AI Model Outputs Audit

AI model-outputs audit (Component 6)

The payoff component: regardless of how the model was built, are its scores good? This is where the AI audit reconnects most directly to classic psychometrics and the SIOP Principles — apply the personnel-selection skills here, on the model's actual operational outputs.

What it is: evaluation of the quality of the predictions the model generates.

Questions to ask: How was prediction quality evaluated — e.g., for psychometric reliability and validity? How was cross-validation conducted, and was it appropriate given the claims about model generalizability?

Apply it (focal example):

• Reliability — Are scores consistent over time and upon multiple administrations? A candidate who records the same interview twice, or whose video is scored repeatedly, should get stable scores.
• Validity — Is there evidence the scores reflect the predicted constructs they claim to, and relate to the intended outcome? Don't accept a predictive R² as construct validity.
• Subgroup differences — Do scores show differences among classes of interest — race, gender, color, national origin, religion, disability, age — and combinations of classes (intersectional subgroups)? AI's high-dimensional inputs make intersectional effects both more likely and easier to overlook.

How to audit outputs (bring the Principles to bear)

1. Reliability evidence. Identify the conditions you must generalize across (occasions, repeated administrations, alternate prompts, re-scoring) and estimate the matching reliability. For stochastic models or pipelines with random elements, test score stability under re-run. See criterion-related-validation (reliability of measures).
2. Validity evidence. Treat the model as a selection procedure: assemble criterion-related and/or construct evidence that scores predict the defensibly defined outcome (tie back to the Component-2 criterion audit). Beware overfitting: a cross-validated estimate beats holdout, which beats temporal — confirm the cross-validation matches the generalizability claim (a tool used on future applicants needs evidence that survives temporal validation, not just k-fold). See criterion-related-validation and internal-structure-validation.
3. Subgroup / bias analysis — run the right test for the right lens. A subgroup mean difference (adverse impact) is a scrutiny trigger, not a verdict of bias (Lens 2/3).
   • Predictive bias / differential prediction — test on the operational score as used (the composite the model outputs), not internal sub-features; ask whether a subgroup is underpredicted, watch power, range restriction, and error-variance homogeneity. See fairness-and-bias-analysis.
   • Measurement bias / invariance & DIF — check whether score meaning is invariant across groups; remember psychometric bias may be expected and "fair" when groups truly differ on the construct (ai-fairness-lenses, Lens 3).
   • Intersectional subgroups — analyze combinations of classes where data and power allow; report what couldn't be analyzed.
4. Generalizability check. Confirm the validation sample and cross-validation support the population where the tool is deployed (the incumbent-vs-applicant / range-restriction problem from Component 1 resurfaces in the output evidence).

Connecting to the developer's claim

The focal developer claims the algorithm "predicts job performance equally well for all groups using appropriate modeling techniques." That claim is evaluated here — and the paper calls it questionable and in need of audit. Translate it into testable pieces: (a) equal predictive accuracy across groups, (b) no problematic differential prediction, (c) appropriate modeling — and require evidence for each.

Pitfalls

• Accepting predictive accuracy as evidence of construct validity.
• Reporting k-fold validity when the deployment claim requires temporal generalization.
• Testing bias on internal features instead of the operational output score.
• Treating a mean difference as proof of bias, or "no measurement bias" as proof of fairness.
• Ignoring intersectional subgroups because each is small (note the limitation explicitly).
• Skipping score-stability/reliability checks for stochastic pipelines.

Checklist

• Reliability estimated for the conditions of intended use; re-run/re-score stability tested
• Validity evidence ties scores to the defensibly defined criterion (not just R²)
• Cross-validation method matches the generalizability claim (temporal where deployment drifts)
• Predictive bias tested on the operational output, framed as underprediction, with adequate power
• Measurement invariance / DIF considered; expected-vs-problematic differences distinguished
• Subgroup differences examined across protected classes and intersections; gaps reported
• Developer's "equal prediction for all groups" claim decomposed and each part evidenced

See also

ai-input-data-and-design-audit · ai-model-development-audit · ai-fairness-lenses · criterion-related-validation · internal-structure-validation · fairness-and-bias-analysis · selection-decisions-and-scoring

Source: Landers & Behrend (2023), Table 1 (Component 6).