EEOC AI Hiring Guidance for Employers: Algorithmic Discrimination and Compliance in 2026
Employers using AI to screen résumés, rank candidates, conduct automated assessments, or analyze video interviews can be held liable for Title VII discrimination even when the bias lives in a vendor's algorithm. The EEOC's guidance is clear: you own the outcome, not your software vendor.
What the EEOC Has Actually Said
The EEOC has issued several guidance documents and technical assistance pieces on AI in employment. The consistent message across all of them: the employer bears liability under Title VII (and the ADA, ADEA) for discriminatory outcomes from AI tools used in employment decisions, regardless of who built the algorithm.
The EEOC's 2022 technical guidance on AI and Title VII established that software and algorithms are selection procedures subject to the Uniform Guidelines on Employee Selection Procedures (UGESP). UGESP requires employers to assess selection tools for adverse impact on protected classes — a requirement that applied to written tests in the 1970s and now applies equally to AI scoring models.
In plain terms: if your AI résumé screener, automated interview platform, or candidate ranking tool disproportionately screens out Black applicants, female candidates, or applicants with disabilities, the EEOC will view that as disparate impact discrimination — and the employer, not the software vendor, faces the primary liability.
How AI Creates Disparate Impact Without Obvious Bias
Algorithmic discrimination is often invisible at the individual decision level and only visible in aggregate outcomes. The mechanisms that create it are worth understanding:
Training data that reflects historical bias
If an AI is trained on résumés of past successful hires, and past hires skewed toward a particular demographic due to prior discriminatory practices, the AI learns to replicate that pattern. The algorithm isn't 'biased' in a human sense — it's accurate to a biased historical record.
Proxy variables for protected characteristics
AI models discover correlations humans wouldn't notice. Attending historically Black colleges and universities, living in certain zip codes, having career gaps consistent with caregiving, or stylistic patterns in writing that correlate with national origin or language background can all serve as proxies. The AI isn't explicitly screening by race — but the practical effect is the same.
Facial expression and vocal analysis in video interviews
AI tools that score candidates on engagement, confidence, or personality based on facial movements or voice patterns often encode cultural norms about how these traits present in specific demographics. Research consistently shows these tools produce racially and culturally disparate outcomes.
Skill-matching models trained on job descriptions with embedded bias
Job descriptions written by male managers have been shown to disproportionately include masculine-coded language. AI trained to match résumés to these descriptions absorbs the bias already embedded in the input.
Employer Liability: The Vendor Defense Doesn't Work
The most common misconception among employers: if the AI tool is commercially available software and the vendor claims it's been validated, the employer is protected. The EEOC disagrees — and courts have generally agreed with the EEOC on this.
Title VII's employer liability attaches to employment decisions and practices — not to whether you built the discriminatory mechanism yourself. Using a vendor's tool that produces disparate impact is equivalent, under the law, to designing a biased written test yourself. The employer remains liable.
Vendor indemnification clauses in HR tech contracts vary widely and often don't fully protect employers. Even if a vendor agrees to indemnify you for claims arising from their algorithm, that indemnification won't necessarily cover EEOC enforcement actions, and it won't undo the reputational damage or hiring outcomes. The only real protection is using validated tools and conducting your own adverse impact analysis.
The ADA Dimension: Disability and AI Screening
The ADA adds a layer of complexity. AI tools that screen candidates based on behavioral or psychological assessments — and many do — may have a disparate impact on candidates with mental health conditions, neurodivergent candidates, or candidates with physical disabilities that affect their performance on AI-assessed tasks.
Automated personality and cognitive assessments
Game-based assessments and cognitive ability tests may disadvantage candidates with ADHD, anxiety disorders, or learning disabilities. Employers must ensure reasonable accommodations are available for automated assessments — not just for in-person interviews.
AI video analysis and physical presentation
Video interview AI scoring on facial expressions, eye contact, or speech patterns can disadvantage candidates with autism spectrum disorder, Tourette syndrome, or speech impediments. These tools have documented disparate impact on disability status.
Accommodation requests and AI workflows
If a candidate requests an accommodation for an automated assessment step, and the AI pipeline has no mechanism to route accommodation requests to a human, the employer is in ADA violation territory. AI hiring workflows need human override and accommodation pathways built in.
NYC Local Law 144: The Strictest AI Hiring Rule in the US
While the EEOC operates through guidance and enforcement, New York City enacted the first specific law mandating AI bias audits for hiring tools: Local Law 144, which took effect in July 2023. For employers hiring in New York City, LL 144 imposes concrete requirements beyond EEOC guidance:
Many employers using national or global hiring platforms are subject to LL 144 without realizing it — if any of the roles they hire for include NYC candidates, the law applies. Several major HR tech vendors have published bias audit results; if yours hasn't, ask.
Compliance Checklist for Employers Using AI in Hiring
Inventory and Vendor Assessment
- ☐List every AI tool involved in candidate evaluation (screening, scoring, ranking, video analysis)
- ☐Request adverse impact testing data from each vendor
- ☐Ask vendors for bias audit results on race, sex, national origin, age, disability
- ☐Review vendor contracts for indemnification and liability allocation
- ☐Assess whether any tool qualifies as an AEDT under NYC LL 144
Adverse Impact Testing
- ☐Run the four-fifths (80%) rule on selection rates for each AI tool by protected class
- ☐Analyze pass-through rates at each AI-gated stage of your hiring funnel
- ☐Document analysis results and keep records for at least two years
- ☐Where adverse impact is found, conduct job-relatedness validation study
- ☐Consider alternative selection tools if adverse impact can't be eliminated
NYC LL 144 Compliance (If Applicable)
- ☐Identify roles where NYC candidates may apply
- ☐Confirm your AI hiring tools qualify as AEDTs under LL 144
- ☐Obtain or commission annual independent bias audit
- ☐Publish audit results on website
- ☐Add AEDT disclosure to job postings and application flows
- ☐Implement opt-out accommodation process
Process and Documentation
- ☐Ensure human review is available at every AI-gated stage
- ☐Build accommodation request workflows for automated assessments
- ☐Train recruiters on AI tool limitations and override procedures
- ☐Document justification for continuing use of any tool with adverse impact
- ☐Review AI tool use at least annually as models and job requirements change
What to Ask Your HR Tech Vendor
Vendors vary enormously in how much adverse impact testing they've done and how transparent they are about results. Here are the direct questions to ask — and what answers to treat as red flags:
Has your tool been tested for adverse impact on race, sex, national origin, age, and disability?
✓ Good: Yes — with published or available-on-request results for each protected class
✗ Red flag: "Our tool doesn't consider protected characteristics" — that's not the same as testing for adverse impact
Can you provide technical documentation on how the model was trained and validated?
✓ Good: Yes — model card or technical whitepaper with training data sources and validation methodology
✗ Red flag: Proprietary / trade secret — reasonable summary should still be available for customer compliance purposes
What data was used to train the model, and how was historical bias addressed?
✓ Good: Training data sources disclosed, bias mitigation steps described, validation on representative datasets
✗ Red flag: Vague references to 'diverse training data' without specifics
Does your platform accommodate opt-outs and reasonable accommodation requests?
✓ Good: Yes — with documented process for routing accommodation requests to human review
✗ Red flag: No alternative assessment pathway available
Frequently Asked Questions
We use LinkedIn Recruiter's AI features to surface candidates. Are we liable for its algorithm?
This is an area of active legal development. Using platform AI to surface candidates is different from using it to make final selection decisions. The EEOC's guidance focuses on selection procedures — tools that screen in or out of a candidate pool for a specific position. Broad sourcing AI is lower risk, but if LinkedIn's AI systematically fails to surface candidates in protected classes for your specific searches, that's worth monitoring. Keep records of your sourcing funnel demographics.
Our AI screening tool vendor told us their product is 'EEOC compliant.' What does that mean?
Not much, unfortunately. There is no EEOC certification or compliance seal. A vendor claiming 'EEOC compliance' is making a marketing claim, not representing that a government agency has reviewed or approved their product. The vendor may have conducted adverse impact testing — which is what actually matters — but ask to see the results, not just the claim.
Can we use AI for some parts of hiring but not others to limit risk?
Yes. AI use in broad résumé keyword matching is much lower-risk than AI that scores personality, predicts performance, or ranks candidates for final-stage selection. Using AI as a search tool (find all résumés mentioning Python) vs. a decision tool (score and rank these 500 candidates) carries very different liability profiles. Keeping humans in consequential selection decisions reduces both adverse impact risk and EEOC exposure.
If we find adverse impact in our AI tool, do we have to stop using it?
Not necessarily. Under UGESP, an employer can continue using a selection tool with adverse impact if they can demonstrate it is job-related and consistent with business necessity. The burden shifts to the employer to validate that the tool's selection criteria are genuinely predictive of job performance — a formal validation study. If you can demonstrate job-relatedness, the tool can continue. If you can't, you need to find an equally valid tool without adverse impact.
The Bottom Line for Employers
AI hiring tools can genuinely improve efficiency and reduce human bias — but they can also encode it at scale. The EEOC's consistent message since 2022 is that employers who adopt AI tools are adopting the legal obligations that come with them: test for adverse impact, document the validation, provide accommodations, and don't assume your vendor's compliance claim covers your liability.
The employers with the most exposure are those who deployed AI hiring tools at scale without testing outcomes, relied on vendor compliance claims without documentation, and have no process for candidates to request human review. Do the adverse impact analysis, keep the records, and build the human override into your pipeline.