AI Reliability

AI reliability (Concern 6)

Like validity, reliability is an absolute requirement for any test. The Uniform Guidelines require reliability to be evaluated and reported (§§14C(5), 15B(7), 15B(8)); the Principles (p. 22) require predictor scores to exhibit adequate reliability and to identify the conditions of measurement across which one wishes to generalize.

For how to estimate reliability, use criterion-related-validation (predictor/criterion reliability) and ai-model-outputs-audit.

What reliability requires of an AI tool

To be reliable, scores must measure KSAOs that are relatively stable and consistent over time and setting:

• An applicant's score should be about the same if they retook the assessment a week later (memorization notwithstanding).
• Scores should not be contingent on day or time of testing, testing conditions, or particular equipment unless those are job relevant.
• Reliable changes in scores should reflect changes in the job-relevant construct (gained skill) or reduced irrelevancy (less anxiety, better understanding of the protocol) — not artifacts.
• A game score should not reflect idiosyncrasies of the game or baseline skills beyond what's normally required (e.g., multikey or mouse facility).

The evidence is thin and mixed

Reliability evidence for technologically enhanced measures is often minimal or difficult to obtain; when available, results are mixed.

• Machine-scored interviews: some evidence shows reliability is high — sometimes higher than interviewer-scored — with substantial agreement with well-trained humans. Because human scorers in operational settings are notoriously prone to error and low interrater agreement, there's long-recognized potential for greater reliability and fairness in algorithmic scoring (Kuncel et al., 2013) so long as job-relevant information is what's being scored.
• Facial-feature / emotion analysis: reliability is particularly questionable. Even if facial expressions can reliably signal categories of emotion in human judgment (Cowen & Keltner, 2019), studies suggest the faces of people with darker skin are harder to evaluate via AI (Singer & Metz, 2019). It is unclear how the tool treats people with injuries, disabilities, medication effects, or altered features (scars, tattoos) — a reliability and fairness problem.

Confounds from how the data are generated

Threats to reliability can come from individual differences in the data collected, not just how they're measured. For example, more extraverted applicants provide more detail in written or spoken responses, giving the algorithm more key words that may relate to other traits (cognitive ability, verbal skills) not necessarily related to extraversion. So big data, ML, or both can end up affecting the reliability of other traits the tool claims to measure — a subtle source of construct-irrelevant variance.

Questions to ask (from the article)

• How should reliability be assessed when AI is used to build the predictive models used for selection?
• What should the minimal requirements for documentation of reliability be?
• Are appropriate and sufficient measures of reliability available and reported for technologically enhanced assessments?
• Are the new forms of employment assessment sufficiently reliable to meet psychometric standards?

Pitfalls

• Assuming algorithmic scoring is automatically reliable (it depends on what is scored).
• Ignoring stability across occasions/conditions (no test–retest evidence).
• Overlooking lower reliability of facial analysis for darker skin, disabilities, or altered features.
• Missing verbosity/extraversion confounds that contaminate other trait scores.
• Accepting "the model is consistent" without documentation of the conditions of measurement.

Checklist

• Reliability evaluated and reported for the conditions of intended use
• Score stability over time/occasions demonstrated (re-administration / re-scoring)
• Job-irrelevant dependencies (day, equipment, game idiosyncrasies) ruled out
• Facial/voice analysis reliability checked across skin tone, disability, altered features
• Data-generation confounds (e.g., verbosity → other traits) examined
• Documentation of reliability sufficient to meet UGESP/Principles/Standards

See also

ai-validity-evidence · ai-ml-methodology-evaluation · ai-dynamic-models-and-revalidation (stability vs. frequent updating) · ai-candidate-data-control (disability/appearance fairness) · criterion-related-validation (reliability estimation) · fairness-and-bias-analysis · ai-model-outputs-audit

Source: Tippins, Oswald & McPhail (2021), Concern: "Reliability."