audit selection

Why this matters

No revenue department can examine every return. Choosing which cases to look at closely has always involved judgement, rules of thumb and sometimes simple scoring. AI and predictive analytics promise to make that choice faster and more consistent by finding patterns in large amounts of data.

Selection for audit is not a neutral act. It costs the taxpayer time, money and sometimes reputation. If a tool selects some groups more often for reasons unrelated to real risk, it can be unfair even when no officer intends it. You may never build such a tool, but you may use its output, supervise people who do or be asked to explain a selection to a taxpayer. This module gives you the questions to ask.

How risk scoring works, in plain terms

A scoring tool is usually trained on past cases. It looks at features of returns or taxpayers and learns which patterns were associated with past findings. It then gives new cases a score or a flag.

Three things follow from this:

• It learns from the past. If past audits focused on certain sectors, areas or types of business, the tool may learn to find problems there simply because that is where people looked.
• It finds correlation, not cause. A pattern linked to past findings is not proof that a particular taxpayer has done anything wrong.
• It can drift. As the economy, the law or taxpayer behaviour changes, a tool trained on older cases can become less accurate without anyone noticing. This is model drift.

Bias: where it comes from

Bias in a scoring tool can enter through:

• Who was audited before. Past selection patterns become the tool's idea of risk.
• Proxy features. Even if a tool does not use a protected characteristic directly, other data such as location, language of filing or type of business can stand in for it.
• Data gaps. Taxpayers with thin or poor records, often smaller or less formal businesses, may be scored unreliably.

Questions to ask: which groups are flagged most often and does that match what audits actually find? Has anyone checked whether the tool treats similar taxpayers differently because of where they are or who they are?

Explainability

If an officer cannot say why a case was selected, the officer cannot judge whether the selection is sensible and the taxpayer cannot meaningfully respond. Explainability means being able to give the main reasons for a score in plain words.

At a minimum, an officer reviewing a flagged case should be able to see which factors raised the score. If a tool gives only a number with no reasons, treat it with extra caution and raise the question with whoever is responsible for the tool.

Human review

A score should start a human process, not replace it. Good practice includes:

• An officer reviews every flagged case before any action that affects the taxpayer, looking at the actual return and records.
• The officer can disagree with the tool and record why, without needing special permission.
• Decisions are recorded with the officer's own reasons, not just "selected by system". An audit trail helps later review.
• Results are fed back. Cases where the flag was wrong are as important as cases where it was right, because they show where the tool is weak.

This is what human oversight means in practice: a person with the authority, time and information to make a real decision.

Appeals and redress

Taxpayers have rights under the law that governs your work, including routes to object and appeal. AI does not change those rights. A selection or decision supported by a tool must still be explainable to the taxpayer and open to challenge through the normal routes. Check your department's procedures for what must be disclosed and how. If a tool is found to have selected cases unfairly, the department should be able to identify and review the affected cases. That is part of redress and accountability.

Try it

Using no real data, write down five features a scoring tool might use to flag a return, such as late filing or a large change from last year. For each, ask: could this feature stand in for location, business size or another characteristic that should not drive selection? Which features would you want explained before acting on a flag?

Check yourself

1. Why might a tool trained on past audits repeat past patterns?
2. What should an officer be able to see when reviewing a flagged case?
3. Name two features of good human review of AI-flagged cases.

Answers: (1) It learns its idea of risk from where people looked before, so past selection patterns become part of the model. (2) The main factors that raised the score, in plain words. (3) Any two of: an officer reviews every case before action, the officer can disagree and record why, decisions are recorded with the officer's reasons, results are fed back.