Skip to content

AI interview compliance: EU, UK, US.

What the EU AI Act, the GDPR, UK law and US city and state laws require when you use AI interviews, with dates, penalties and a checklist per region.

Updated 17 November 2026 · 22 min read

On this page

In short

  • EU: an AI interviewer that evaluates candidates is high-risk under the AI Act. Your duties as deployer apply from 2 December 2027; emotion recognition at work has been banned since 2 February 2025.
  • GDPR, now: tell candidates, run a DPIA, and never reject anyone on the AI's output alone.
  • UK: no AI act. Solely automated decisions need safeguards since 5 February 2026, and the Equality Act requires reasonable adjustments, such as typing instead of speaking.
  • US: Title VII and the ADA apply to AI tools. New York City requires a bias audit and notice 10 business days ahead; Illinois requires notice, and consent for AI-analyzed video; Colorado and California add duties from 1 January 2027; New Jersey's rules treat an untested automated tool as a disparate-impact risk.
  • Everywhere: tell candidates, keep a trained person in every decision who can overrule the AI, offer an alternative to voice, and keep records.

Not legal advice. A summary of the law as of 17 November 2026. Check your situation with a lawyer and re-check dates before relying on them: several of these rules are new.

Before you start

  • You carry the duties. Under each of these laws the employer using the tool is responsible: the deployer (AI Act), the controller (GDPR), the employer (Equality Act, Title VII). The vendor's duties come on top, not instead.
  • The candidate's location counts. The EU rules apply to candidates in the EU, New York City's to jobs in the city, Illinois's to candidates applying for jobs in the state, wherever your company is.
  • Dates are as of 17 November 2026. Several rules are new or not yet enforced. Re-check before relying on a date.

European Union

The AI Act: AI interviews are high-risk

Annex III, point 4(a) lists AI systems used "to analyse and filter job applications, and to evaluate candidates" as high-risk. An AI interviewer that scores or summarizes answers is one. The Article 6(3) way out (a narrow procedural task, or only preparing a human assessment) doesn't help: a system that profiles people is always high-risk, and rating a candidate's answers against job criteria is profiling.

The Act reaches you if the system's output is used in the EU (Article 2(1)(c)), wherever your company or the vendor is based.

DateWhat applies
2 February 2025Banned practices (Article 5), including emotion recognition at work; AI literacy (Article 4)
2 August 2025Penalties for banned practices become enforceable
2 August 2026Article 50: an AI system that talks to people must make clear it is an AI
2 December 2027High-risk duties for Annex III systems, for providers and deployers (Articles 8 to 27, 49, 86)

Two sets of Commission guidelines matter here. The guidelines on the Article 50 transparency duties, published in July 2026 and adopted as C(2026) 7702 on 5 November 2026, explain how a system that talks to people makes clear it is an AI. The guidelines on high-risk classification, which include employment, were still a draft (May 2026) on 17 November 2026. Until they are final, read Annex III as written and treat an AI interviewer that evaluates candidates as high-risk.

Banned since February 2025: emotion recognition

Article 5(1)(f) bans AI that infers emotions of people in the workplace, except for medical or safety reasons. It covers inferring enthusiasm, stress, confidence or honesty from a face or a voice. The Commission's guidelines on prohibited practices (February 2025) count candidates in a selection process as part of the workplace. A tool that scores "tone", "energy" or "sentiment" from video or audio is not usable in the EU.

Who does what: provider and deployer

The vendor is the provider; your company, using the system under its authority, is the deployer. If you put your name on the system, change it substantially or use it for a purpose the vendor didn't intend, you take on the provider's duties (Article 25).

DutyArticleWho
Risk management; training data checked for bias9, 10Provider
Technical documentation, automatic logging, instructions for use11, 12, 13Provider
Human oversight built into the design; accuracy and robustness14, 15Provider
Conformity assessment, CE marking, EU database registration43, 48, 49Provider
Use the system as the instructions say26(1)Deployer
Assign oversight to people with the competence, training and authority to overrule it26(2)Deployer
Make sure the input data you control is relevant to the role26(4)Deployer
Monitor it; on a risk, suspend use and tell the provider26(5)Deployer
Keep the logs it generates for at least six months26(6)Deployer
Inform workers' representatives and affected workers before use at work26(7)Deployer
Tell candidates they are subject to a high-risk AI system26(11)Deployer
On request, explain the AI's role in a decision about the candidate86Deployer

The GDPR applies now

  • No decision on the AI alone (Article 22). Candidates have the right not to be subject to a decision based solely on automated processing with significant effects. Rejecting on a score nobody reviews is such a decision. In SCHUFA (C-634/21) the Court of Justice held that a score counts as the decision when it plays a determining role, so a reviewer who rubber-stamps it doesn't take you out of Article 22.
  • Tell candidates (Articles 13 and 14). What you collect, why, the lawful basis, who receives it, how long you keep it and their rights; and, where Article 22 applies, meaningful information about the logic involved. In Dun & Bradstreet Austria (C-203/22), February 2025, the Court said that means the procedure and principles actually applied, in a form the person can understand.
  • Run a DPIA first (Article 35). Evaluating candidates with new technology meets the EDPB's criteria for a data protection impact assessment. Article 26(9) of the AI Act tells deployers to use the provider's instructions for it.
  • Answer requests. Access (Article 15) includes the transcript and the AI's output about the person; erasure (Article 17) once you no longer need the data.

Works councils

Article 26(7) of the AI Act adds to national rules that often apply already. In Germany the works council co-determines selection guidelines (§95 BetrVG) and technical systems able to monitor performance (§87(1) no. 6), and must be informed of planned AI (§90). France's CSE is consulted before new technology is introduced (Code du travail L2312-8). In the Netherlands the works council must consent to personnel assessment and data-processing systems (WOR, Article 27). Start these talks before the pilot, not after.

What to do in the EU

  1. Check for emotion recognition now. Ask every vendor in writing whether its tool infers emotions, personality or honesty from face or voice. If yes, stop using it for EU candidates.
  2. Keep a person in every decision. Name the reviewers, train them on what the output means and doesn't, and give them the authority to overrule it. Record who decided.
  3. Update the candidate notice. Say that an AI conducts or evaluates the interview, what it assesses, that a person decides and how to ask for a human review.
  4. Run or update the DPIA before the first EU candidate, using the vendor's documentation.
  5. Inform your works council or worker representatives before rollout, where national law or Article 26(7) requires it.
  6. Keep logs for at least six months and set retention limits for everything else.
  7. Get the AI Act documents from the vendor: instructions for use, its classification of the tool, and its plan for conformity by 2 December 2027.

United Kingdom

The UK has no AI act. AI interviews fall under the UK GDPR, the Data Protection Act 2018 and the Equality Act 2010, which apply to them as to any other selection method.

Automated decisions since 5 February 2026

Section 80 of the Data (Use and Access) Act 2025 replaced Article 22 of the UK GDPR with Articles 22A to 22D, in force since 5 February 2026. A decision is "solely automated" when there is no meaningful human involvement in it; rejecting a candidate is a significant decision.

  • Ordinary data: solely automated significant decisions are now allowed on any lawful basis, except recognised legitimate interests, if the safeguards below are in place.
  • Special category data (health, ethnicity, religion and the like): only with explicit consent, or when needed for a contract or required by law with a substantial public interest condition (Article 22B).
  • The safeguards (Article 22C): tell the candidate about the decision, let them make representations, give them human intervention and let them contest it.

A reviewer who reads the evidence and can overrule the result keeps the decision out of these rules. The ICO's draft guidance on automated decision-making, consulted on from 31 March to 29 May 2026, is not yet final.

What the ICO found in AI recruitment

In November 2024 the ICO published the outcomes of its audits of AI sourcing, screening and selection tools: 296 recommendations, addressed to providers and recruiters alike.

  • Inferred characteristics. Some tools estimated gender or ethnicity from a name or an application. The ICO's view: inferred data is not accurate enough to monitor bias, and was often processed without a lawful basis.
  • Filtering by protected characteristics. Some tools let recruiters filter candidates out on them.
  • Too much data. Some collected far more than needed, or combined it with data scraped from social media.
  • Unclear roles. Providers called themselves processors while acting as controllers, and pushed responsibility onto recruiters through vague contracts.

Its key recommendations: monitor fairness and bias with accurate data; explain to candidates what the AI processes and the logic behind its output; collect only what you need; complete a DPIA before processing; fix controller and processor roles in the contract; give the provider written processing instructions; and record a lawful basis. Its March 2026 report Recruitment rewired, on more than 30 employers, found that many believed a person reviewed outcomes when the review wasn't meaningful, and that candidates were rarely told how automation was used.

The Equality Act 2010

  • Indirect discrimination (section 19). A method that applies to everyone but puts a group at a particular disadvantage is unlawful unless it is a proportionate means of a legitimate aim. A voice interview that scores fluency disadvantages candidates with a speech impairment and non-native speakers; scoring what the job needs, from what was said, does not.
  • Reasonable adjustments (sections 20 and 21). You must adjust for disabled candidates: typed answers instead of speech, extra time, a different format or a human interview. Tell candidates how to ask before the interview starts.
  • Your liability. The employer answers for its selection method, whoever built the tool.

What to do in the UK

  1. Make human review meaningful: a trained reviewer reads the evidence, can overrule the AI, and the review is recorded.
  2. Update the candidate notice: AI use, what it assesses, how results are used, and how to ask for an adjustment or a human review.
  3. Offer an alternative to voice and agree any adjustment before the interview.
  4. Complete a DPIA before the first candidate, and update it when the tool changes.
  5. Settle roles in the contract: the vendor as processor, with written instructions on data, purposes and outputs.
  6. Monitor outcomes for bias with data candidates gave you themselves, never inferred data.

United States

No federal law regulates AI in hiring. Federal anti-discrimination law applies to an AI interview as to any test, and a city and five states add their own duties. Which rules apply depends on where the job and the candidate are, not where your company is.

Federal: Title VII and the ADA

  • Title VII. A selection method with a disparate impact on a race, color, religion, sex or national origin is unlawful unless it is job-related and consistent with business necessity (42 U.S.C. §2000e-2(k)). The employer is liable for a vendor's tool.
  • The ADA. Candidates with disabilities are entitled to reasonable accommodation in the interview (29 CFR 1630.11), and a test must measure the skill, not the impairment. A voice-only interview with no other way to answer fails that for some candidates.
  • What changed at the EEOC. In January 2025 the EEOC removed its technical assistance on AI in hiring (the 2022 ADA document and the 2023 Title VII document) from its website; the Justice Department's 2022 page on AI and disability discrimination in hiring is still on ada.gov. Executive Order 14281 (April 2025) told federal agencies to deprioritize disparate-impact enforcement, and the EEOC is rescinding parts of the Uniform Guidelines on Employee Selection Procedures: a final rule has been under White House review since October 2026 and, as of November 2026, is not published.
  • What didn't change. The statutes. Candidates can still sue: in Mobley v. Workday a federal court let an age discrimination claim against an AI screening vendor proceed as a nationwide collective action in May 2025; the opt-in period closed in March 2026, and as of November 2026 there is no finding of liability. State agencies enforce their own laws.

New York City: Local Law 144

Local Law 144 applies when an automated employment decision tool substantially assists a decision on a job in the city. A tool that scores candidates and drives who moves on is one. Before using it you must:

  • Have a bias audit done by an independent auditor within the year before use: selection or scoring rates and impact ratios by sex, race/ethnicity and their intersections. Neither you nor the vendor can be the auditor.
  • Publish the audit summary on your careers site: its date, the impact ratios, the number of candidates in unknown categories and the tool's distribution date.
  • Notify candidates who live in the city at least 10 business days before use: that the tool is used, the job qualifications and characteristics it assesses, and how to request an alternative process or an accommodation. The type and source of the data collected and your retention policy go on your website or to anyone who asks in writing, within 30 days.

Illinois: video interviews and the Human Rights Act

  • AI Video Interview Act (820 ILCS 42), since 2020. If AI analyzes recorded video interviews: tell the candidate before the interview, explain how the AI works and what it evaluates, and get their consent. Share the videos only with those whose expertise is needed to evaluate them, and delete them, with all copies, within 30 days of the candidate's request. If you rely solely on AI to choose who gets an in-person interview, report the race and ethnicity of those chosen and not chosen to the Department of Commerce and Economic Opportunity every year.
  • Human Rights Act amendment (HB 3773, 775 ILCS 5/2-102(L)), since 1 January 2026. Tell candidates when you use AI in recruitment and hiring decisions; don't use AI that has the effect of discriminating on a protected class; don't use zip codes as a proxy for one. The Department of Human Rights proposed notice rules in May 2026 and withdrew them in June; until new ones are adopted, follow the statute.

Colorado: SB 26-189

Colorado's 2024 AI Act, due on 30 June 2026, never took effect. SB 26-189, signed on 14 May 2026, repealed and replaced it from 1 January 2027, with a narrower set of duties for automated decision-making technology that materially influences a hiring decision:

  • Notice before use, clear and conspicuous.
  • After an adverse decision, within 30 days: a plain-language explanation of the decision, the technology's role and how to get more information.
  • Rights: to correct inaccurate personal data (not scores or opinions) and to ask for human review of an adverse decision, to the extent commercially reasonable.
  • Records kept for three years.

California: two sets of rules

  • Civil Rights Council regulations, since 1 October 2025. The Fair Employment and Housing Act applies to automated-decision systems. A vendor acting for you is your agent, so its tool's discrimination is yours. Tools that analyze voice, facial expression or reaction time can discriminate on disability or other protected traits. Anti-bias testing (its quality, recency and what you did with the results) counts for or against you in a claim. Keep the system's data and other employment records for four years.
  • CCPA rules on automated decision-making technology (ADMT), from 1 January 2027. If a tool replaces or substantially replaces human decision-making in hiring: a pre-use notice (its purpose, how it works, the data that affects the output, the candidate's rights), a right to access information about how it was used on them, a right to opt out, and a risk assessment before use. For hiring, you may decline the opt-out if the tool works as intended and doesn't discriminate. A reviewer who understands the output, analyzes it with other information and has the authority to change the decision takes the tool outside these rules.

New Jersey: disparate-impact rules

Since 15 December 2025 the Division on Civil Rights' disparate impact rules (N.J.A.C. 13:16) apply the Law Against Discrimination to automated employment decision tools, defined to include tools used for screening and interviewing. A tool that hasn't been adequately tested and shown not to adversely affect a protected class before use "may have a disparate impact"; a tool that screens out applicants by schedule must let them request a reasonable accommodation. A practice with a disparate impact is unlawful unless it is job-related and consistent with a legitimate business necessity and no less discriminatory alternative would do. The rules add no notice or audit duty; the practical reading is to test for adverse impact before use and keep the results.

Maryland: facial recognition

Since 1 October 2020 an employer may not use facial recognition during an interview to create a facial template unless the candidate signs a waiver giving their name, the date of the interview, their consent and whether they read the waiver (Labor and Employment §3-717).

What to do in the US

  1. Map where you hire: New York City, Illinois, Colorado, California, New Jersey and Maryland each add duties; elsewhere, federal and state anti-discrimination law applies.
  2. For New York City jobs, commission the independent bias audit, publish its summary and send the notice 10 business days before use.
  3. Give the Illinois notice now, and get consent if AI analyzes video.
  4. By 1 January 2027, prepare the Colorado notice and adverse-decision explanation and the California pre-use notice, access process and risk assessment.
  5. Offer accommodations and an alternative to voice, and say how to ask, before the interview.
  6. Check adverse impact before use (New Jersey expects it) and on your own outcomes, with demographic data candidates provide voluntarily, kept apart from the decision.
  7. Keep records four years for California, three for Colorado.
  8. Don't use facial recognition in interviews unless you have a signed Maryland waiver.

The regions side by side

EUUKUS
Notice to candidatesYes: GDPR now; AI Act from 2 Dec 2027 (high-risk use), Article 50 since 2 Aug 2026 (talking to an AI)Yes: UK GDPR, including the logic of any automated decisionNYC 10 business days ahead; Illinois since 2026; Colorado and California from 2027
ConsentNot the usual basis; explicit consent is one way to allow a solely automated decisionExplicit consent needed for solely automated decisions on special category dataIllinois for AI-analyzed video; Maryland for facial recognition
Bias audit or testingProvider tests its data (Article 10); you monitor in useICO expects fairness monitoring with accurate dataNYC independent audit each year; New Jersey expects testing before use; elsewhere adverse-impact checks are your defense
Human reviewOversight by trained people with authority (Article 26); no decision solely by the AI (GDPR 22)Meaningful human involvement, or the Article 22C safeguardsColorado on request; California's ADMT rules don't apply when a person meaningfully decides
Candidate rightsAccess, erasure, explanation (Article 86), contest an automated decisionAccess, erasure, representations, human intervention, contestNYC alternative process; Illinois video deletion in 30 days; Colorado correction and review; California access and opt-out
RecordsAI logs at least 6 monthsAs long as needed, set in your retention policyCalifornia 4 years; Colorado 3 years
Maximum penaltyAI Act €35m or 7%; GDPR €20m or 4%£17.5m or 4%; uncapped tribunal awardsNYC $1,500 per violation per day; CCPA $7,988; Colorado $20,000 per violation

How testee supports each duty

testee makes several duties easier to meet. It doesn't meet them for you: your company is the deployer, the controller and the employer, and testee doesn't claim to be certified or to make you compliant.

DutyWhat testee doesWhat stays with you
Tell candidates an AI is usedThe invitation page says the interview is with an AI interviewer and lists what is recorded and analyzed. Candidates accept the terms before Start; the server refuses a start without it and stores the accepted text's version and time.Your privacy notice, the NYC notice 10 business days ahead, and the wording your state requires
Recording consentCamera and screen recording is off by default. When on, candidates see what is recorded, why, who can watch it and for how long, and agree first.Deciding whether recording is needed at all
An alternative to voiceText mode: the candidate can choose to type every answer instead of speaking, on the invitation page. Captions are always on.Other adjustments, such as a human interview
People decideThe AI coverage score is labeled "for review, not a hiring decision". The decision (advance, hold, reject) is set by a person, needs a note, and its history of changes is kept. testee never rejects a candidate on its own.Training reviewers and giving them the authority to overrule
Explain the resultEach expectation's status comes with the exact quotes behind it, the transcript and recording to check them, and a Candidate explanation PDF for the candidate who asks.Answering a candidate's request for an explanation
No emotion or biometric analysisNo emotion, accent, fluency or personality assessment, no face analysis, no biometrics. Video is for people to watch; no AI processes it.Checking the other tools in your process
Records and logsThe consent record on every session, the decision history, and an audit log of who did what in your company, kept 365 days.Keeping exports longer where your law asks for it
RetentionBy default, candidates with no activity are erased after 365 days, and audio and recordings are deleted 180 days after the interview; each company can change both.Choosing periods that fit your law
Erase and exportErase a candidate everywhere in one step; export their data as a file for an access request.Handling the request and its deadline
Bias audittestee doesn't ask for, infer or store candidates' protected characteristics.Collecting voluntary demographic data and commissioning the NYC audit

Data is stored in the EU; the voice interviewer, transcription and analysis run on OpenAI's API in the US, and no customer data trains AI models. Details are on the Trust page and in the data processing agreement.

Questions and answers

Are AI interviews legal?

Yes, in the EU, the UK and the US, when you meet the duties that come with them: tell candidates, keep a person in the decision, offer an alternative to voice, and keep records. Some practices are not allowed: in the EU, AI that infers emotions from a candidate's face or voice has been banned since 2 February 2025; in Maryland, facial recognition in an interview needs a signed waiver; in Illinois, AI analysis of video interviews needs notice and consent first.

Is an AI interview high-risk under the EU AI Act?

Yes, if it evaluates candidates. Annex III, point 4(a) lists AI used to analyse and filter applications and to evaluate candidates, and the Article 6(3) exception never covers a system that profiles people. The deployer's high-risk duties apply from 2 December 2027, after the Digital Omnibus moved them. The Commission's classification guidelines were still a draft on 17 November 2026, so we read Annex III as written and treat such a tool as high-risk.

What does NYC Local Law 144 require?

Before using an automated employment decision tool for a job in New York City, you need an independent bias audit done within the past year, a published summary of its results, and a notice to candidates who live in the city at least 10 business days before use, saying how to ask for an alternative process or an accommodation. Each day of use without an audit, and each candidate not notified, is a separate violation, with penalties of up to $1,500 each.

Does the GDPR allow AI interviews?

Yes, with a lawful basis (consent is rarely the right one between an employer and a candidate), a clear notice under Articles 13 and 14, and a DPIA before you start. Article 22 gives candidates the right not to be subject to a decision based solely on automated processing, so a person must review the evidence and be able to overrule the AI before anyone is rejected.

What does the Illinois AI Video Interview Act require?

If AI analyzes recorded video interviews for a job in Illinois: tell the candidate before the interview, explain how the AI works and what it evaluates, and get their consent; share the videos only with those needed to evaluate them; delete them within 30 days of the candidate's request; and, if AI alone decides who gets an in-person interview, report demographic data every year. The Act speaks of video. For a voice or text AI interview we take the conservative reading and give the same notice, explanation and consent; the Human Rights Act's AI notice applies since 1 January 2026 in any case.

Does a human have to make the hiring decision?

In the EU, in practice yes: the GDPR's Article 22 and the AI Act's human oversight duty mean a trained person with the authority to overrule the AI decides, and a rubber stamp doesn't count. The UK allows solely automated decisions on ordinary data with safeguards, and California's and Colorado's rules apply less, or not at all, when a person meaningfully decides. The safe course everywhere is the same: a person reviews the evidence and makes, and records, every decision.

Sources

  1. Regulation (EU) 2024/1689, the AI Act
  2. Regulation (EU) 2026/1744, the Digital Omnibus on AI
  3. European Commission, Guidelines on the transparency obligations under Article 50 of the AI Act, C(2026) 7702
  4. European Commission, Draft guidelines on the classification of high-risk AI systems (May 2026)
  5. Regulation (EU) 2016/679, the GDPR
  6. Court of Justice, C-634/21 SCHUFA Holding
  7. Court of Justice, C-203/22 Dun & Bradstreet Austria
  8. Data (Use and Access) Act 2025, section 80
  9. The Data (Use and Access) Act 2025 (Commencement No. 6) Regulations 2026, SI 2026/82
  10. Equality Act 2010
  11. ICO, AI tools in recruitment: audit outcomes report (November 2024)
  12. ICO, Recruitment rewired (March 2026)
  13. Title VII, 42 U.S.C. §2000e-2
  14. ADA regulations, 29 CFR 1630.11
  15. NYC DCWP, Automated employment decision tools (Local Law 144)
  16. NYC DCWP, penalty schedule for automated employment decision tools (6 RCNY §6-81)
  17. New York State Comptroller, Enforcement of Local Law 144 (December 2025)
  18. Illinois Artificial Intelligence Video Interview Act, 820 ILCS 42
  19. Illinois Public Act 103-0804 (HB 3773)
  20. Colorado SB 26-189
  21. Colorado Session Laws 2026, chapter 131 (SB 26-189 as signed)
  22. Colorado Department of Law, proposed ADMT rules, 4 CCR 904-6 (August 2026)
  23. New Jersey Division on Civil Rights, disparate impact rules, N.J.A.C. 13:16
  24. California Civil Rights Council, automated-decision systems regulations
  25. California Privacy Protection Agency, CCPA regulations
  26. CPPA, 2025 adjusted fines
  27. Maryland Labor and Employment §3-717

Try it onyour next role.

Build a structured AI interview from your job description in minutes.14 days free, no card needed.