OpenMatter-Network avatar

ai-selection-legal-ethical

Evaluating AI-based selection tools against scientific, legal, and ethical standards (Uniform Guidelines, Title VII, APA Ethics Code), per Tippins, Oswald & McPhail (2021).

GitHub stars
10

Install

/plugin install ai-selection-legal-ethical@io-skills

Run it inside Claude Code after adding the marketplace with /plugin marketplace add OpenMatter-Network/agent-io-skills.

Contents

Discovered from source

Skills 12

AI applicant reactions & communications (Concerns 9–10)

Two linked concerns about information flow around an AI selection tool: how applicants experience and react to it, and what the organization tells candidates and other stakeholders.

Applicant experience and reactions (Concern 9)

Employers want selection that is simple, quick, and engaging to attract qualified candidates, and technologically enhanced assessments are often highly engaging with little applicant effort. But innovative methods raise reaction concerns.

Pitfalls in reaction metrics

Many vendors collect applicant-reaction data, but the metrics are weak:

  1. They rarely incorporate the full range of organizational considerations — vendors ask if the experience was engaging, but seldom whether applicants felt job-relevant KSAOs were measured.
  2. Good comparative data are rarely available — reactions are used as a marketing tool, seldom compared to reactions to other tools.
  3. It's hard to know the local range of reactions (how positive/negative) even with meta-analytic work (Hausknecht et al., 2004).
  4. Reactions are affected by how well candidates think they performed — harder to gauge for novel games or tools with no obviously correct answers.
  5. Whether a job offer followed is a huge driver of reactions — so consider measuring reactions before offers are extended.
  6. Simply asking about the experience afterward may alter perceptions (e.g., prompting reflection on fairness/invasiveness of a video interview or scraped data).

Impact on well-qualified candidates and the faking question

The applicant reaction–behavior link is tenuous — the "Achilles heel" of applicant-reactions research (Sackett & Lievens, 2008). Still, organizations worry about effects on the quality and quantity of applicants they attract. It's unclear how candidates react on learning their selection hinged on an unknown weighted combination of facial expressions, voice quality, mouse clicks, and other data — versus, say, their MBA from a top school. Reactions may matter more now because applicants amplify them via social media (Twitter, Facebook, LinkedIn). A largely unresolved issue: whether training for a video interview is possible, and if so, whether it produces invalid variance (faking/lying) or valid variance (by ensuring candidates understand what's expected). Some organizations sidestep specifics by simply informing candidates whether the outcome indicates they met the employer's needs.

Questions to ask (Concern 9)

  • How should applicants prepare for the assessment (are practice sessions allowable)?
  • What is being measured, is it relevant to the job, and does the applicant know it's being measured?
  • Why was an applicant not selected — can big data and the ML algorithm provide explanations?
  • How can applicants improve to become more qualified upon retesting?
  • How do applicants evaluate organizations that minimize personal interaction (chatbots, avatars)?
  • What are applicant reactions to innovative approaches, and how do they affect the employer's ability to attract qualified candidates and its reputation?

Communications (Concern 10)

Managers (whose success depends on a competent workforce) care that job-critical skills are being measured; labor organizations and advocacy groups care about job relevance and fairness; and enforcement officials have a statutory/regulatory interest in what is measured and how.

Sharing selection-procedure information

A key question is what to tell people about how others were selected. There have always been limits:

  • Organizations are unwilling to share anything that jeopardizes test use (item content, scoring keys) or increases legal/administrative challenge (e.g., adverse-impact data).
  • The technical aspects of evaluating measurement or predictive bias are beyond the comprehension of most applicants and hiring managers (regression slopes, factor loadings), and bias in ML models is even harder to explain.
  • Most applicants want to know at a basic level: what KSAOs are measured, how they're evaluated, and — if unsuccessful — what they can do to improve next time.
  • Test-prep materials typically describe the process, give tips, and sometimes offer practice questions; but it's unclear what preparation can be offered when selection rests on face or voice characteristics.

Questions to ask (Concern 10)

  • What information can and should be provided to unsuccessful applicants?
  • What aspects of a selection procedure should an organization share with a range of stakeholders (manager, industry, clients, customers, shareholders)?

Pitfalls

  • Treating "engaging" reaction data as evidence of fairness or job-relevance perceptions.
  • Using vendor reaction stats (marketing) without comparative or local data.
  • Ignoring that AI decisions are hard to explain — leaving rejected candidates with no actionable feedback.
  • Encouraging "training" for video interviews without resolving whether it introduces faking.
  • Over-sharing (compromising test security) or under-sharing (eroding trust/justice).

Checklist

  • Reaction measurement captures job-relevance perceptions, not just engagement
  • Reaction data benchmarked (comparative/local), not taken as marketing
  • Offer-status and timing effects on reactions accounted for
  • Faking-vs-training question for any video/behavioral interview addressed
  • Basic, honest information (what's measured, how, how to improve) prepared for candidates
  • Security-sensitive details (keys, items, adverse-impact data) protected
  • Stakeholder communication plan defined (managers, regulators, advocacy groups)

See also

ai-candidate-data-control · ai-selection-ethics · ai-selection-legal-landscape · ai-claims-and-stakeholder-audit (second-party effects, justice) · administration-documentation (candidate communications, feedback)

Source: Tippins, Oswald & McPhail (2021), Concerns: "Applicant Experience and Reactions" and "Communications."

AI candidate data control (Concern 8)

Traditionally, applicants control to a large degree what they present to an employer — effort on ability tests, answers on personality/SJT measures, demeanor in interviews, resume and application content. Using information outside those sources is not new ("word of mouth," references, background and credit checks). What's new with AI is the scale and the loss of applicant control.

What changed

  • Scraped data outside the applicant's control. Employers can search the Internet and large databases for evidence of "inappropriate behavior" (poor judgment), but such data often contain irrelevant information — demographics (Zhang et al., 2020), political affiliation (Roth et al., 2020). Applicants have increasingly less control over the type and relevance of personal data organizations extract from social media. They may try impression management via profiles (Schroeder & Cavanaugh, 2018), but in some cases they did not post the information themselves, it was substantially altered, or it was posted without intent to be shared. Online information is often suspect, dated, or lacking context.
  • Reputation-scrubbing services and adverse impact. Companies that help manage online reputations are not free; to the extent their availability/affordability varies by race/ethnicity or other demographics, these "scrubbing" services may contribute to adverse impact that is difficult to detect.
  • Incidental data that candidates cannot alter. Images, video, audio, facial micro-expressions, and voice purport to convey job-relevant emotions — but physical appearance beyond grooming is outside most people's control (skin color, voice timbre, basic speech patterns, the features of one's face). This raises special problems for people who look or speak differently due to cultural differences (minorities, immigrants), physical differences (disabilities, diseases, injuries), gender, and age.

The crux: "Is it fair?"

Irrelevant variables may well predict performance; the essential question is "Is it fair?" There is no law or guideline requiring an employer to use only data presented by the candidate (except in the realm of privacy statutes), and no specific ethical standard requiring it either. Yet there is a moral dilemma. The article lays out a spectrum of approaches:

  • One extreme — press ahead. Use this kind of data, on the rationale that applicants have never controlled everything an employer sees and uses.
  • Other extreme — use no data beyond the candidate's control. This would entail careful review even of traditional forms like biodata.
  • A moderate approach — informed consent. Being legislated in laws like the Illinois AI Video Interview Act (ai-selection-legal-landscape): require informed consent before an employer bases a selection decision on data beyond the applicant's control.

Questions to ask (from the article)

  • Is it fair to use data that are outside the control of an applicant?
  • Should employers seek out data on the Internet at all?
  • Would there be legal issues associated with not seeking information about some behaviors (e.g., poor judgment, behavioral deviancy, CWBs)?
  • How long should applicants' past failures or mistakes affect their future job prospects, and what mistakes should be considered (criminal history, online behavior, early-life behavior)?

Pitfalls

  • Using scraped data without confirming relevance, recency, accuracy, or that the candidate authored it.
  • Ignoring that paid reputation-scrubbing can create undetectable adverse impact.
  • Scoring appearance/voice features that candidates cannot change (disability/cultural/age fairness).
  • Assuming "no law against it" settles the fairness/ethics question (it doesn't).
  • Treating implied consent as covering data the candidate doesn't know is being collected (see ai-selection-ethics).

Checklist

  • Data sources classified by applicant control (provided vs. scraped vs. incidental/uncontrollable)
  • Job relevance of out-of-control data interrogated ("Is it fair?")
  • Accuracy/recency/authorship/context of scraped data verified
  • Adverse-impact risk from scrubbing-service disparities considered
  • Appearance/voice/disability/cultural/age fairness evaluated
  • Consent approach chosen along the spectrum; statutory consent (e.g., Illinois) satisfied
  • Privacy-statute compliance confirmed with counsel

See also

ai-selection-ethics (informed consent) · ai-applicant-reactions-and-communications · ai-selection-legal-landscape (Illinois AIVI Act, privacy) · ai-reliability (appearance/disability) · candidate-accommodations (disability, linguistic/cultural) · ai-claims-and-stakeholder-audit

Source: Tippins, Oswald & McPhail (2021), Concern: "Control Over the Data Presented to an Employer."

AI dynamic models & revalidation (Concern 7)

ML-derived selection procedures enable analysis that wasn't feasible before, and many vendors refresh the algorithm frequently — sometimes after every test administration. Such "dynamic" procedures update validation evidence and normative data in near real time. That creates a dilemma and several operational problems.

The core dilemma

When a dynamic model changes, the change could reflect either:

  • a real change in the nature of the applicant sample and/or the job (which should drive an update), or
  • sample idiosyncrasies or instabilities over time that should not be capitalized on.

Distinguishing the two is hard — and capitalizing on instability degrades the model.

Why "dynamic" raises problems

  1. Documentation must keep up. Validity evidence must be documented in a technical report. So every update to the underlying selection process requires updating the technical report (validity and normative-data characteristics). Out-of-date reports are problematic for legal defensibility and for HR records maintenance over time. (See technical-validation-report.)
  2. Score adjustments & grandparenting. Each new selection process may require adjusting scores already in the database, or a policy on how to treat scores produced by different processes at different times. Continual adjustment creates administration problems: a candidate qualified today may be unqualified tomorrow (or vice versa). Changes can include altering predictor weights, adding/removing predictors, and changing interpretations (e.g., adjusting cutoff scores). Grandparenting policies are administratively hard to manage in high-volume programs. (See selection-decisions-and-scoring.)
  3. Disparate-treatment / disparate-impact risk. A result of frequent change is that different candidates are evaluated on different variables depending on when they applied. If that variation relates to a protected characteristic, there's a disparate-impact specter; even without group-level disparate impact, disparate-treatment concerns can arise — and the appearance of such treatment can trigger applicant dissatisfaction and complaints.
  4. Applicant-pool shifts. Large shifts in the applicant pool can materially affect reliability, validity, range restriction, or range enhancement on big data — so monitor key applicant-pool characteristics (demographic, educational). Future shifts in available technologies/data/algorithms may change the applicant database and indicate the ML model should be updated.
  5. Defining group comparisons. Frequent algorithm changes make it harder to define relevant applicant pools for adverse-impact analysis or appropriate normative groups for comparison.

Revalidation & norms updating

Employers have always revalidated and updated norms; with AI the difference is the frequency. Traditionally, revalidation was triggered when the job changed, the test was compromised, the applicant pool shifted substantially, or enough time elapsed to question validity in a legal challenge — and it was undertaken at well-spaced intervals because it was laborious. Today's computing power makes continuous updating far less laborious, which raises a genuinely open question: how often should validation be refreshed to accommodate the nature of new applicant data?

Questions to ask (from the article)

  • Are dynamic algorithms and norms useful?
  • How should results from dynamic algorithms be documented to comply with existing and future legal and administrative requirements?
  • How frequently should tests be revalidated and norms updated?
  • What are the indicators that revalidation and updates to norms are needed?

Pitfalls

  • Capitalizing on sample instability and mistaking it for real change.
  • Letting technical reports/normative documentation fall out of date as the model drifts.
  • Unmanaged score adjustments → candidates flipping qualified/unqualified; messy grandparenting.
  • Evaluating different applicants on different variables → disparate-treatment exposure (and optics).
  • Not monitoring applicant-pool shifts that invalidate prior validity/range assumptions.
  • Unstable pool/normative definitions that make adverse-impact analysis incoherent.

Checklist

  • Each model/norm update justified as real change, not instability
  • Technical report and normative documentation updated with every change
  • Score-adjustment / grandparenting policy defined and administrable
  • Risk that candidates are scored on different variables (disparate treatment) assessed
  • Applicant-pool characteristics monitored for shifts affecting validity/range
  • Revalidation/norms-update cadence and trigger indicators defined
  • Adverse-impact pools and normative groups remain definable under the update regime

See also

ai-validity-evidence · ai-selection-legal-landscape · generalizing-validity-evidence · technical-validation-report · administration-documentation (review/updating, records) · selection-decisions-and-scoring (cutoffs, norms)

Source: Tippins, Oswald & McPhail (2021), Concern: "Changes to Technologically Enhanced Systems" (Dynamic models and norms; Revalidation and norms updating).

Extending professional standards (Call to Action)

The article's thesis and orchestration point: AI/technologically enhanced selection should be held to the same established professional standards as any other employment test, and I-O psychologists should lead in working out how. Use this skill to orient a project — or a professional-policy discussion — to the governing documents and the collaborative path forward. It ties the whole collection together.

The two guiding documents

Two documents guide research and practice in employee selection regardless of the form of assessment:

  • Principles for the Validation and Use of Personnel Selection Procedures (SIOP, 2018) — see the personnel-selection collection.
  • Standards for Educational and Psychological Testing (AERA, APA, NCME, 2014).

Both adopt the same definition of validity — the degree to which accumulated evidence and theory support specific interpretations of test scores for proposed uses — which is exactly why a technology is never "universally valid" (ai-selection-tech-data-algorithms) and why validity/reliability/fairness evidence is required for AI tools.

Why I-O psychologists are well-equipped — but not sufficient alone

I-O psychologists bring deep grounding in the factors critical for employment testing: psychological constructs (knowledge, personality, interests, engagement, teamwork, safety, performance, turnover), theories of testing and assessment (construct-oriented test development, psychometric modeling, appropriate scoring/interpretation), the types of evidence that support inferences (selection decisions, validity), psychometric properties (internal consistency, test–retest, alternate-forms reliability), and the evaluation of subgroup differences (differential prediction, measurement invariance, adverse impact). SIOP also has a long history of documenting consensus in the Principles.

But this knowledge must be supplemented by others in the field: data scientists and software developers (acquire/store/analyze data, build and evaluate algorithms), web designers and IT professionals (build engaging, effective interfaces), and the legal profession (compliance with federal/state/local law and regulatory requirements). I-O psychologists cannot regulate others' practice, but many serve as experts advising organizations and government and testifying about assessments — supporting and challenging them.

The Call to Action

  • Develop interpretive guidance — don't rewrite the Principles. The recommendation is for SIOP to develop interpretive guidance that applies the Principles to technologically enhanced assessments, guiding developers and users in best practices and addressing the open questions the paper raises. The Principles already reflect the established science of selection; the goal is interpretation and consistency, not replacement.
  • Collaborate across disciplines. Engage applied statistics, computer science, and other fields to learn about ML applications. Together, identify the strengths, critique the weaknesses, and understand appropriate vs. inappropriate applications. Interpretive guidance should help fill knowledge gaps among the participating parties.
  • Engage proactively — not only selection specialists, but also those in recruiting, diversity and inclusion, and leadership — because doing so can improve assessment and promote the future relevance of the profession.

The guardrail: no "escape velocity"

The overarching responsibility: ensure that progress does not approach escape velocity from its moorings in scientific, psychometric, and practical knowledge; understanding of legal guidelines and professional/ethical obligations; and the many hard lessons learned in the employment-testing arena. New tools offer real advantages for employers and applicants — and we are responsible for keeping them anchored. Now is the time to consider how the Principles should be applied to new and evolving forms of assessment to reflect the research literature and best practices.

How to use this skill

  • Orient any AI-selection project to the Principles and Standards as the benchmark, then route to the specific concern skills for the evaluation.
  • Frame professional/policy discussions around applying (not rewriting) the Principles and building interdisciplinary collaboration.
  • Audit your team composition: do you have psychometric, data-science, IT/UX, and legal expertise at the table?
  • Apply the escape-velocity test: is any practice drifting away from scientific, legal, or ethical moorings?

Pitfalls

  • Proposing to rewrite the Principles for AI rather than developing interpretive guidance that applies them — the established science of selection still holds.
  • Treating a technology as "validated" rather than validating the inferences about constructs measured in a specific use.
  • Assembling a team with psychometric expertise but no data-science, IT/UX, or legal voices (or vice versa) — the paper stresses no discipline suffices alone.
  • Inventing ad hoc, tool-specific rules disconnected from the Principles and Standards.
  • Letting innovation outrun (reach "escape velocity" from) scientific, legal, and ethical moorings in the name of efficiency.

Checklist

  • Project benchmarked against the Principles and Standards
  • Validity framed as evidence + theory for a specific inference (not a property of the technology)
  • Interdisciplinary expertise assembled (psychometrics + data science + IT/UX + legal)
  • Stance taken: interpret/apply the Principles to the tool, not invent ad hoc rules
  • Open questions from the 11 concerns logged for the developer/vendor
  • "Escape velocity" check applied to scientific, legal, and ethical moorings

See also

All skills in this collection (this is the orchestration point) · personnel-selection (the Principles operationalized) · ai-personnel-assessment (the audit framework) · validation-planning · technical-validation-report

Source: Tippins, Oswald & McPhail (2021), "Standards" and "A Call to Action," and the Conclusion.

AI job analysis & job relevancy (Concerns 2–3)

Professional standards, legal guidelines, and case law all require selection systems to be tied to job requirements, established through an analysis of work (Morgeson et al., 2020). For AI tools built from convenient or scraped data, this is frequently the weakest link. (The article treats these two concerns together because requirements for a job analysis are hard to separate from job relevancy.)

This skill is the AI-specific framing; for how to conduct a work analysis, use work-analysis.

Why job analysis matters here

  • Traditionally, job analysis is the basis for identifying the KSAOs required and the appropriate variables for selection, and for developing criteria / defining the performance domain.
  • Uniform Guidelines §14B(2) requires a review of job information for criterion-related validity; the Principles state job analysis defines appropriate predictors and establishes the relevance of the criteria.
  • In practice, many test users skip or abbreviate job analysis — assuming, say, that all sales jobs are alike and a generic "sales test" transfers. Without a job analysis you cannot know (a) that the test's purported skills are the ones predicting in this organization, or (b) that the skills the job requires match those the test measures. Predictors and criteria may then be wrong, inadequate, or unfair.

Is job analysis still needed when predictor–criterion relationships are strong?

A live question: if a predictor set empirically predicts a criterion (locally or meta-analytically), does job relatedness still need a job analysis? Risks of saying "no": a spurious relationship can masquerade as job relatedness. Example: "leadership experience" correlates with managerial performance, but gender also correlates with both (men more often given leadership opportunities) — gender should be irrelevant, yet the relationship may be found. Like ZIP-code-as-race in credit scoring, AI's millions of correlations can rest decisions on hidden, illegitimate relationships. Without a job analysis to identify, define, and measure the KSAOs, predictors and criteria "may be used that are wrong, inadequate, or unfair."

Forms and rigor of job analysis

  • Legal/professional guidelines allow many acceptable methods but are silent on how comprehensive the analysis must be. Guardians (2d Cir. 1980) holds job analysis should be systematic and accurate regardless of methodology.
  • Competency models: many test publishers apply a standard competency model across organizations, with loose ties to the actual tasks/behaviors of incumbents. Competencies may be specific KSAOs, broad KSAOs, or even organizational aspirations ("Demonstrates amazing customer service at all times"). Generic competency lists (e.g., Hunt, 1996) used for all jobs/levels for efficiency may not be complete relative to a specific job. Test publishers often omit a critical competency the test doesn't measure.
  • O*NET is a useful starting point for tasks/KSAOs but should not substitute for a complete job analysis (it has limitations given how the data were collected).
  • Collecting task information (frequency/importance ratings): UGESP requires measures of task importance/criticality and documents them as Essential (§15B(3)). This step is often skipped, yet collecting task ratings orients SMEs to actual job requirements and away from stereotypes — and reliability/validity of KSAO judgments rises when KSAOs are concrete and tied to specific tasks (Morgeson et al., 2020, p. 375).
  • SME judgment quality: interrater agreement (ICC) assumes SMEs are similarly knowledgeable or trained — an assumption rarely evaluated in practice.

Job relevancy / job relatedness (the legal anchor)

  • Job relevancy = the core features of a job (KSAOs, work behaviors, environment), determined by job analysis. U.S. law requires job relatedness for tests with adverse impact (Title VII disparate impact → "job related for the position and consistent with business necessity").
  • The Uniform Guidelines establish the explicit requirement to conduct a job analysis whenever undertaking a criterion-related validation study (§3.14(A)).
  • The Principles allow several routes to job relatedness: appropriate criterion-related validity coefficients, the job relevance of the content, or the construct measured (Principles, p. 90). Validity is a function of evidence supporting interpretation for specific purposes.
  • The AI gap: technologically enhanced tools may predict outcomes ("job related" in a statistical sense) yet not be based in a job analysis — leaving job relevance unestablished and legal defensibility weak.

Questions to ask (from the article)

  • Is a job analysis necessary if predictors and criteria are strongly related in a criterion-related study?
  • Is a job analysis necessary to justify an operational performance measure (e.g., KPIs) used as the validation criterion?
  • To what extent is a competency model an adequate substitute for a job analysis?
  • Is it important to have a complete list of competencies?
  • How much rigor in the job-analysis methodology is necessary to establish job requirements?
  • Is O*NET an acceptable source for a complete list of KSAOs?
  • Is it important to collect task information (e.g., frequency and importance ratings)?
  • If a job analysis is not conducted, can job relevancy be demonstrated for the assessments or the criteria? Is predictor–criterion correlation alone sufficient to establish job relatedness under the Uniform Guidelines?

Pitfalls

  • Deploying a generic "sales test" (or vendor competency model) without analyzing the local job.
  • Treating a strong predictor–criterion correlation as proof of job relatedness (spurious/proxy risk).
  • Using O*NET or an aspirational competency list as a complete job analysis.
  • Skipping task importance ratings, then over-relying on stereotype-tinged SME KSAO judgments.
  • Never checking whether SMEs are comparably knowledgeable before trusting interrater agreement.

Checklist

  • A systematic, accurate job analysis underlies predictors and criteria (or its absence is flagged)
  • Spurious/proxy relationships ruled out via job-analytic grounding
  • Competency model's rigor and completeness for the specific job evaluated
  • O*NET used only as a starting point, not a substitute
  • Task importance/criticality information collected and documented
  • SME comparability assessed before relying on agreement statistics
  • Job-relatedness route (criterion / content / construct) identified for legal defensibility

See also

work-analysis (how-to) · ai-predictor-theoretical-basis · ai-validity-evidence · ai-selection-legal-landscape · ai-input-data-and-design-audit

Source: Tippins, Oswald & McPhail (2021), Concerns: "Job Analysis" and "Job Relevancy."

AI methodology evaluation (Concern 4)

Technologically enhanced selection may apply ML to thousands of data points, weighting and combining them into predictions that are (a) complex (interactive, nonlinear) and (b) intended to hold up in new datasets (cross-validation). Evaluating whether that methodology is appropriate requires understanding methods that are unfamiliar to many I-O psychologists.

The interpretability problem

  • ML results can be obscure, counterintuitive, or impossible to interpret: random forests base predictions on hundreds of "trees"; neural nets tune arbitrary, layered configurations of "neurons." Understanding what goes on inside the black box is problematic.
  • Explainable AI (XAI) is advancing (Ribeiro et al., 2016; Kaleghi, 2019), but the authors are unaware of breakthroughs in complex selection prediction that yield new insight for theory or practice. Continued XAI development will be extremely important for achieving transparency in personnel selection.
  • When AI methods are used in talent assessment, the underlying predictive structures are generally inaccessible or proprietary, making it hard to interpret results for ourselves, stakeholders, the legal community, and beyond.

Evaluation metrics (know them, or get help)

AI models generate metrics unfamiliar to many I-O psychologists:

  • Mean squared error (MSE) — minimized in many estimation methods.
  • Confusion matrix — actual vs. predicted classifications.
  • ROC curve and area under the curve (AUC) — correct/incorrect classifications under different cut scores.

Many I-O psychologists lack the background to interpret and evaluate these and need additional methodological education.

The variable-to-case ratio

In big-data applications the ratio of variables to cases is often high (e.g., 30 variables per case), the reverse of traditional analyses (≈1 variable per 30 cases). Traditional statistics are literally impossible when variables exceed cases (e.g., the variance–covariance matrix won't invert) — which is why ML is necessary to operate on big data (unless variables are reduced via composites, factor/scale scores, etc.). Even with cross-validation, variable-driven interpretation (as in regression coefficients) differs and often remains in a black box.

You can't yet compare AI to traditional methods cleanly

  • For traditional measures, there's a large meta-analytic literature giving typical correlation ranges by instrument/construct (Schmidt & Hunter, 1998), plus effect-size benchmarks (Cohen, 1988; Bosco et al., 2015). I-O psychologists would be highly skeptical of a .75 correlation between a structured interview and overall performance.
  • But the available literature on ML application and efficacy in selection is insufficient to provide a basis for comparison. Because many I-O psychologists lack a fundamental understanding of how different ML algorithms work — their assumptions, boundary conditions, and metrics — it is challenging to compare ML results to one another and to traditional multiple regression.

The I-O education gap (a recurring theme)

ML methods may be unfamiliar or "completely foreign" to many I-O psychologists, yet their strong training in psychological measurement and psychometrics positions them to extend into ML and participate in critical conversations: whether big-data analysis is necessary, whether it provided meaningful prediction and a substantial improvement, and whether/when predictions generalize. More ML education is "clearly needed" (Aiken & Hanges, 2015; Oswald & Putka, 2016, 2017), including changes to I-O graduate curricula.

Questions to ask (from the article)

  • What are the appropriate criteria for evaluating technologically enhanced models?
  • What is the best approach to choosing and using an ML algorithm?
  • What defines acceptable results when ML is used? What is an acceptable level of prediction?
  • How can we determine if predictive results are based on idiosyncrasies of the sample the model was built on — how generalizable are they to other samples?
  • What continuing education and changes to I-O graduate education are needed to prepare psychologists to develop, research, and evaluate these tools?

Pitfalls

  • Accepting a model's reported accuracy without understanding its metric (e.g., AUC) or cross-validation.
  • Comparing an ML model's performance to traditional validities without a common, defensible baseline.
  • Treating a proprietary black box as evaluable from outputs alone.
  • Mistaking in-sample fit for generalizable prediction (sample idiosyncrasy).
  • Deferring entirely to data scientists instead of building enough ML literacy to ask hard questions.

Checklist

  • Model interpretability/XAI status assessed; black-box limits documented
  • Reported metrics (MSE/confusion matrix/ROC-AUC) understood and scrutinized
  • Variable-to-case ratio and its analytic implications considered
  • Generalizability vs. sample idiosyncrasy probed (cross-validation appropriateness)
  • Comparison to traditional/meta-analytic baselines attempted (and its limits acknowledged)
  • ML expertise secured (own or collaborator) sufficient to evaluate the methodology

See also

ai-validity-evidence · ai-reliability · ai-model-development-audit · ai-model-outputs-audit · criterion-related-validation (data analysis, cross-validation) · ai-extending-professional-standards

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

AI predictors: theoretical basis (Concern 1)

Many technologically enhanced assessments use a wide variety of data scraped from applications, resumes, social media, emails, or the Internet, then run through hundreds of candidate ML algorithms. The substantive nature of the included variables and their linkages to job requirements are often unknown.

The problem

  • No substantive rationale. Past-employer data scraped from resumes might show that employment at Employers A, B, C predicts future performance while D, E, F do not — with no substantive post hoc explanation, even when all six are in the same business. Theology coursework might "predict" sales success. Voice or facial characteristics may have no obvious theory linking them to KSAOs or performance — justification is "inferred at best and unknown at worst."
  • Proxy variables. Atheoretical predictors readily encode protected characteristics. In credit scoring, ZIP code is a known proxy for race; AI using millions of correlations can base decisions on such hidden relationships. A predictor can "work" statistically while being a construct-irrelevant proxy.
  • Interpretability ≠ relevance. Even when relationships are interpretable, they may have little practical or conceptual relevance to the work performed (Braun & Kuljanin, 2015). Big data is often massive, messy, and missing.

The debate (represent both sides honestly)

I-O psychology has long debated whether predictors need a theoretical basis:

  • "Theory matters." The traditional basis for including a measure is the extent to which it reflects a KSAO necessary to perform the job, as determined by a job analysis. The Standards and Principles embed this in the very definition of validity: "the degree to which accumulated evidence and theory support specific interpretations of scores… entailed by the proposed uses" (Principles, p. 96; Standards, p. 225; emphasis added).
  • "Prediction is enough." Others hold that if scores correlate with a relevant criterion (performance, engagement, turnover), they're useful predictors and the rationale is merely "nice to know."

The article's resolution: if the only purpose is mechanical prediction, studying constructs and jobs is "merely a response to regulatory requirements." But if the purpose looks beyond simple prediction, understanding the predictive relationship yields improved measures, broader coverage of the performance domain, greater generalizability, and assurance that the system is sensible with respect to recruiting, training, diverse applicant pools, and change over time. Systematic research also surfaces additional variables and data sources that may predict, mediate, or explain work behavior (Rotolo & Church, 2015).

How adverse impact changes the calculus

  • Without adverse impact, even a job-irrelevant predictor would not legally require the "job related / business necessity" defense (Title VII) and is not unlawful per se. From this view, theory can look like an avoidable intellectual exercise.
  • With adverse impact, job relatedness — which rests on relevance/theory established via job analysis — becomes a legal requirement (see ai-selection-legal-landscape, ai-job-analysis-and-relevancy).

So the theoretical-basis question is partly scientific (do we want to understand prediction?) and partly contingent on adverse impact (do we legally have to?). The deeper issue: is selection research propelled by science, prioritizing understanding applicants' suitability through the lens of job requirements — or is it an atheoretical, purely empirical activity to maximize predicted outcomes?

Questions to ask (from the article)

  • Are theoretical justifications necessary in employment testing?
  • Is a technologically enhanced measure that predicts organizational outcomes sufficient, or does one need to understand why that prediction occurs?
  • Do theoretical justifications improve practice in employment testing?
  • Do the considerations about theoretical justification change when there is adverse impact versus when there is not?

Pitfalls

  • Accepting "it predicts, so who cares why" when the use goes beyond one-shot mechanical prediction.
  • Missing proxy variables that encode protected characteristics through hidden correlations.
  • Assuming interpretability implies job relevance.
  • Treating a serendipitous, unreplicated correlation as a justified predictor.

Checklist

  • Each predictor's substantive link to a job-relevant KSAO examined (or its absence noted)
  • Proxy-variable risk for protected characteristics assessed
  • Purpose clarified: mechanical prediction only, or understanding/coverage/generalizability?
  • Adverse-impact status checked and its legal implications applied
  • Atheoretical/serendipitous relationships flagged for replication before use

See also

ai-job-analysis-and-relevancy · ai-selection-legal-landscape · ai-validity-evidence · work-analysis · criterion-related-validation (predictor choice, rationale) · ai-input-data-and-design-audit

Source: Tippins, Oswald & McPhail (2021), Concern: "Lack of a Theoretical Basis for Predictors."

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."

AI selection ethics (Concern 11)

SIOP members are bound by the APA Ethical Principles of Psychologists and Code of Conduct. A review of Section 9 (Assessment) surfaces several ethical concerns for newer forms of assessment. The central message: reliability, validity, and fairness are not separate from ethics — they are intertwined with it. It is the professional responsibility of I-O psychologists to require information about reliability, validity, and fairness when deciding whether a selection system — technology-enhanced or otherwise — can be used to make inferences about job performance.

The governing standards (APA Ethics Code, Section 9)

  • 9.01 Bases for Assessments. Psychologists base the opinions in their recommendations, reports, and evaluative statements on information and techniques sufficient to substantiate their findings. → An opinion that an AI tool selects good employees must rest on sufficient evidence, not vendor assertion.
  • 9.02 Use of Assessments. Psychologists use instruments whose validity and reliability have been established for the population tested. When validity/reliability have not been established, they describe the strengths and limitations of results and interpretation. → If an AI tool lacks established reliability/validity for your applicant population, that must be disclosed and its limitations described — not glossed.
  • 9.03 Informed Consent. Psychologists obtain informed consent for assessments (per Standard 3.10), except when (1) testing is mandated by law or governmental regulation; (2) consent is implied because testing is a routine educational, institutional, or organizational activity (e.g., when participants voluntarily apply for a job); or (3) the purpose is to evaluate decisional capacity. Informed consent includes an explanation of the nature and purpose of the assessment, fees, third-party involvement, limits of confidentiality, and a sufficient opportunity to ask questions and receive answers.

Where AI strains these standards

  • Implied consent may not reach incidental data. Implied consent covers routine testing when someone applies for a job — but it is not clear that implied consent extends to data the candidate may not be aware is being obtained (scraped social media, facial/voice analysis). This links directly to ai-candidate-data-control.
  • Different consent standards for builders vs. users. The informed-consent standards differ for researchers developing selection tools (Ethics Code 8.05) versus those employing them (Guzzo et al., 2015; Dekas & McCune, 2015). Know which role you're in.
  • Sufficient evidence to substantiate findings. If an ML algorithm infers that applicants will have a higher likelihood of good performance, what is the quality and strength of the evidence behind that inference (9.01)? Establishing reliability and validity is an ethical, not merely technical, obligation, because the recommendation made from the test score must be supported.

The integrating point

I-O psychologists must determine whether ethical standards are being met or can be met when working with AI tools. That means they must:

  • establish (or require evidence of) validity and reliability of the instruments used for selection (9.02), and
  • have the evidence to support the recommendation made from the test score (9.01), and
  • ensure appropriate consent (9.03), recognizing the gaps around uncontrolled/unaware data.

In short, the ethical duty operationalizes the scientific concerns: you cannot ethically deploy a tool whose reliability, validity, and fairness you cannot vouch for.

Questions to ask

  • Is there sufficient evidence to substantiate the findings/recommendations the tool produces (9.01)?
  • Have validity and reliability been established for the population tested — and if not, are strengths/limitations clearly described (9.02)?
  • Is informed consent properly obtained or appropriately implied — and does any implied consent actually cover data the candidate is unaware of (9.03)?
  • Are you acting as a developer (8.05) or a user of the tool, and have you applied the right consent standard?

Pitfalls

  • Relying on vendor claims rather than evidence "sufficient to substantiate findings" (9.01).
  • Using a tool without reliability/validity established for your applicant population, and not disclosing the limitation (9.02).
  • Stretching "implied consent" to cover scraped or incidental data the candidate never knew about (9.03).
  • Confusing the consent obligations of tool developers with those of employers/users.
  • Treating ethics as separate from psychometrics rather than as their enforcement.

Checklist

  • Evidence is sufficient to substantiate the tool's recommendations (9.01)
  • Reliability/validity established for the tested population, or limitations described (9.02)
  • Informed consent obtained or properly implied; coverage of unaware/uncontrolled data examined (9.03)
  • Developer (8.05) vs. user consent role identified and correct standard applied
  • Reliability/validity/fairness evidence required as an ethical precondition for deployment

See also

ai-candidate-data-control (consent for uncontrolled data) · ai-validity-evidence · ai-reliability · ai-selection-legal-landscape (APA Code as professional, not legal) · ai-fairness-lenses (legal/ethical/moral lens) · ai-audit-meta-components (respect / ethical-standards conformance)

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

AI selection: legal landscape

The legal/regulatory frame an AI selection tool must survive. This is decidedly U.S.-centric because the legal requirements are, but many of the concerns apply globally (validity as business justification exists everywhere). Not legal advice — consult qualified counsel.

Core principle: AI/ML selection tools are "tests" under the law. Uniform Guidelines Section 2B defines a test as any selection procedure used as the basis for an employment decision — which sweeps in games, video interviews, facial/voice scoring, resume screeners, and big-data models. They are governed by the same rules as any other test.

The federal framework (U.S.)

Uniform Guidelines on Employee Selection Procedures (EEOC, CSC, DoL, DoJ, 1978)

  • Broadened "test" to any selection procedure used for an employment decision (Sec. 2B; the 1979 Q&A further expanded it to application forms, interviews, training/probationary performance, etc.).
  • Adopted 1978, with very few changes and none since the 1981 Q&A (the OFCCP's 2019 FAQs are not part of the Guidelines). Theoretically and practically dated — yet they remain the controlling administrative rules for Title VII litigation and are deeply embedded in case law. They must still be considered when evaluating any procedure that results in adverse impact.
  • Require a job analysis whenever conducting a criterion-related validation study, and specify the documentation expected in technical reports.

Title VII, Civil Rights Act of 1964 — disparate impact

A disparate-impact violation is established when a complaining party shows a practice causes disparate (adverse) impact on the basis of race, color, religion, sex, or national origin and the respondent fails to demonstrate the practice is job related for the position and consistent with business necessity (Sec. 2000e-2(k)(1)(A)(i)). So:

  • Adverse impact shifts the burden to the employer to show job relatedness / business necessity — which in practice means validity evidence (see ai-validity-evidence).
  • Differential treatment (treating class members differently — e.g., awarding bonus points to a group, or modeling class membership) is a distinct, separate legal problem.

OFCCP (2019 FAQ)

"Irrespective of the level of technical sophistication involved, OFCCP analyzes all selection devices for adverse impact." If a contractor's AI-based procedure has adverse impact, the contractor must validate it using an appropriate validation strategy. AI buys no exemption.

Search for less-adverse alternatives

U.S. employers are obligated to consider alternative selection procedures with substantially equal or greater validity and less adverse impact (Uniform Guidelines §3B). This makes comparative data important — an AI tool should be compared against alternatives, including traditional measures whose meta-analytic validity provides a reasonable baseline the AI must beat.

Case law — job analysis in content validation

Guardians Association of NYC Police Dept. v. Civil Service Commission of NYC (2d Cir. 1980) and related cases establish that job analysis is central in content-validation disputes and should be systematic and accurate regardless of the methodology used. (See ai-job-analysis-and-relevancy.)

State law (proliferating)

  • Illinois Artificial Intelligence Video Interview Act (passed May 29, 2019; effective Jan. 1, 2020): employers using AI to evaluate video interviews must (a) notify applicants in writing that AI may be used and what characteristics it evaluates, (b) provide information on how the technology works and what characteristics it uses, and (c) obtain written consent before the interview. They may not share the video except with those with expertise to evaluate it, and must destroy the video within 30 days of a request.
  • Other state legislatures are weighing in; legislation and policy on the nature, transparency, and privacy of applicant data will continue to emerge and evolve. Re-check current jurisdictional law.

Professional (not legal) — the APA Ethics Code

SIOP members subscribe to the APA Ethics Code, which governs the treatment of candidates and what psychologists say about tools — an obligation independent of, and additional to, the law. See ai-selection-ethics.

The "no adverse impact ≠ valid" rule

A crucial argument to deploy: reduced or no adverse impact does not, by itself, justify use. A random-number generator can winnow a large applicant pool quickly without producing adverse impact — yet it lacks the reliability, validity, and utility an organization needs to identify capable candidates and achieve an acceptable return. It is incumbent on developers (and users) to provide sufficient evidence that AI selection tools meet these requirements — not merely to show they don't create adverse impact. Vendors often tout "reduced adverse impact" while the empirical validity evidence is unavailable, making the relevant comparison impossible.

Pitfalls

  • Treating an AI tool as exempt from test-validation law because it's "just an algorithm."
  • Relying on "no/low adverse impact" as if it established validity or utility.
  • Forgetting the duty to consider less-adverse, equally valid alternatives.
  • Overlooking state-specific notice/consent/destruction requirements (e.g., Illinois).
  • Assuming the OFCCP 2019 FAQs are part of the Uniform Guidelines (they are not).
  • Conflating disparate impact with disparate (differential) treatment.

Checklist

  • Tool classified as a "test" under UGESP §2B
  • Adverse-impact exposure assessed; validity-defense obligation understood
  • Job-relatedness/business-necessity evidence path identified (→ ai-validity-evidence)
  • Less-adverse-alternative comparison considered (§3B)
  • Job-analysis adequacy checked (Guardians) (→ ai-job-analysis-and-relevancy)
  • State-law notice/consent/destruction requirements checked (e.g., Illinois AIVI Act)
  • "No adverse impact ≠ valid" argument applied to vendor claims
  • APA Ethics Code obligations noted (→ ai-selection-ethics)
  • Counsel engaged for jurisdiction-specific questions

See also

ai-validity-evidence · ai-job-analysis-and-relevancy · ai-candidate-data-control (consent) · ai-selection-ethics · fairness-and-bias-analysis · criterion-related-validation (alternatives, validation)

Source: Tippins, Oswald & McPhail (2021), "New Forms of Assessment" (legal framing), "Disadvantages," "Purpose," and the legal threads woven through the concerns.

AI selection: technologies, data, algorithms

The starting lens for the whole "scientific, legal, and ethical concerns" framework. Before asking "is this AI tool any good?", decompose it into three independent parts and evaluate each separately (the "modular approach," Lievens & Sackett, 2017). Conflating them is the root error behind most overclaiming.

The three parts

1. Technologies — independent of the constructs measured

Examples: online games, video interviews, social media, gamification, VR. Technologies are independent from the constructs being measured and should not be confused with them (Arthur & Villado, 2008; Campbell & Fiske, 1959). Most formats can measure a wide range of constructs, and most constructs can be assessed by many technologies.

Consequence (state this in any evaluation): a new technology cannot be said to be universally valid. "Our video-interview platform is validated" is a category error — validity is a property of the inferences about the constructs measured in a specific use, not of the technology. Demand evidence about which job-relevant constructs are measured, which then informs validity and fairness.

2. Data — a continuum from intentional to incidental

Data vary along a continuum (Oswald, 2020):

  • Intentional — traditional, controllable responses to a prompt (a test item, an interview answer).
  • Incidental — less intentional/controllable, requiring little or no applicant effort (and sometimes little applicant control): social-media posts, facial movements, voice characteristics in a video, mouse clicks, response times.

Less obtrusive technologies tend to collect more incidental data, in massive amounts (game data = every click/decision/scenario; video = continuous voice and facial features; "big data" pulled from resumes, emails, social media). The more incidental the data, the more the concerns about job relevance, control, consent, and fairness intensify (see ai-candidate-data-control).

3. Algorithms — AI ⊃ ML ⊃ deep learning

  • AI — broad term for computer procedures that mimic human decisions/processes/outcomes closely enough to appear intelligent.
  • Machine learning (ML) — subset: the mathematical/statistical procedures underlying these tools.
  • Deep learning — subset of ML: neural-network-based.
  • Supervised learning predicts a criterion (the criterion "supervises" how predictors are used — e.g., predicting supervisory performance ratings); unsupervised learning groups people/cases into clusters with no criterion (e.g., applicants like/unlike high performers).

The "learning" happens when algorithms are first exposed to a training set; the model is judged on an independent test set (a hold-out sample, k-fold folds, or newly collected data).

A key claim to carry forward

There are hundreds of ML algorithms, and different algorithms often make highly similar predictions with similar overall accuracy (Domingos, 2012). So in personnel selection, the effectiveness of ML prediction/clustering is more likely driven by the availability of high-quality data than by which ML algorithm is chosen. Whether the advantages of a large number of predictors offset the disadvantages of "messy" data must be determined case by case. Accurate, well-justified predictions depend on good measurement processes and good data — not on algorithm sophistication. Don't let a vendor's algorithm story distract from data-quality and construct questions.

How to use this skill

  1. Name the three parts of the tool under review (which technology, what data on the intentional–incidental continuum, which algorithm type — supervised/unsupervised).
  2. Reject "the technology is valid" framing; redirect to what constructs are measured and is that measurement job-relevant, reliable, valid, and fair?
  3. Locate the data on the continuum — the more incidental, the more you escalate the control/consent/fairness concerns.
  4. Treat algorithm choice as secondary to data quality; probe the data-generation and construct-relevance story.
  5. Proceed to the legal frame (ai-selection-legal-landscape) and the 11 concerns.

Pitfalls

  • Accepting "this technology is validated" as if validity were a property of the medium.
  • Being dazzled by the algorithm while ignoring data quality and construct relevance.
  • Missing that incidental data (face, voice, social media) drives the hardest legal/ethical issues.
  • Assuming supervised = good and unsupervised = bad (each has distinct evaluation needs).

Checklist

  • Technology, data type, and algorithm type named and evaluated separately
  • "Universally valid technology" claim rejected; construct/inference framing substituted
  • Data located on the intentional–incidental continuum; control/consent flags raised if incidental
  • Supervised vs. unsupervised identified; training/test-set basis confirmed
  • Data-quality story probed as primary; algorithm choice treated as secondary

See also

ai-selection-legal-landscape · all 11 concern skills · ai-validity-evidence · ai-input-data-and-design-audit (the audit counterpart) · validation-planning

Source: Tippins, Oswald & McPhail (2021), "New Forms of Assessment."

AI validity evidence (Concern 5)

Among I-O psychologists there is no doubt about the importance of validity. Evidence of validity is necessary to demonstrate a selection procedure's effectiveness, and it is a legal requirement when adverse impact is observed. For business and legal-defensibility reasons, validity is a sine qua non of selection research — including for AI tools.

This is the AI-specific framing; for how to gather validity evidence, use criterion-related-validation, content-based-validation, and generalizing-validity-evidence.

AI tools are "tests" — so they need validity

AI-based predictors, game-oriented tools, and evaluations of facial features and voice qualities used to decide whom to hire are tests under both the law and professional guidelines (UGESP §2B: a test is any selection procedure used for an employment decision). Therefore:

  • In the U.S., their use must be supported by validity evidence if adverse impact results.
  • In all countries, validity evidence provides the business justification for use.

Which validity strategy — and the AI wrinkle

  • Criterion-related — a statistical relationship between scores and a relevant criterion (correlation/regression). Report strength via size, confidence interval, and significance; regression should indicate imprecision (CIs, bootstrapping, cross-validation). Traditional selection rarely models nonlinearities (linearity affords interpretability and needs less power); ML does model and cross-validate nonlinearities, but interpretation is difficult and power may or may not be an issue depending on the number of variables and cases.
  • Content — SMEs relate the content of the test to job requirements. Because of the nature of AI-based tests, content validation evidence is often not used at all, or only to supplement a criterion-related study. (When content can't carry the argument, criterion evidence and job analysis must.)

When overall model fit is the only handle

Some ML methods are based on regression (logistic, lasso, ridge, elastic net) and yield coefficients — a "demonstration of a relationship." But complex models (random forests, neural nets) or clustering approaches don't express relationships as apparent, direct coefficients. There, overall model fit (analogous to R² in regression) may be the only basis for establishing the predictor–criterion relationship — which makes job analysis all the more important to ensure job-relevant predictors and criteria are the ones in the model, even when their weighting is opaque.

Comparative data and the alternatives duty

Because U.S. employers are obligated to search for alternative procedures with equal or greater validity and less adverse impact (UGESP §3B), comparative data are particularly important. The empirical, meta-analytic history of validities in I-O psychology can serve as a reasonable baseline the ML model should have to beat — but vendor claims of "higher validity / lower adverse impact" are frequently unaccompanied by the evidence needed to make the comparison.

Fundamental requirements and documentation

A key open question: should users of algorithm-based procedures be required to provide specific conceptual and empirical evidence of predictor–criterion relationships, in addition to overall model metrics? The Uniform Guidelines, Principles, and Standards all enumerate the components a technical report should document — yet critical information is often omitted, and the problem is acute for AI because so many decisions are made in selecting and tuning the model. Documentation should let an external party understand, evaluate, and reproduce the work. Ask:

  • How were data obtained, prepared, cleaned, transformed, and combined? What approach handled missing data?
  • With text mining / written responses, were effects of vocabulary knowledge or verbal fluency controlled where the text wasn't meant to measure verbal ability?
  • How was conflicting applicant data weighted, discarded, or cleaned?
  • Was one algorithmic approach taken or several, and could an external party reproduce the analysis?
  • Were algorithm choices a priori or post hoc? Were predictions combined across algorithms?
  • Was there post hoc filtering — inspecting results to pick the algorithm that "looked good" (a suspect practice)?

Questions to ask (from the article)

  • What is sufficient evidence of validity when ML models are used?
  • What details of the AI research must be documented?
  • What details of the AI research must be shared with users?

Pitfalls

  • Treating "no adverse impact" as a substitute for validity (see ai-selection-legal-landscape).
  • Accepting overall model fit (R²/AUC) without job-analytic grounding of the predictors/criterion.
  • Accepting "more valid, less adverse impact" claims without the comparative evidence.
  • Tolerating post hoc algorithm selection / filtering without cross-validation safeguards.
  • Thin documentation that no external party could reproduce.

Checklist

  • Tool treated as a "test" requiring validity (esp. if adverse impact)
  • Validity strategy chosen (criterion-related / content / generalization) and justified for AI
  • If only overall model fit is available, predictor/criterion job-relevance secured via job analysis
  • Imprecision reported (CIs, bootstrapping, cross-validation)
  • Comparative data vs. traditional/meta-analytic baselines and less-adverse alternatives provided
  • Data prep, missing-data, weighting, and algorithm-selection decisions documented and reproducible
  • Post hoc filtering ruled out or controlled; evidence shared with users

See also

criterion-related-validation · content-based-validation · generalizing-validity-evidence · technical-validation-report · ai-selection-tech-data-algorithms (why a technology is not "universally valid") · ai-ml-methodology-evaluation (model-fit metrics) · ai-job-analysis-and-relevancy · ai-selection-legal-landscape · ai-model-outputs-audit

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

Sponsors

Details

Marketplace
OpenMatter-Network/agent-io-skills
Source
ai-selection-legal-ethical
Category
Community

More from OpenMatter-Network/agent-io-skills

All 3