EU AI Act Article 50: What the Transparency Rules Mean for Recruitment

On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act began to apply. The European Commission’s new Guidelines explain how they should work in practice, alongside a Code of Practice covering AI-generated content.

For recruitment and talent acquisition teams, the timing matters. Candidate-facing chatbots, AI sourcing agents, automated interview tools and synthetic employer-brand content are already in use. Some fall directly within Article 50. Others do not.

Article 50 is not a general rule requiring employers to announce every use of AI in hiring. It contains four separate obligations, aimed at different systems and assigned to different actors. A blanket “we use AI” notice will often be too vague. In other cases, it may address the wrong legal requirement.

Key takeaways for recruitment teams

  1. Candidate-facing AI chatbots, voice assistants and conversational agents will generally require an in-context disclosure that the candidate is interacting with AI.
  2. AI agents contacting candidates must disclose their artificial nature and the person or organisation on whose behalf they are acting.
  3. A back-end CV screening or ranking system is not automatically covered by the interactive-AI rule, although separate high-risk AI and data protection duties may apply.
  4. Emotion recognition in recruitment should be treated first as a legality question. Disclosure does not make a prohibited or otherwise unlawful use lawful.
  5. Realistic synthetic recruiter, employee or executive videos may qualify as deep fakes and require a clear label at first exposure.
  6. Ordinary AI-assisted job adverts and candidate emails are not automatically caught by the public-interest text rule. Purpose, audience and human review matter.

Article 50 is four different transparency rules

Article 50(1) applies to providers of AI systems that interact directly with people. Article 50(2) requires providers of systems generating or manipulating synthetic content to make outputs machine-readable and detectable as AI-generated or manipulated. Article 50(3) applies to deployers using emotion recognition or biometric categorisation. Article 50(4) applies to deployers using AI to create deep fakes or certain public-interest text.

The provider–deployer split is not just vendor terminology. A provider develops an AI system, has it developed, or places it on the EU market under its own name or trademark. A deployer decides to use a system under its authority and determines its purpose and manner of use. An employer may become a provider where it develops an in-house system or materially modifies an existing one and puts it into service under its own name.

Responsibility cannot be settled by saying, “the vendor handles compliance”. The vendor may carry the design obligation for a candidate chatbot, while the employer still needs to verify that the disclosure appears in the configured candidate journey. The employer may separately carry deployer duties for synthetic media or biometric tools.

Candidate-facing chatbots and voice assistants

Article 50(1) is the most obvious point of contact with recruitment.

It applies where an AI system is intended to interact directly with a person through a genuine conversational or responsive exchange. The interaction can be written, spoken, visual or physical, and may be a single exchange or a longer conversation.

A generative career-site chatbot answering questions about vacancies is likely to be in scope. So is a voice assistant conducting an initial screening call, an interview-scheduling bot that responds contextually, or an AI avatar guiding someone through an application.

A simple rule-based tool is different. The Guidelines distinguish AI systems from traditional automated responses such as out-of-office emails or fixed, rule-based quick replies. They also distinguish direct interaction from back-end assistance. Where a recruiter uses AI to draft a response, reviews it and sends it as the human interlocutor, that does not necessarily become direct interaction between the candidate and the AI system.

Where Article 50(1) applies, the candidate should be explicitly informed about the artificial, non-human nature of the interaction no later than the first interaction.

For a recruitment journey, the wording could be:

You are interacting with an AI recruitment assistant provided by [Vendor] on behalf of [Employer].

For voice:

This call is being handled by an AI-powered recruitment assistant acting on behalf of [Employer].

Terms and conditions are not enough. Nor is a generic website statement saying that the company uses AI. The Guidelines also warn against ambiguous labels such as “assistant”, technical descriptions such as “powered by an LLM”, or machine-readable markings that candidates cannot perceive.

A single prominent notice at the start will often be sufficient. Longer or changing interactions may justify a persistent badge or reminder, particularly where a voice or avatar becomes increasingly human-like.

AI sourcing and scheduling agents need to identify their principal

The section on AI agents deserves particular attention from recruitment businesses.

The Guidelines state that an AI agent interacting with people should disclose its artificial nature and the person on whose behalf it is acting. That maps directly onto sourcing agents that identify prospects, send outreach, answer questions, negotiate interview availability or manage correspondence.

An outbound message that uses a recruiter’s name and profile while hiding that the exchange is being conducted by AI is difficult to reconcile with that approach.

Recruitment teams should establish whether the system itself is communicating with the prospect and whether the recipient can understand both that it is AI and who authorised it to act. This also matters in agency and RPO arrangements, where the agent may be acting for the staffing company, the end client, or each at different stages.

A CV ranking system is not a chatbot

One of the easiest mistakes is to use Article 50 as the legal basis for every candidate notice about automated hiring.

The Guidelines place back-end decision-support systems outside the Article 50(1) direct-interaction rule where the person only sees the output and cannot interact with the AI system. A CV matching or ranking tool behind an applicant tracking system may therefore fall outside Article 50(1), even though it may be highly consequential.

That does not mean no transparency duty exists. Deployers of high-risk AI systems may have a separate obligation to inform people when a high-risk system makes or assists with decisions about them. The notices are complementary and can be combined, but they concern different facts.

A chatbot notice tells the candidate what they are talking to. A decision-system notice tells them that AI is being used in a process affecting them. One should not be used as a proxy for the other.

Article 50 compliance also does not answer questions about discrimination, bias testing, human oversight, data protection or the wider high-risk AI regime. A well-designed banner cannot compensate for a poorly governed selection model. The Commission itself makes clear that satisfying Article 50 does not make an otherwise prohibited system lawful or replace obligations arising elsewhere in the AI Act.

Emotion recognition: disclosure is not permission

Article 50(3) requires deployers of emotion recognition and biometric categorisation systems to inform every person exposed to the system, whether analysis happens in real time or after a recording is collected.

An emotion recognition system identifies or infers emotions or intentions from biometric data. In recruitment, that may include tools claiming to infer confidence, enthusiasm, honesty or engagement from facial movement, voice, posture or other biometric signals.

The legal issue does not begin with the wording of the notice. The Commission is explicit that Article 50 compliance does not make the underlying use lawful. It points to emotion recognition in workplace and education settings as an example of a use that may be prohibited under Article 5.

For a video-interview product making emotion or intention inferences, the sensible sequence is to establish what it infers and from which data, assess whether the use is prohibited or restricted, and only then consider the Article 50(3) notice and separate data protection information.

Product labels are not determinative. Calling a feature “behavioural insight” or “candidate engagement analysis” does not resolve the issue if its function is to infer emotion or intention from biometric data.

Biometric categorisation needs similar care. The Guidelines give age or gender classification from biometric data as examples. Even where a use is not prohibited, the information duty sits alongside data protection and non-discrimination requirements.

Synthetic recruitment videos and employer-brand content

Article 50(4) reaches beyond political deep fakes.

A deep fake is AI-generated or manipulated image, audio or video content that resembles an existing person, object, place, entity or event and would falsely appear authentic or truthful. The Commission’s examples include a realistic synthetic avatar of a company CEO and an AI-generated celebrity in promotional content.

That makes the rule relevant to employer branding. A cloned video of a chief people officer introducing a graduate programme, a synthetic recruiter giving application advice, an AI-generated employee testimonial or cloned voice used in a recruitment advert could fall within the definition. Realism, the subject represented, the audience and whether the content could be taken as authentic all matter.

Where the content qualifies, the deployer must disclose that it was artificially generated or manipulated. The Code of Practice suggests an “AI” icon or equivalent label positioned so it is immediately visible. For video, it recommends disclosure at the beginning and, where appropriate, at intervals or after interruptions. Audio-only content should begin with a short, plain-language disclaimer.

A footer on a careers page, an obscure caption or a label that disappears before the viewer can process it is unlikely to meet the stated standard. Accessibility must also be considered, including contrast, screen-reader compatibility and alternatives for audio or visual cues.

What about AI-written job adverts and candidate emails?

The public-interest text limb of Article 50(4) is narrower than many organizations may assume.

It applies where AI-generated or manipulated text is published to inform the public on matters of public interest. The Guidelines say that private professional correspondence and internal company communications are not “published”. They also indicate that ordinary company advertising or product descriptions generally fall outside the rule unless they contain claims relating to matters such as health, consumer safety or sustainability.

Most candidate emails will therefore sit outside this part of Article 50. A standard job advert will also usually look more like advertising than a public-interest publication. That is a reasonable reading of the Guidelines, not a universal exemption. A corporate report about workforce restructuring, pay practices, public-sector recruitment or discrimination could present a different case.

Where the rule is engaged, there is an exception if the text has undergone substantive human review or editorial control and a person or organisation holds editorial responsibility.

The review cannot be cosmetic. The Commission says fact-checking is a minimum element. Spell-checking, a quick grammatical pass, automated review by another AI system or cursory approval will not be enough. Substantive AI intervention after sign-off can also remove the benefit of the exception.

Even when Article 50(4) is not triggered, recruitment content should still be reviewed for accuracy, discriminatory wording, misleading claims, salary information, eligibility criteria and consistency with the real hiring process.

A practical Article 50 map for recruitment

Recruitment use caseLikely Article 50 issuePractical responseGenerative careers chatbotArticle 50(1): direct AI interactionClear first-turn disclosure; verify it in the live journeyAI sourcing agent emailing prospectsArticle 50(1): agent identity and principalState that it is AI and who it acts forBack-end CV screening or rankingUsually not Article 50(1) because there is no direct interactionAssess separate high-risk, data protection and employment-law dutiesVideo tool inferring emotion from face or voiceArticle 5 legality question and Article 50(3) transparencyAssess legality before designing a noticeSynthetic executive or recruiter videoArticle 50(4) if it qualifies as a deep fakeApply a visible, accessible label at first exposureAI-assisted job advert reviewed by a recruiterUsually outside the public-interest text limbKeep meaningful human review and evidence of approval

“Likely” matters. Classification depends on the system, configuration, audience and use.

Which Article 50 deadline applies?

Article 50 generally applies from 2 August 2026, but the Commission Guidelines identify a limited transition for certain existing generative AI systems.

Providers of generative AI systems placed on the market or put into service before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking and detection requirements under Article 50(2). This is a targeted grandfathering provision. It is not a four-month delay to Article 50 as a whole.

For recruitment teams, the practical distinction is important. A candidate-facing chatbot that also generates text may benefit from the December transition in relation to the provider’s technical marking of its outputs. The requirement to inform candidates that they are interacting with AI still applies from 2 August.

The same is true of other deployer-facing obligations. The December date does not postpone disclosures concerning emotion recognition, biometric categorisation, deep fakes or qualifying public-interest text.

Content generated before 2 August does not generally have to be marked or labelled retrospectively. However, qualifying public-interest text generated before that date but first published on or after 2 August must be labelled.

What talent acquisition leaders should do now

Build an Article 50 inventory. List every AI-enabled candidate touchpoint and process producing synthetic text, audio, images or video. Record the owner, vendor, intended audience, EU exposure, provider/deployer role and possible trigger.

Test the live interface. Capture screenshots, recordings and transcripts showing what a candidate sees or hears at first interaction. Check mobile views, voice journeys, embedded tools and redirects. A disclosure in a policy but missing from the configured interface is not an effective control.

Separate interaction transparency from decision transparency. Document which notice explains an AI conversation and which explains AI-assisted assessment or decision-making. Combine them where useful, but do not collapse them into vague wording.

Re-open vendor due diligence. Ask suppliers to explain their Article 50 role, disclosure mechanism, accessibility testing, AI-agent identification, machine-readable marking, change-control procedures and customer evidence. Confirm whether customer modifications could shift provider responsibilities.

Review synthetic media workflows. Employer-brand and marketing teams need a route for identifying deep-fake content before publication, applying a label, preserving it across channels and correcting missing or incorrect disclosures.

Ring-fence emotion and biometric inference. Where a tool claims to evaluate emotional state, intention, honesty, confidence or similar traits from biometric data, pause deployment or restrict the feature until its function and legal status have been assessed and documented.

Keep an evidence pack. Retain the use-case assessment, role allocation, disclosure wording, interface evidence, accessibility checks, approval records, staff guidance, exception rationale and remediation log. The Code emphasises internal processes, representative examples, personnel awareness and correction mechanisms.

The real standard is operational

The Commission Guidelines are non-binding, and authoritative interpretation remains with the Court of Justice of the European Union. The Code of Practice offers a recognised route for demonstrating compliance with the AI-generated content rules, but it does not replace the Act.

Infringements of Article 50 can attract fines of up to €15 million or 3% of worldwide annual turnover, subject to the Act’s rules for smaller organizations and the circumstances of the case. That should not be the only reason to act.

For recruitment, the immediate issue is whether candidates understand when AI is speaking to them, acting on someone else’s behalf, analysing biometric information or presenting synthetic content as if it were real.

The right response is neither to label every algorithm nor to leave transparency entirely to the supplier. Map the system. Identify the legal role. Place the notice inside the real candidate experience. Test it. Keep the evidence.

Not necessarily under the interactive-system rule.

A back-end screening, matching or ranking system will generally not fall within Article 50(1) where the candidate does not interact directly with the system. The Commission specifically distinguishes direct interaction from back-end decision-support tools where a person only encounters the resulting output.

That does not remove other obligations. Recruitment and selection systems may be classified as high-risk under the EU AI Act, and separate transparency, governance, human oversight, data protection and non-discrimination requirements may apply.

No. Article 50 does not create a single, general obligation to announce every use of AI in a hiring process.

Its requirements depend on the type of system, what it does and whether the organisation is acting as a provider or deployer. A candidate-facing AI chatbot may trigger the interactive-system transparency rule, while a back-end screening tool may fall outside that specific obligation but still be subject to separate requirements for high-risk AI systems, data protection and employment law.

Usually, yes.

Where a chatbot or conversational agent uses AI to interact directly with a candidate, the provider must design the system so that the candidate is informed that they are interacting with AI. The disclosure should appear no later than the beginning of the first interaction and should be clear enough for the candidate to understand the artificial nature of the exchange.

A label such as “virtual assistant” may not be sufficient on its own. A clearer notice would say:

You are interacting with an AI recruitment assistant acting on behalf of [Employer].

No. A disclosure hidden in terms and conditions, a privacy notice, product documentation or several layers of an online menu is unlikely to satisfy the Article 50 standard by itself.

The information must be noticeable, easy to understand and distinguishable from surrounding information. For an interactive recruitment tool, it should normally appear within or immediately beside the candidate interaction.

Privacy information may still be required, but it serves a different purpose.

Both dates appear in the Commission Guidelines, but they apply to different circumstances.

2 August 2026 is the general application date for Article 50. Candidate-facing AI interaction disclosures and the relevant deployer obligations therefore apply from that date.

2 December 2026 is a limited transition date for providers of generative AI systems placed on the market or put into service before 2 August. It gives those providers additional time to implement the machine-readable marking and detection requirements under Article 50(2).

The December transition does not postpone the obligation to tell a candidate that they are interacting with an AI system. A system that is both interactive and generative may therefore have two timelines: candidate-facing disclosure

Article 50 began to apply on 2 August 2026.

The obligation applies to in-scope systems operating in the EU on that date, including systems that were placed on the market or put into service earlier. Existing recruitment technology is not automatically exempt because it was purchased or deployed before August 2026.

It depends on how the workflow operates.

Where AI prepares a draft and a recruiter substantively reviews, approves and sends it as the main human interlocutor, the candidate is not necessarily interacting directly with the AI system. The Guidelines distinguish this from situations where the system itself sends contextual responses to the candidate.

The practical question is not simply whether AI assisted with the message. It is whether the candidate is communicating with the AI system or with a human who has used AI as an internal drafting tool.

Yes, where the agent is directly interacting with a person.

The Commission’s Guidelines state that AI agents should disclose both their artificial nature and the person on whose behalf they are acting. In recruitment, this could apply to agents that send sourcing messages, answer candidate questions, arrange interviews or manage correspondence.

The message should make it clear whether the agent is acting for an employer, a recruitment agency, an RPO provider or another organization.

Only where the review is substantive and someone holds editorial responsibility for the publication.

The Commission says that human review should involve deliberate examination by someone with relevant knowledge and professional judgement. Fact-checking is a minimum element. Spell-checking, a superficial grammatical review or approval by another AI system will not satisfy the exception.

The organiztion should also be able to identify the person or legal entity holding ultimate editorial responsibility.

They may do.

A realistic synthetic video, image or audio recording that resembles an existing or plausible person, object, place, entity or event and could falsely appear authentic may qualify as a deep fake. The Commission gives a realistic synthetic company CEO as an example of content that may fall within the definition.

This could include a cloned executive welcome video, synthetic recruiter, AI-generated employee testimonial or voice clone used in employer-brand advertising.

Where the content qualifies, the disclosure should be visible or audible at first exposure and should remain perceivable under normal viewing or listening conditions.

Disclosure alone does not make the use lawful.

Article 50(3) requires people exposed to emotion recognition or biometric categorisation systems to be informed about the system’s operation. The Commission also makes clear that complying with this information duty does not legitimise a use that is prohibited or unlawful under another part of the AI Act or under data protection and discrimination law.

Recruitment teams should first establish whether the tool is inferring emotions or intentions from biometric data and whether the proposed workplace use is legally permitted. The notice comes after that assessment, not instead of it.

The responsibility may be shared, but it is not necessarily identical.

A vendor that develops and markets a candidate-facing AI system will commonly be the provider responsible for designing the relevant transparency mechanism. The employer or recruitment business will often be the deployer deciding why and how the system is used.

The employer should therefore verify that the disclosure works in its configured candidate journey rather than relying only on contractual assurances. An organization may also become a provider if it develops an in-house system or substantially modifies an existing system and puts it into service under its own name.

Usually not under the public-interest text provision, although the facts matter.

Article 50(4) applies to AI-generated or manipulated text published to inform the public on matters of public interest. The Guidelines indicate that ordinary company advertising and product descriptions generally fall outside this category, subject to the nature of the claims being made.

A conventional job advert is therefore unlikely to require an Article 50 label solely because AI helped to write it. It should still undergo meaningful human review for accuracy, discriminatory language, misleading claims and consistency with the actual role.