Annex III point 4 lists AI systems intended to be used for recruitment or selection, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates, as high-risk use cases under Article 6(2), subject to the Article 6 conditions and the profiling limit. That listing covers the function, not the employer's label for the tool: a system that filters applications or evaluates candidates must be tested against the listing even if it is described as assistive. Article 5 screening, including biometric and emotion inference prohibitions, and Article 50 transparency duties must be tested separately. This guide does not determine whether any recruitment tool is high-risk, and it does not give employment law advice.
Which recruitment functions are listed in Annex III?
Annex III point 4 lists AI systems intended to be used for recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates. Each of those functions must be tested on its own terms: targeted advertising that selects who sees a vacancy, filtering that removes applications from consideration, and evaluation that scores or ranks candidates are each within the listing's language, and a system that performs any of them should be treated as a candidate for high-risk classification. A system that only hosts vacancies, only stores applications or only schedules interviews, without analysing, filtering or evaluating, is assessed differently, and the distinction must be recorded from the system's actual function. The record should describe what the system does to an application or a candidate, what output it produces and what decision follows, rather than relying on the vendor's product category or the employer's description of the tool as assistive.
How is a recruitment system classified?
Classification follows the Article 6 route: the system's intended purpose is tested against the Annex III listing, the Article 6(2) conditions are applied, and any Article 6(3) question is tested against the narrow tasks and the profiling limit. Profiling is decisive: an Annex III system that performs profiling of natural persons is always high-risk, and the Article 6(3) route is not available for it. A recruitment system that scores, ranks or filters candidates on the basis of their characteristics will often involve profiling, and the classification record should address that question directly, with the definition of profiling and the system's actual operation cited. The provider's assessment under Article 6(4), where it relies on Article 6(3), must be documented before market placement and the system registered, and the deployer should not assume the provider's conclusion without seeing the reasoning. Where classification is contested, qualified review should be sought before the system is used to make or influence hiring decisions.
For the broader regulatory framework, see healthcare and biometrics under the AI Act.
What must be screened before classification?
The Article 5 screen comes before any classification work, and in recruitment it has real bite. Video interviews analysed for facial expression or voice patterns, biometric identification at an assessment centre, and any inference of emotion in a workplace context must each be tested element by element against the consolidated text, because a prohibited use must not proceed to conformity assessment at all. The screen should record, for each prohibition considered, the input data, the inference made and the decision that follows, including where the tool only records video for later human review. Data protection law adds its own conditions on special-category data and on solely automated decisions, and employment law can require information and consultation duties that sit alongside the AI Act. A tool described by its supplier as assistive should be screened on what it does to an application, not on the label it carries.
What oversight and candidate information are needed?
Oversight in hiring has to be more than a signature on a system output. The person reviewing a candidate needs the competence, time and authority to weigh the output against information the system never saw, and to decide differently where the facts require it. Candidates need a clear account, in plain terms, of the AI use, what the system does, what data it uses and how a decision can be challenged, and workers' representatives need the information and consultation employment law requires. Every element should be traceable: who was assigned oversight, what information was given, what decisions were taken, and where a human overrode the system. A hiring process in which a person rubber-stamps a ranked list is not the deployment the Act describes, and the record should make that visible rather than hiding it behind a claim of human involvement.
What evidence should a recruitment deployment keep?
A recruitment file should let a stranger reconstruct the decision. It needs the system and version, the provider and the provider's classification and conformity information, the intended purpose and the functions actually used, the Article 5 screen, the Article 6 classification and reasoning, the data protection assessment, the oversight design, the information given to candidates and workers, the monitoring of outcomes including any disparate impact observed, and the complaints and challenges received. Monitoring should track whether the system performs consistently across groups, whether outputs are followed or overridden, and whether hiring outcomes shift in ways nobody intended. An unexpected pattern should go to the provider, trigger a reassessment and be recorded with the action taken. A file holding only a vendor description and a signed contract does not support a decision that affects people's livelihoods, and will not withstand scrutiny from a candidate, a workers' representative or an authority.
Continue with worker management and monitoring.
What does this guide not decide?
This guide assesses no real recruitment system and judges no hiring decision. Whether a system is high-risk or prohibited, and whether a hiring decision was lawful, depends on the function, the data and inferences, the decision process, and the employment, equality and data protection law that applies, worked through on the specific facts with qualified advice. The validity or fairness of a recruitment tool is a technical and legal question for the evaluation evidence, not something this guide can settle. Before a recruitment system influences a hiring decision, the classification, prohibition, transparency, consultation and data protection positions should be confirmed with a competent authority or legal adviser.
Frequently asked questions
Is a recruitment chatbot high-risk?
It depends on its function. A chatbot that only answers questions about a vacancy is assessed differently from a system that analyses applications, filters candidates or evaluates them, which are listed functions in Annex III point 4. The actual function must be tested.
Does describing a tool as assistive avoid high-risk status?
No. Classification follows the intended purpose and function, not the label. A system that filters or evaluates candidates must be tested against the Annex III listing even if a human makes the final decision.
Must candidates be told AI is used in hiring?
Transparency, data protection and employment law duties can all require information to candidates and workers. The deployer should plan the information, consultation and challenge route before the system is used, and record what was given.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689, as amended ↗AI Act Service Desk implementation timeline ↗