What AI hiring mistakes create discrimination risk?
The costliest mistake is assuming the vendor tested it. The second is never computing your own ratio.
Last updated: August 02, 2026
Direct Answer
Five mistakes account for most AI hiring discrimination risk: deploying tools with no adverse impact testing on your own applicant flow; providing no accommodation alternative for applicants with disabilities; letting models learn from proxy variables that track protected traits; rejecting applicants through decisions nobody can explain; and keeping no records of what the tool recommended versus what humans decided. Each is preventable with governance the employer controls.
Controlling authority: Title VII and the ADA, which govern by disparate impact regardless of intent, and TRAIGA (Tex. H.B. 149), which is intent-based and enforceable only by the Texas Attorney General.
The Five Mistakes
Untested deployment leads the list because scale changes everything: a biased human screener touches dozens of candidates, while a biased model touches all of them, uniformly, with perfect consistency. Testing pass-through rates by protected group on your actual applicant data, at deployment and periodically after, is the control. The second mistake is skipping accommodations: timed assessments, video scoring, and games can screen out disabled applicants who would perform the job fine, and the fix is a real human alternative offered up front.
The third is proxies. Models trained on your historical hires learn your historical patterns, including zip codes, school names, employment gaps, and activity keywords that correlate with race, sex, age, or disability. The fourth is unexplainability: when a rejected applicant asks why and the honest answer is nobody knows, that answer performs poorly in front of agencies and juries. The fifth is missing records: without logs of tool recommendations and human decisions, you cannot demonstrate oversight even where it existed.
| Source | What it does | What it does not do |
|---|---|---|
| Title VII and the ADA | Still the operative risk. Disparate impact liability attaches regardless of intent, and regardless of whether a vendor built the tool | It does not excuse you because the algorithm is a third party’s. |
| TRAIGA (effective 1 January 2026) | Prohibits developing or deploying AI with the specific intent to discriminate on protected characteristics, plus a narrow set of other prohibited uses | It is intent-based, not impact-based. It was pared back substantially before passage and imposes far less on private employers than commentary suggests. |
| TRAIGA enforcement | The Texas Attorney General only | There is no private right of action. Your TRAIGA exposure is regulatory, not litigation. |
| The four-fifths rule | The screen that will actually surface your problem | It is a triage indicator, not a safe harbour above 0.80. |
| EEOC AI technical assistance | Removed from the EEOC website on 27 January 2025 | Those were non-binding technical assistance documents. Removing them changed the explanation, not the obligation. Title VII and the ADA are untouched. |
| Your vendor contract | Allocates cost and cooperation between you and the vendor | It does not transfer liability to the vendor. The employer is the one that made the decision. |
What the Prevention Program Looks Like
The controls mirror the mistakes: outcome testing on a calendar, an accommodation path advertised before automated steps, vendor answers in writing about training data and proxy management, a human decision layer for rejections in protected-risk zones, and retention of the decision trail. None of this requires a data science team; it requires an owner and a checklist.
This connects to the broader Texas picture: TRAIGA governance, your AI use policy, and your hiring compliance process are one system. Employers who run that system get AI's speed without inheriting its silent failure modes.
| Requirement | What to ask for | Why |
|---|---|---|
| Adverse impact testing | Results by race, sex, ethnicity and age, on your applicant pool | A vendor’s aggregate testing says nothing about your population. |
| Validation evidence | Job-relatedness and business necessity, documented | This is the defence if impact appears. |
| What the model actually scores | The features and their weights, at least in summary | You cannot defend a decision you cannot describe. |
| Accommodation pathway | A documented alternative for candidates who cannot use the tool | An ADA obligation the vendor will not discharge for you. |
| Human review point | Where a person can override, and on what basis | A fully automated rejection is the hardest fact pattern to defend. |
| Audit and data rights | Your right to test, and to export your own data | Without it you cannot run the four-fifths screen at all. |
| Change notification | Notice before the model is retrained or changed | A silent model update can move your selection rates overnight. |
AI Hiring Risks to Watch
These conditions predict the charge before it arrives. Watch for these.
- Pass-through rates by protected group never measured
- Assessments with no accommodation alternative
- Tools trained on historical hiring data with no proxy review
- Rejections no human can explain
- No retained record of tool output versus human decisions
Compute your own ratios before the next hiring cycle
Pick your highest-volume role and audit its funnel end to end: every automated gate, its measured impact, its accommodation path, and its records. Fix that funnel, then replicate the fixes.
Add outcome testing to the annual compliance calendar so it survives staff turnover.
| Step | The calculation | Worked example |
|---|---|---|
| 1. Selection rate per group | Selected ÷ applicants, for each group | Group A: 60 of 100 = 60%. Group B: 30 of 100 = 30%. |
| 2. Identify the highest rate | The comparison base | Group A, at 60%. |
| 3. Compute the ratio | Lower rate ÷ highest rate | 30% ÷ 60% = 0.50. |
| 4. Apply the threshold | Below 0.80 is generally regarded as evidence of adverse impact | 0.50 is well below 0.80. This selection procedure needs examination. |
| 5. Do not stop at 0.80 | Smaller differences may still be adverse impact where statistically and practically significant, or where the employer’s conduct discouraged applicants disproportionately | A ratio of 0.85 is not a clean bill of health. |
| 6. If impact appears | Validate the procedure for job-relatedness and business necessity, or find a less discriminatory alternative | Document the analysis either way. The undocumented analysis is treated as no analysis. |
Job Description Reality Gap Scorecard
Scores the distance between what the job description says and what the person actually does.
When a tool is already in production
Get help structuring the first audit if the funnel has never been examined, because the first pass sets the template for everything after.
If a demand letter or EEOC charge referencing your screening tools has arrived, coordinate the response with counsel and preserve the tool records immediately.
Get a Straight Answer for Your Situation
General rules only go so far. If this question is live in your organization right now, talk it through with a senior HR consultant before you act. One conversation now costs less than one claim later.
Contact UsThis page provides general HR information for employers and is not legal advice. For legal interpretation or representation, consult qualified employment counsel.