EU AI Act Deferral Left Article 50 Live: What Landed Aug 2

Most recruiters read one headline in June ("EU AI Act delayed to 2027") and closed the tab. That was a mistake. The Digital Omnibus deferred the heavy Annex III paperwork, but Article 50 candidate disclosure duties and the February 2025 workplace prohibitions on emotion recognition are fully live as of last week, and they hit the exact tools sitting in your sourcing stack right now.

What actually took effect on August 2, 2026

The Article 50 transparency obligations for deployers of AI systems became enforceable on August 2, 2026, on the original schedule. Only the high-risk Annex III obligations for recruitment tools moved, and only for the vendors building them.

Here is the split that matters for anyone hiring into or from the EU:

• Article 50(1) chatbot disclosure: live August 2, 2026. Any AI system interacting with a natural person must tell that person, at first contact, that they are talking to an AI.
• Article 50 deployer transparency (deepfakes, synthetic media, AI-generated text): live August 2, 2026. Visible or audible labels, not metadata watermarks nobody sees.
• Article 50(2) provider-side watermarking: postponed to December 2, 2026. Shorter than the six-month delay originally floated.
• Annex III high-risk hiring obligations (standalone systems): deferred to December 2, 2027.
• Annex I high-risk AI embedded in regulated products: deferred to August 2, 2028.
• Article 5 prohibitions (emotion recognition, biometric categorization in workplace and recruitment): unchanged, live since February 2, 2025.
• Article 4 AI literacy duty: unchanged, live since February 2, 2025.

The European Parliament endorsed the Omnibus on June 16, 2026 by 423 to 57 with 174 abstentions. The Council of the EU gave final approval on June 29, 2026. The deferral is real. It just does not cover what most recruiters assumed it covered.

Recruiters are deployers, not providers

Article 50 draws a sharp line. Providers build the model. Deployers put it in front of a candidate. Almost every transparency duty that landed on August 2 falls on the deployer, meaning the employer or agency running the outbound agent, not the vendor selling it. The one duty that got postponed (Article 50(2) watermarking of synthetic output at generation time) is a provider duty. Recruiters do not get the extension.

The tools in your stack that now require disclosure or removal

Three concrete tools in a typical 2026 sourcing stack now require disclosure or removal: AI outreach agents, AI interview chatbots, and any product that scores candidates on emotion, voice tone, or "cultural fit" from video.

Here is the checklist, in the order most funnels break:

1. AI outreach agents: If a candidate replies to your sequence and an LLM writes the reply, disclose it in that reply, not on the careers page. First contact means first message.
2. Screening chatbots: The intake bot on your careers site must say it is an AI in the interaction itself. A privacy footer does not count.
3. AI-generated job descriptions and outreach copy sent to EU candidates: These count as AI-generated text under Article 50 and require deployer-side transparency where the content is not obviously AI-authored.
4. Interview scoring tools that read facial expression, voice tone, or body language: These are not a disclosure problem. They are prohibited under Article 5. Turn them off.
5. Boolean-only sourcing: Not regulated, but the more you lean on plain-English search that never touches a candidate, the smaller your Article 50 surface area. This is a genuine reason the market is shifting toward tools like Refolk, where the AI ranks profiles from public data rather than sending messages that trigger disclosure duties.

€35M

First contact means the interaction, not the ToS

The Commission's draft Article 50 guidelines are explicit: a statement buried in terms and conditions, a metadata watermark on its own, or a vague reference to an "assistant" does not satisfy the chatbot disclosure duty. The information has to be perceivable in the interaction itself. That kills the common 2025 practice of hiding "AI-assisted" in a footer.

The emotion-AI ban has been live for 18 months and vendors are still selling banned features

AI systems that infer emotions of employees or candidates in workplace contexts have been prohibited since February 2, 2025, except for medical or safety purposes. The European Commission's guidelines apply a broad interpretation of "workplace" that explicitly covers recruitment, meaning job applicants are protected before they are ever hired.

The features that are now illegal in EU hiring:

• Facial expression analysis during video interviews
• Voice tone or prosody scoring for "enthusiasm" or "confidence"
• Body language scoring for "cultural fit"
• Any biometric categorization inferring emotional state from a candidate

Companies including HireVue, Cognisess, Emotiv, and Cogito have built or claimed to build exactly these capabilities. The Dutch Data Protection Authority (Autoriteit Persoonsgegevens), through its Department for the Coordination of Algorithmic Oversight, published a February 2025 summary (DCA-2025-02) confirming workplace and education emotion recognition as a priority. Fines under Article 5 reach €35 million or 7% of global annual turnover, whichever is higher, and they land on the deploying employer.

The vendor built the feature. The employer pressed the button. The 7% penalty follows the button.

Who has to change their outbound this week

Roughly 1,790 technical recruiters, sourcers, talent acquisition specialists, and heads of talent across the nine largest EU economies need to update disclosure language in their outbound and interview flows this week. That count comes from Refolk's index of professional profiles across DE, FR, NL, IE, ES, IT, BE, SE, and PL, filtered to those titles.

The concentration matters more than the total:

Two things jump out. First, the EU-9 recruiter base is roughly 119 times the size of the AI-sourcing-specialist pool in the US and UK combined. Most EU recruiters are generalists, not AI specialists, so they need clear disclosure guidance, not a research paper. Second, there are 1.54 recruiters for every AI governance professional in the EU-9. The people who need to change wording outnumber the people writing the policy. Expect uneven rollout for months.

1.54x

Agencies cascade fastest

K2 Partnering Solutions is the single largest employer of technical recruiters in the EU-9 sample. Staffing agencies punch above their weight here: one agency's chatbot disclosure decision ripples across dozens of client hiring funnels within a week. If you run an in-house talent team and you use an agency, ask them today what their outreach agent says on first contact.

Berlin, Madrid, and Dublin lead the EU recruiter concentration. Those three cities are also where you will find the highest density of English-language AI startups running US-style outbound sequences against EU candidates, which is exactly the failure mode Article 50 was written for.

Non-EU employers are covered the moment a candidate is in the EU

The obligation applies to any provider whose product reaches EU users, not just EU-headquartered companies. A US-based recruiter using an AI outreach agent that messages a Berlin engineer is inside the scope of Article 50.

This is where the deferral becomes a trap. By pushing the heavy Annex III audit paperwork to December 2027, the Omnibus removed the forcing function that was going to make employers review their hiring AI this summer. Meanwhile the February 2025 prohibitions and the August 2026 disclosures are still enforceable. Employers now have 16 more months of exposure with less scheduled internal review.

Fines under Article 50 run up to €15 million or 3% of global annual turnover, whichever is greater. Fines under Article 5 for the emotion-AI ban run to €35 million or 7%. Both apply extraterritorially where the output reaches an EU user.

The label question

The Commission has floated a standardised EU label for AI-generated content, currently proposed as a visual "AI" tag (localised as "KI" in German, "IA" in French). A Code of Practice on AI-generated content is under development. Do not wait for the final artwork. A plain sentence at the top of an outreach email ("This message was drafted with AI assistance") satisfies the deployer duty today and is easy to update to whatever glyph the Code lands on.

What to do this week

Do these five things before Friday. None of them require legal sign-off. All of them shrink your exposure.

1. Audit every candidate-facing surface for AI touchpoints: careers-site chatbots, screening bots, scheduling agents, outbound sequencers, interview transcription and scoring.
2. Add a first-contact disclosure line to every AI-authored outbound message and every chatbot opener. One sentence. Not a footer.
3. Kill emotion, voice tone, and body language scoring in any video interview tool. Confirm with the vendor in writing that these features are disabled for EU candidates.
4. Shift more of your sourcing to read-only tools that rank public profiles rather than message candidates. Plain-English search across public professional data (the surface Refolk covers) has zero Article 50 disclosure surface because no candidate ever interacts with an AI in the search step.
5. Document your AI literacy program. Article 4 has been live since February 2025 and is often forgotten. A one-page internal doc on how your team uses AI in hiring, refreshed quarterly, is the minimum defensible artifact.

The through-line: the more of your sourcing pipeline is AI-ranking against public data and the less of it is AI-messaging into candidate inboxes, the smaller your compliance surface. That is what the text of Article 50 actually regulates. Treat "does this tool talk to candidates" as the first filter when you rebuild your stack.

FAQ

Did the EU AI Act get delayed for hiring?

Partially. The Digital Omnibus, approved by the Council on June 29, 2026, deferred the Annex III high-risk obligations for standalone recruitment AI from August 2, 2026 to December 2, 2027, and the Annex I obligations for AI embedded in regulated products to August 2, 2028. It did not defer Article 50 transparency duties for deployers, the Article 4 AI literacy duty, or the Article 5 prohibitions on workplace emotion recognition and biometric categorization. Recruiters are deployers, so most of what applies to them is live now.

What has to be disclosed to candidates under Article 50?

Any AI system interacting with a candidate must disclose that it is an AI at the first point of contact, perceivably, in the interaction itself. That covers screening chatbots, AI outreach agents that reply on the recruiter's behalf, and synthetic media. A statement in the terms of service, a metadata watermark, or a vague word like "assistant" does not satisfy the duty. Provider-side watermarking under Article 50(2) is postponed to December 2, 2026, but deployer disclosure is not.

Is HireVue-style video scoring still allowed in the EU?

Not for emotion, tone, or cultural fit inference. Since February 2, 2025, AI systems that infer emotions of candidates in the recruitment process have been prohibited under Article 5, except for narrow medical or safety purposes. Fines reach €35 million or 7% of global annual turnover. Structured video interviews without emotion inference remain allowed. The specific features that are banned are facial expression analysis, voice tone analysis, and body language scoring aimed at emotional or personality traits.

Do US employers hiring EU candidates have to comply?

Yes, when the AI system's output reaches a person in the EU. A US-based recruiter running an AI outreach agent that messages a Berlin engineer is within scope of Article 50, and a US employer using banned emotion-AI on an EU applicant is within scope of Article 5. The extraterritorial reach was one of the design goals of the AI Act and was not touched by the Omnibus. Practically, if any part of your funnel touches EU candidates, apply the disclosure line globally rather than trying to route around it.