EU issues draft high-risk AI guidance

The European Commission has started to put detail around one of the EU AI Act’s most contested questions: which systems actually count as “high-risk.” On May 19, the Commission opened a targeted consultation on draft guidelines meant to help providers, deployers and other actors decide whether an AI system falls into that category. The move follows a separate May 8 publication of draft guidance on transparency duties for certain AI systems. Together, the documents give companies more formal direction as the bloc continues to revise the law’s rollout. ### Which guidance did Brussels publish, and when? The European Commission published draft transparency guidelines on May 8 and opened a consultation on draft high-risk classification guidelines on May 19, according to documents on its digital policy site. The transparency text covers Article 50 of the AI Act and is intended to help competent authorities, providers and deployers comply with disclosure obligations for certain AI systems. (digital-strategy.ec.europa.eu) The May 19 consultation says the high-risk draft is designed to clarify Article 6 of the AI Act and includes practical examples across different sectors and use cases. The Commission said the exercise is meant to collect feedback on both the clarity of the guidance and the usefulness of its examples. (digital-strategy.ec.europa.eu) ### Why does the “high-risk” label matter so much? Article 6 of the AI Act determines whether an AI system falls into the category that triggers the law’s strictest compliance duties. The Commission’s AI Act overview says the regulation sorts systems into categories that include prohibited practices, high-risk systems and systems subject to transparency obligations. (digital-strategy.ec.europa.eu) Providers and deployers have been waiting for that line-drawing because the obligations for high-risk systems are more extensive than the lighter disclosure rules that apply elsewhere in the law. The Commission’s consultation notice does not change the law itself, but it gives companies a draft reading of how Brussels expects the classification test to work in practice. That is an inference from the consultation’s stated purpose and the structure of the Act. (digital-strategy.ec.europa.eu) ### What do the transparency rules cover? The Commission’s May 8 draft says the transparency guidance is meant to support “consistent, effective and uniform” compliance with Article 50. That article covers obligations for certain AI systems, including disclosure-related requirements rather than the full compliance regime reserved for high-risk applications. (digital-strategy.ec.europa.eu) The consultation attached to that draft runs until June 3, according to legal analyses that cite the Commission text. Those analyses describe the guidance as practical help for businesses trying to work out how Article 50 applies before the underlying obligations take effect. (digital-strategy.ec.europa.eu) ### What changed in the May 7 political deal? EU member states and European Parliament negotiators reached a provisional agreement on May 7 on targeted amendments to simplify and streamline parts of the AI Act, the Council of the EU said. The Council said the proposal is part of the “Omnibus VII” package and updated its release on May 18 with a letter sent by the Council presidency to Parliament for a first-reading agreement. (globalpolicywatch.com) That provisional deal would push back some key compliance dates if it is formally adopted. Reporting and legal summaries published after the agreement say the obligations for standalone high-risk systems in Annex III would move to December 2, 2027, while some other deadlines would move to 2028. ### Where do “nudifier” apps fit into this rewrite? (consilium.europa.eu) The European Parliament said in a May press release that the revised text includes a ban on so-called nudifier apps. Parliament’s release says the law also delays the application of watermarking obligations on AI-generated content until December 2, 2026. (computerworld.com) The same release says the ban would apply from December 2, 2026, and would cover AI systems used to create or alter images, audio or video to generate false intimate content without a real person’s consent. That provision sits alongside the broader timetable changes rather than replacing them. ### What happens next for companies and regulators? (europarl.europa.eu) June 3 is the deadline for feedback on the draft transparency guidelines, and June 23 is the deadline for feedback on the draft high-risk classification guidance, according to the Commission consultation pages. Those consultations are the immediate channel for companies, lawyers and trade groups to try to shape the final texts. (europarl.europa.eu) The Council of the EU said the May 7 provisional deal still needs formal adoption by the Parliament and the Council. Until that happens, companies are working from draft guidance on one side and a not-yet-final revised compliance calendar on the other. (consilium.europa.eu) (digital-strategy.ec.europa.eu)