BRUSSELS — The European Council and the European Parliament reached political agreement on Wednesday on a package of amendments to the European Union’s Artificial Intelligence Act, in legislation that the institutions have collectively described as the “A.I. omnibus” on simplification.
The package, the first significant adjustment to the Act since its provisions on general-purpose artificial-intelligence models took effect last August, makes three substantive changes. It extends the list of prohibited A.I. practices to include so-called “nudifier” applications — systems that generate potentially harmful intimate content from non-intimate inputs — with the prohibition to take effect on December 2, 2026. It postpones the deadline by which member states must establish national A.I. regulatory sandboxes to August 2, 2027, in order, the Council said, to give national competent authorities additional implementation time. And it shortens, from six months to three, the grace period that A.I. providers receive to comply with the Act’s transparency obligations on artificially generated content, with the new deadline now set for December 2, 2026.
The Council’s release framed the package as an effort to simplify the Act’s administrative burden on small and medium-sized enterprises while sharpening the most consumer-facing of its obligations. The accompanying language was conciliatory in tone — a contrast to the more adversarial framing the omnibus had received in earlier-stage commentary from European industry associations, which had argued that the Act risked making European A.I. developers uncompetitive against their American and Chinese counterparts.
The European Commission separately opened two consultations during the month. On May 8, it sought stakeholder input on draft guidelines for A.I. transparency obligations. On May 19, it sought feedback on the draft guidelines for classifying high-risk A.I. systems.
The omnibus does not amend the core obligations applicable to providers of general-purpose A.I. models, which entered into force on August 2, 2025, and continue to apply. The penalties applicable specifically to those providers remain postponed until August 2, 2026.
For the international competitive picture, the omnibus is a signal that Brussels intends to keep the Act as the world’s most comprehensive A.I. legislation while making targeted adjustments to keep it implementable. American observers will note that the Brussels agreement arrived in the same week that the Trump administration postponed an executive order in Washington that would have created a federal pre-release review process for frontier models — a contrast that suggests the two sides of the Atlantic are now moving in opposite directions on the regulatory question.
The omnibus must still be formally adopted by both institutions, a step the Council described as a technical follow-up.