Skip to main content

AI hiring rules slip to December 2027. Some bans already apply

Ilustrační obrázek
In 2018 Amazon scrapped an AI recruiting tool that had learned to penalise references to women’s organisations and graduates of women’s colleges. Eight years later, 87% of companies use AI somewhere in hiring, according to a Forbes-cited survey. Europe’s answer is not a single hiring law but a stack of rules that treat recruitment algorithms as high-risk — and an OnLabor analysis argues the United States is still far behind.

Europe puts hiring AI in the high-risk box

The centrepiece is the EU Artificial Intelligence Act, adopted in 2024. It classifies any AI software used in hiring as “high-risk”. That label is not a ban. It is a set of duties that attach before and after the tool reaches an employer.

Developers carry the heaviest load. Under the Act, they must test their models, identify disparate outcomes and ensure that training data are “relevant, sufficiently representative and . . . free of errors”. They must document how the algorithm was built and which data sources it relies on, so regulators can audit a model suspected of producing discriminatory results. If they become aware of a risk that a system is unfair, they must immediately investigate the causes and take corrective action, including retraining or recalling the model.

Employers have duties too. They must keep detailed logs automatically generated by the system, recording every instance of an AI tool’s use. Those logs create an audit trail. They must also provide human oversight by “at least two natural persons with the necessary competence, training and authority” who can properly interpret the system’s output and intervene if outcomes look biased or unreliable. If intervention fails, the model must be retrained or withdrawn.

Enforcement is not left to goodwill. Each member state must designate a single national supervisory authority, while the European Artificial Intelligence Board coordinates oversight across the bloc. Developers must affix a CE marking to signal compliance to business customers. Under Article 99, the €35 million or 7% of worldwide annual turnover cap is the top tier and covers prohibited practices under Article 5. Breaches of the high-risk obligations described here — testing, documentation, logging, human oversight or corrective action — fall in the next tier, capped at €15 million or 3% of worldwide annual turnover. That top-tier figure should not be read as the penalty for a deployer failing to act on a biased model.

The delay — and the bans already in force

Recruitment and selection AI sits in Annex III of the AI Act, so it is high-risk. Those high-risk obligations were originally due to apply from 2 August 2026 and have been pushed to 2 December 2027 by the digital omnibus package (Parliament approved 16 June 2026, Council 29 June 2026). Article 5 has applied since 2 February 2025, and it already bans AI systems that infer emotions in the workplace, with only a narrow medical or safety exception — a prohibition that directly touches the AI video-interview tools described below.

What a rejected candidate can ask for

The AI Act does not stand alone. The GDPR gives people the right not to be subject to a decision based solely on automated processing when it has legal or similarly significant effects. In practice, a job applicant can ask a company how an algorithm affected their outcome and request a human re-evaluation.

That right is more than a complaint mechanism. It forces employers to be able to reconstruct a decision. The AI Act’s logging duty and the GDPR’s explanation right fit together: a rejected candidate should be able to learn that automated decision-making exists, what logic was involved and what consequences followed. They will not receive source code, but they should receive an explanation a person can follow.

The human oversight requirement also sets a practical bar. A signature at the end of a pipeline is not enough if the reviewer has no authority to change the outcome or access to the reasoning. The Act asks for competent people with the authority to intervene; a rubber stamp does not satisfy that.

The American patchwork

The contrast with the United States is stark. Congress has stayed silent, leaving states and cities to fill the gap. New York City’s Local Law 144 mandates annual bias audits and disclosure for employers using AI tools. California requires employers to keep four years of decision-making data from AI systems and to give 30 days’ notice when AI systems are used in hiring.

Those laws push in the right direction, but they remain a geographical patchwork. Most American workers are outside them. Litigation is testing whether existing anti-discrimination law can reach algorithmic bias. In EEOC v. iTutorGroup, the Equal Employment Opportunity Commission brought its first publicly known AI-bias enforcement action, alleging that software automatically rejected female applicants aged 55 or older and male applicants aged 60 or older. The company agreed to pay $365,000 in August 2023.

In Mobley v. Workday, a federal court allowed age discrimination claims to proceed on the theory that Workday’s AI software could act as an “agent” of employers who delegate hiring functions to it. That signalled AI hiring vendors could face direct liability. Other cases are earlier. Harper v. Sirius XM Radio alleges an AI-powered system downgraded applications using proxies such as education, zip code and work history that disproportionately disadvantage Black candidates. ACLU v. Aon Consulting alleges that Aon’s “personality” assessment and automated video interviewing tool, marketed as “bias-free”, screened out racial minorities and people with disabilities at disproportionate rates.

These cases matter because they show existing statutes can reach algorithmic bias. They also show how rare such claims remain relative to AI’s spread through hiring.

Why the hiring pipeline is hard to see

AI enters hiring at several points. Some systems use data from past “successful hires” to target job ads at people with similar profiles, favouring candidates who mirror the existing workforce. Others extract data from resumes and rank candidates on historical patterns. Some go further with video interviews that analyse “non-verbal cues, communication style, professionalism, and listening”. Those criteria can disadvantage candidates whose cultural background sits outside the model’s norm. To the extent such tools infer a candidate’s emotions, Article 5’s workplace ban already applies, subject only to the narrow medical or safety exception.

In many cases, applicants are screened out before reaching a human reviewer. That is the invisible influence the OnLabor analysis highlights. The Amazon episode is the clearest warning: a model trained on a decade of old resumes mostly from men learned to penalise references to women’s organisations and graduates of women’s colleges. The company abandoned the tool, but the lesson remains. Automating hiring can also automate bias.

What European employers should take from it

Buying a compliant tool does not transfer the employer’s responsibility as deployer. A supplier’s compliance folder is a starting point, not a defence. Employers using high-risk hiring systems need to keep the logs, ensure the human reviewers are genuinely competent and empowered, document the training data and test for disparate outcomes. They should also be able to explain, in plain language, how a candidate’s application was assessed.

The market is already moving. The StepStone Group publicly audited its AI recommendation engine for bias, an effort legal experts praised as a model for responsible compliance. TechWolf, a Belgium-based HR tech firm, described the law as managing risks explicitly and safeguarding fairness and transparency. Amazon, Microsoft and OpenAI have voiced support for the Act as a whole.

The OnLabor analysis makes a narrower legal point too: the EU framework combines ex ante rules demanding fairness before deployment with ex post rights letting individuals challenge unfair results. The US has neither a comprehensive federal regime nor a consistent state one. For European employers, the message is less about waiting for a deadline and more about being able to show, today, how a hiring algorithm was tested, logged and overseen.

Does the EU AI Act ban AI in hiring?

No. It classifies any AI software used in hiring as high-risk. That means developers and employers must meet obligations on testing, documentation, logging and human oversight, but the Act does not prohibit hiring algorithms outright.

Can I find out why an AI rejected my application?

Under the GDPR you can ask for information about automated decision-making, the logic involved and its consequences, and you can request human intervention and contest the result. You will not receive source code, but you should receive an explanation a person can follow.

Who is responsible if a recruitment tool misbehaves, the vendor or the employer?

Both can be, but the obligations attached to high-risk use fall on the employer that deploys the system. Buying a tool with a compliance certificate does not move that responsibility to the supplier.

Europe’s approach is not perfect. Enforcement across 27 national authorities will vary, and researchers have not yet been able to assess the Act’s full impact. But the direction is clear: hiring AI is treated as a decision-making system with legal consequences, not as a neutral filing cabinet. For candidates, that means a route to ask questions. For employers, it means the paperwork starts before the shortlist, not after the rejection.

Discussion

No comments yet — be the first to share your thoughts.
X

Don't miss out!

Subscribe for the latest news and updates.