AI compliance for HR & hiring
Whether you build HR AI or deploy it, Fairly AI tests it for bias, stops prohibited interview questions, monitors every AI hiring law where you operate, and keeps the evidence buyers, regulators and courts ask for.
Who it’s for
AI hiring laws put different duties on the company that builds a tool and the company that uses it. Start with your side.
HR AI solution developers
Enterprise buyers now ask for bias audits, EU AI Act documentation and regulatory disclosures before they sign. Fairly AI gives you independent, shareable evidence, so compliance closes deals instead of stalling them.
HR AI solution deployers
When a hiring tool discriminates, regulators and plaintiffs come to the employer, not the vendor. Vendor contracts often disclaim compliance, and laws in California and New Jersey reach tools you bought. Fairly AI lets you prove every AI-assisted decision was fair.
Both
Build your own screening model and buy others? You carry developer and deployer duties at once, and Colorado voids contract terms that shift liability for a party’s own violations.
AI hiring rules change month by month. Our research team tracks new laws, regulations and court decisions in every jurisdiction you operate in, and updates your controls and evidence requirements, so your program never goes stale.
The U.S. and beyond
Only New York City requires an independent bias audit. Everywhere else, bias testing is your evidence in a discrimination claim or something the regulator expects — and anti-discrimination law reaches AI outcomes in every market. What really changes by region is notices, human review, and whether you may collect demographic data to test at all.
No federal AI hiring statute — but Title VII, ADEA and ADA apply to AI selection tools like any other test, and in Mobley v. Workday a court let claims proceed against a vendor as the employer’s agent. The new rules are state and local.
Independent bias audit within 12 months of use, a public summary, and notice to candidates 10 business days ahead.
Discriminatory automated-decision outcomes are unlawful; anti-bias testing counts as evidence; keep decision data four years.
Pre-use notice, opt-out, access rights and a written risk assessment before using automated decisions in hiring.
Developer disclosures; deployer notice, adverse-outcome explanation within 30 days, three-year records. Enforcement currently stayed by court order.
No AI with discriminatory effect, no zip codes as a proxy for protected classes, notice of AI use.
“AI is not a defense” (CT); disparate impact unlawful even when the tool is vendor-supplied (NJ); intentional AI discrimination banned (TX).
Outside the EU, few countries have AI-specific hiring laws. The duties come through data protection — automated-decision rights, notices and limits on collecting the very attributes a bias test needs.
Recruitment and promotion AI is high-risk. Providers and deployers each carry their own duties, and it reaches any tool sold into the EU.
UK GDPR Arts. 22A–22D: tell candidates, let them contest, provide meaningful human review. The ICO expects bias monitoring.
Ontario job postings must disclose AI screening (25+ staff); Quebec Law 25 gives notice, explanation and human review of automated decisions.
Privacy policies must disclose decisions made solely or substantially by automated means.
DPDP Act notice and consent duties — including a lawful basis before reusing candidate data for bias testing.
Singapore’s Workplace Fairness Act targets end-2027; Japan sharply limits which candidate attributes can be collected; DIFC Regulation 10 governs autonomous systems in Dubai.
Duties follow where the job and the candidate are, not where you’re headquartered. Planning summary as of November 2026, not legal advice.
Beyond bias
Generative AI now writes job ads, asks interview questions and explains decisions. Each of those words is regulated somewhere. Fairly AI checks every stage of the hiring flow against every jurisdiction you hire in.
Ontario: disclose AI use and drop “Canadian experience” requirements. Japan: no age limits or indirect discriminatory criteria. Singapore: no age, race, gender or religion criteria unless job-related.
NYC audit of scoring and ranking. Illinois bans zip codes as a proxy for protected classes. California and Colorado automated-decision rules apply. Candidate sourcing in Japan may need government notification.
U.S.: ADA, GINA and state inquiry laws bar probing disability, genetic information, age and other protected traits. Japan: fair-selection guidance lists topics interviewers must avoid, from family origin and birthplace to religion and beliefs. Illinois and Maryland require consent for video and facial analysis; Connecticut requires notice when AI talks to applicants.
Transcription error rates by accent, dialect and speech disability. Impact ratios on scores. Explanation rights in the UK, Quebec and Colorado.
Human review must be meaningful (UK), the decision must not be “exclusively automated” (Quebec), and candidates can ask for review (Colorado, California). In Australia, a human final step is not a safe harbour.
A red-team suite tries to make your interview AI ask prohibited questions, jurisdiction by jurisdiction, on every release.
Generated questions are filtered in real time by the candidate’s jurisdiction — and every block is logged as evidence.
How it works
Three stages turn each obligation into reusable control objectives — set by jurisdiction and by role, and enforced from design through production.
01 · Asenion Assess
02 · Asenion Verify
03 · Asenion Witness
Built for: resume screening · video interview analysis · performance appraisals · retention and turnover prediction
What we test and assure
Five assurance tests, each built from the requirements of the jurisdictions you hire in, so every result maps to a control.
01 · Prohibited-question detection
Every AI-generated question is checked against a library of requirements by jurisdiction. It catches questions on topics a law restricts, and flags when one is asked too early in the process.
02 · Speech-to-text bias
We measure transcription error rates by accent, dialect and speech disability. A candidate who is misheard is scored unfairly, even when the scoring model itself passes its bias test.
03 · Summary accuracy
AI summaries are checked against the source transcript for invented, omitted or distorted statements, before a recruiter relies on them to make a decision.
04 · Video interview transparency
We test the candidate flow against each rule that applies: the Illinois AI Video Interview Act and BIPA, Maryland facial-recognition consent, Connecticut’s AI-interaction notice, and EU AI Act transparency duties.
05 · Bias testing, two ways
A single bias test answers only one of those questions. We run both and report both.
Equal population distributions, with matched profiles that vary one protected attribute at a time. This isolates what the model itself does.
Populations that mirror each jurisdiction’s regional demographics. This shows the adverse impact real candidates would see in that market.
U.S. obligations matrix
Every U.S. test and audit for hiring AI, labeled by owner and by how binding it is: mandated, effectively required, incentivized, or conditional.
Obligations in effect or with a confirmed effective date, as of September 2026. A planning reference, not legal advice.
Why Fairly AI
Hiring is one of the most ambiguous, highest-impact areas of AI: every decision changes someone’s livelihood. Our diverse, global team brings ethics, compliance and AI assurance research from leading institutions, and contributes to the bodies shaping how AI is tested.
Ethics, compliance and AI assurance research
Contributing member
Case studies
The question: is first-round candidate scoring biased, and does the process meet employment and privacy rules?
The question: what can go wrong when a recruiting agent acts on its own — and how do you prove it doesn’t?
Usually not. Colorado looks at whether AI materially influences the decision, and the EU AI Act classifies tools by their intended purpose. Human review is a required control, not an exemption.
Not always. The obligation sits with the employer, and a generic audit may not cover your data or jurisdictions. Fairly AI shows where a vendor audit holds and where it doesn’t.
No. Fairly AI runs the program and holds the evidence; your counsel reviews it with far less chasing.
The rules of wherever the job and the candidate are — not where you’re headquartered. Fairly AI maps each tool to every jurisdiction it touches, including what demographic data you may lawfully use to test it there.
A 30-minute call to map your hiring AI against the laws that apply to you.
Schedule a discovery callFairly AI is part of Asenion
In 2025, Fairly AI joined with anch.AI to form Asenion. Fairly AI is Asenion’s solution for HR AI — built for the companies that develop hiring AI and the employers that deploy it. For AI governance, assurance and security beyond hiring, across 15 industries from banking to healthcare, visit Asenion.