IIF Certified Assessment
Start IQ Test
IIF Certified Assessment
Recruitment screening and candidate assessment are high-risk uses under the EU AI Act, and their obligations were set to begin on 2 August 2026. A regulation in force since late July moved that date to 2 December 2027 — while leaving one workplace ban already biting.
If you applied for a job in the European Union this month, there is a fair chance something automated read your application before a person did. From 2 August 2026 that was supposed to sit inside a compliance regime: risk assessments, technical documentation, bias testing, logged human oversight, information given to the people affected. The date arrived on schedule. The obligations did not.
The EU AI Act entered into force on 1 August 2024 and sorts systems by the risk of the use rather than by the sophistication of the technology. Employment is one of the categories it treats most seriously. Annex III lists AI used in recruitment and selection — publishing targeted job advertisements, analysing and filtering applications, evaluating candidates — as high risk, and extends the same classification to systems that make or inform decisions about promotion and termination, the allocation of tasks, and the monitoring and evaluation of performance at work.
High risk is not a ban. It is a set of duties falling mainly on whoever builds the system and, in narrower form, on whoever deploys it: a documented risk-management process, data governance covering the training data, technical documentation, record-keeping, accuracy and robustness requirements, meaningful human oversight, and information supplied to the people the system is used on. For an assessment vendor this is close to the discipline a psychometrician would recognise from a test manual — evidence that the instrument measures what it claims, gathered before it is used to make decisions about people.
That is precisely the discipline missing from most automated screening today, which is why the 2 August 2026 date mattered to candidates and not only to compliance teams. It was the point at which a scored, ranked or filtered application was meant to come with an evidence trail behind it.
The European Commission proposed a package of simplification measures — the Digital Omnibus — on 19 November 2025, one element of which was a deferral of the high-risk deadlines. Parliament and Council reached a provisional political agreement in early May 2026, the Council gave final approval at the end of June, and the regulation was published in the Official Journal on 24 July 2026 and entered into force on 27 July. Legal analyses of the final text put the new dates at 2 December 2027 for stand-alone Annex III systems, the category that covers recruitment, and 2 August 2028 for high-risk AI embedded in products already regulated under other EU legislation.
The last line is the one most relevant to assessment. Among the prohibited practices is the use of AI to infer emotions from a person in the workplace or in education, outside narrow medical and safety exceptions. A vendor offering to read confidence, enthusiasm or stress from a candidate's face or voice during a video interview is not waiting for a 2027 deadline. That practice has been unlawful in the EU for more than eighteen months.
The deadline moved. The decisions being made about candidates did not pause for it.
Nothing, which is the point. Screening tools stayed in service throughout the legislative process, and the duties that would have made them auditable are now scheduled for December 2027. National employment and equality law still applies, and the General Data Protection Regulation still gives a data subject rights in relation to decisions based solely on automated processing. But the specific machinery the AI Act was going to install — the bias testing, the technical file, the demonstrable human in the loop — is not yet a legal requirement anywhere in the Union.
It is worth being precise about what the delay does and does not do, because both sides of this argument have an interest in overstating it. It repeals nothing: the classification of recruitment AI as high risk survives untouched, and the obligations are the same obligations, arriving sixteen months later. Nor does it create a vacuum, since data protection law, national employment law and equality law were never suspended. What it postpones is the specific evidence requirement — the documentation a vendor must hold, the bias testing it must have carried out, the oversight a deployer must be able to demonstrate — which is the only part of the regime that would have made an automated rejection auditable from outside the company that made it.
It is worth separating two things the phrase "AI hiring test" runs together. A cognitive ability test is a measuring instrument with published norms, a documented standardisation sample, a reported reliability coefficient and a standard error of measurement — the figure telling you how much a person's score would move between sittings for reasons that have nothing to do with ability. A screening model is a prediction engine trained on one employer's past decisions, and it inherits whatever was in them.
The evidence for what a cognitive score predicts at work is real, and considerably more modest than vendor material implies: we have gone through what the research supports and what it does not, and the honest summary is a moderate group-level association with training and job performance, with very wide spread around it for any individual. A screening model has no equivalent literature, because it is a different model in every company that runs one.
The second difference is one candidates can act on. Standardised tests have a known and measurable practice effect — scores rise on a second sitting of the same instrument for reasons unrelated to ability, which is why publishers set retest intervals and why an employer should be asking when you last sat the thing. A screening model publishes no such interval and gives no such feedback. You cannot practise for it, and you will not be told how it scored you.
A third difference decides whether either number means anything, and it is the one to ask about. A normed test tells you what its score is a position in: 118 on a mean-100, standard-deviation-15 scale corresponds to a definite share of the reference population, and the percentile calculator converts it in a line. A screening model outputs a rank inside one applicant pool for one vacancy, which is not a property of you at all — change the pool and the number changes while you stay exactly as you were. Anyone told they have been scored by an automated system is entitled to ask which of those two things they were handed. If you have never seen the first kind, sitting a properly normed reasoning test once is the fastest way to learn what the difference looks like from the candidate's side.
If an employer asks you to sit a reasoning or aptitude test, useful preparation is familiarity with the format rather than cramming. Sit one timed test under quiet conditions so the clock is not a surprise, then read the result as a range rather than a point — our percentile calculator shows what a score means against a reference population.
Find your IQ score now! →The broader lesson is one this desk keeps returning to. A number produced about a person — by a test, by a model, by a recruiter's spreadsheet — is only as good as the evidence that it measures what it claims and the conditions under which it was collected. Regulation is one way of forcing that evidence into the open, and it has just been deferred by sixteen months. Until it arrives, the questions are the ones you would ask of any assessment: what was measured, against which reference group, with how much error, and who checked. That checklist is not specific to employers — it is the same one worth running over any test you are asked to sit online, whoever is doing the asking.
Yes, but the main obligations are not yet in force. AI used for recruitment, candidate filtering, evaluation, promotion, termination, task allocation and performance monitoring is classified as high risk under Annex III of the EU AI Act. Those duties were scheduled to apply from 2 August 2026 and were deferred to 2 December 2027 by the Digital Omnibus regulation, which was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.
Yes. The Article 50 transparency obligations were not postponed and applied from that date, with a short grace period into December 2026 for marking requirements on systems already on the market. The Article 5 prohibitions have applied since 2 February 2025 and were not affected either.
Not in the EU. Inferring emotions from a person in the workplace or in education is among the practices prohibited by Article 5 of the AI Act, outside narrow medical and safety exceptions, and that prohibition has applied since 2 February 2025. It was not deferred by the 2026 simplification regulation.
No. A cognitive ability test is a measuring instrument with published norms, a documented standardisation sample, a reliability coefficient and a standard error of measurement, and there is a substantial research literature on what its scores predict. A screening model is trained on one employer's own past decisions, has no published norms, and is not comparable between companies.
Corrections: spotted an error? Email corrections@iqmetrics.org and we will update this story and note the change here.
Aptitude tests are among the better-studied hiring tools, and for decades the headline validity figures were quoted with more confidence than the corrections behind them deserved. A 2022 reanalysis pulled those numbers down.
Sitting the same test twice usually produces a higher second score. Almost none of that gain is ability — which is why employers use alternate forms, and why the second number is a worse estimate than the first.
Our IIF-certified assessment reports your score with its scale, percentile and confidence range — and a breakdown of the cognitive domains behind it.
Start IQ Test →
