Counterfactual Fairness

A practical guide to counterfactual fairness for AI audit and assurance practitioners.

What This Lesson Covers

Counterfactual Fairness is a key topic within Fairness Audit. In this lesson you will learn the underlying audit and assurance discipline, the controlling standards and frameworks, how to apply the procedures to real AI systems, and the open questions practitioners are actively working through. By the end you will be able to engage with counterfactual fairness in real AI audit and assurance work with confidence.

This lesson belongs to the Technical AI Audits category of the AI Audit & Assurance track. AI audit sits at the intersection of internal audit, IT audit, model risk management, AI governance, and emerging conformity-assessment regimes. Understanding the underlying discipline is what lets you build audit programs that survive board scrutiny, regulator inquiry, and certification audits.

Why It Matters

Conduct a fairness audit beyond bias metrics. Learn procedural fairness (process), distributive fairness (outcomes), counterfactual fairness, and intersectional analysis.

The reason counterfactual fairness deserves dedicated attention is that AI audit and assurance is a young discipline whose standards are landing every quarter (ISO/IEC 42001 audits going live, EU AI Act conformity assessment, AICPA AI assurance, ISACA AI audit toolkit, NYC AEDT bias audits, CO AI Act assessments). Auditors and management who can reason from first principles will navigate the next standard or attestation requirement far more effectively than those who only know current rules.

How It Works in Practice

Below is a practical AI audit framework for counterfactual fairness. Read through it once, then think about how you would apply it to a real engagement on an AI system in your portfolio.

# Fairness audit beyond bias metrics
FAIRNESS_LENSES = {
    "procedural": [
        "Was protected attribute considered in design?",
        "Was an impact assessment performed?",
        "Was an opportunity to contest/appeal provided?",
        "Was the model documented (model card / FRIA)?",
    ],
    "distributive": [
        "Did outcomes differ by protected group beyond model error?",
        "Are benefits/harms equitably distributed?",
        "Are second-order effects considered (e.g., feedback loops)?",
    ],
    "counterfactual": [
        "If protected attribute were different, would outcome change?",
        "Implementation: Kusner et al. CFR / matched-pair tests",
    ],
    "intersectional": [
        "Sub-groups defined by 2+ protected attributes",
        "Crenshaw lens: hidden harms when categories combine",
        "Power analysis to ensure adequate sample for sub-groups",
    ],
}

TRADE_OFF_DOCUMENTATION_TEMPLATE = '''
Fairness criterion selected: 
Reason: 
Other criteria considered + rejected: 
Trade-off accepted: 
Approver + date: 
Re-evaluation date: 
'''

Step-by-Step Analytical Approach

1. Establish the criteria — What standard, framework, or policy will this audit measure against (NIST RMF, ISO 42001, EU AI Act Article 9, internal policy, contractual commitment)? Document the criteria up front; auditing without explicit criteria is opinion, not assurance.
2. Plan the procedures — Map criteria to procedures (inquire, observe, inspect, recalculate, reperform, analytics). For AI specifically, prefer reperformance (rerun the eval) over inquiry (“trust the team”).
3. Sample appropriately — Statistical for control-pass-fail tests, judgmental for corner cases, stratified for fairness, adversarial-seed for robustness. Document the sampling rationale.
4. Collect sufficient appropriate evidence — Multiple sources, time-stamped, hash-pinned, secured. The bar is what a sophisticated reviewer would expect to support the conclusion.
5. Form the conclusion — Compare evidence to criteria; identify exceptions; quantify if possible; classify by severity.
6. Communicate and track — Findings + recommendations + management response; tracker through validated closure; periodic aging report to audit committee.

When This Topic Applies (and When It Does Not)

Counterfactual Fairness applies when:

• You are providing assurance over AI systems (internal audit, external audit, certification, regulator)
• You are subject to a standard that requires AI audit (EU AI Act conformity, ISO 42001 certification, sector regulator audit)
• You need to demonstrate AI controls operate effectively to the board, customers, regulators, or in litigation
• You are consuming third-party AI assurance reports (SOC 2, ISO 42001 certificate, AICPA attestation)

It does not apply (or applies lightly) when:

• The activity is design-stage advisory rather than independent assurance
• The AI system is genuinely low-stakes with no audit obligation
• The work is consulting / co-sourcing rather than independent audit (independence rules differ)

Practitioner Checklist

• Are the criteria for this engagement explicit, written, and agreed with management?
• Are procedures designed to give sufficient appropriate evidence (not just inquiry)?
• Is the sample defensible (rationale documented, stratified where relevant)?
• Is evidence preserved with integrity (timestamp, hash, immutable storage)?
• Are findings traceable from evidence to criteria to conclusion?
• Do you have a written management response with owner and due date?
• Is closure validation tested, not self-attested?

Disclaimer

This educational content is provided for general informational purposes only. It does not constitute audit, legal, regulatory, or professional advice; it does not create a professional engagement; and it should not be relied on for any specific audit, certification, or compliance matter. AI audit standards and regulations vary by jurisdiction and change rapidly. Consult qualified professional auditors and counsel for advice on your specific situation.

Next Steps

The other lessons in Fairness Audit build directly on this one. Once you are comfortable with counterfactual fairness, the natural next step is to combine it with the patterns in the surrounding lessons — that is where doctrinal mastery turns into a working audit program. AI audit is most useful as an integrated discipline covering planning, fieldwork, evidence, conclusion, reporting, and remediation.