Europe moved its hardest AI deadlines weeks before they applied. The open record shows which industry requests the new text granted, which it refused, and why “Big Tech captured Brussels” is too simple a story.
Mercury Brief · As of 18 June 2026
Status. Parliament approved the compromise on 16 June 2026, but it is not yet law — Council adoption, publication in the Official Journal and entry into force are still outstanding. Until then the current AI Act dates stand.
On 16 June 2026, the European Parliament approved a package that rewrites parts of the EU AI Act less than seven weeks before its first major high-risk obligations were due to apply. Supporters call it simplification; critics call it a rollback. Both labels skip the more useful question: who asked for these changes, and what did they actually get?
Start with why a delay was on the table at all. The Act’s rules for high-risk systems — the AI that screens job applicants, scores borrowers or sits inside critical infrastructure — were built to rest on detailed technical standards written by Europe’s standardization bodies. That layer was not ready. The standards were commissioned in 2023 with an April 2025 deadline; by May 2025 only around eight of roughly thirty-five work items had been published, and national governments were also late naming the authorities the law depends on. Europe set itself a deadline its own machinery could not meet and chose to move the date rather than enforce rules whose scaffolding did not yet exist. That part of the story is genuinely about capacity, not lobbying.
Three requests on the public record
A defensible delay was also exactly what many companies wanted, and the clearest evidence sits in published submissions. The Association for Financial Markets in Europe, the banks’ lobby, called delaying the high-risk deadline its “highest priority” and pressed for a fixed December 2027 date — and, separately, for removing the duty to register systems that providers self-assess as not high-risk. The appliedAI Institute for Europe asked to postpone both the high-risk and the transparency timelines, to grandfather older systems, and to extend relief to deployers, not just providers. And a group of industrial manufacturers, with DigitalEurope and companies including Siemens and Bosch behind it, argued that AI built into regulated products should be governed mainly through existing sectoral law — a dispute that became one of the central obstacles in the April negotiations.
What was granted — and what was not
Set those requests against the Parliament-approved text and the result is a mixed ledger, which is the whole point. The askers won on timing and sectoral scope: the high-risk deadline became a fixed 2 December 2027 (and 2 August 2028 for AI inside regulated products), and AI in machinery moves toward the machinery-law route. But they lost on oversight. The registration duty was not deleted but kept in simplified form, reversed back in after civil-society resistance. Most of the remaining transparency duties stayed on the original 2 August 2026 timetable; only Article 50(2)’s machine-readable marking duty for older generative systems received a short grace period. A proposal to loosen the safeguard on using sensitive data for bias testing was refused, the strict standard kept. Lawmakers also added two new prohibitions — covering AI-generated non-consensual intimate imagery and child sexual abuse material — that were not identified in the industry submissions examined for this reporting. Taken together, the documented requests succeeded on timing and sectoral scope but failed on registration, most transparency delays and the proposed bias-data change. That is a reorder, not a surrender.
The line that matters: alignment is not causation
Watchdogs at Corporate Europe Observatory and LobbyControl traced most of the changes to long-standing industry positions and reported that the Commission’s meetings on the file leaned heavily toward business. That is an attributed watchdog finding. It shows reported access and alignment, not authorship. An industry ask that matches the final wording does not prove the industry wrote it, and a meeting is not a clause. The honest verdict the open record supports is narrower and more durable than either headline: a redistributed regime, assembled fast under a deadline the EU itself could not meet, shaped by documented requests that succeeded on timing and failed on oversight. We can show clearly who asked. We cannot, from the public record alone, prove who caused.
And the part of the risk that moved did not disappear. People screened, scored or sorted by high-risk systems now wait up to sixteen months longer for the Act’s protections — while the law that already governs AI harms today, from data protection to anti-discrimination to product safety, does not move just because the AI Act’s dates did.
Read the full investigation below. Who Sought to Reorder the AI Act? sets each documented request against the agreed text request-by-request, grades the strength of every link in the chain, and marks exactly where the evidence stops short of causation — with the full request-to-outcome evidence appendix. Available to Mercury Research subscribers.
sources
European Parliament. (2026, June 16). AI Act: EP approves simplification measures and “nudifier” app ban [Press release]. https://www.europarl.europa.eu/news/en/press-room/20260611IPR45207/ai-act-ep-approves-simplification-measures-and-nudifier-app-ban
Council of the European Union. (2026, May 7). Artificial intelligence: Council and Parliament agree to simplify and streamline rules [Press release]. https://www.consilium.europa.eu/en/press/press-releases/2026/05/07/artificial-intelligence-council-and-parliament-agree-to-simplify-and-streamline-rules/
Council of the European Union. (2026, May 13). Digital Omnibus on AI: Compromise text agreed with the European Parliament (ST 9247/26). https://data.consilium.europa.eu/doc/document/ST-9247-2026-INIT/en/pdf
Association for Financial Markets in Europe. (2026, January). EU Digital Omnibus on AI [Position paper]. https://www.afme.eu/media/hviabfe5/afme-paper-on-digital-omnibus-ai-proposals.pdf
appliedAI Institute for Europe. (2026, March). Input on the AI Omnibus (Klein, T., Niemeyer, D., Deo, A., Fischer, W., & Goll, F.) [Position paper]. https://www.appliedai-institute.de/media/downloads/Omnibus/Position_on_Digital_Omnibus_March_2026.pdf
Corporate Europe Observatory & LobbyControl. (2026, January 14). Article by article, how Big Tech shaped the EU’s roll-back of digital rights. https://corporateeurope.org/en/2026/01/article-article-how-big-tech-shaped-eus-roll-back-digital-rights
Kilian, R., Jäck, L., & Ebel, D. (2025). European AI standards: Technical standardisation and implementation challenges under the EU AI Act. European Journal of Risk Regulation, 16(3), 1038–1062. https://doi.org/10.1017/err.2025.10032
Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union, L 2024/1689. http://data.europa.eu/eli/reg/2024/1689/oj
— Viki / Viki Bakos · The Ethics Engine · FairByDesign
Continue to the full paper
This public brief shows what changed and why the simple explanations do not hold.
The full Mercury Research paper, Who Sought to Reorder the AI Act?, follows each documented request through the legislative record, compares it with the Parliament-approved compromise, grades the strength of the evidence, and shows exactly where access and alignment stop short of proven causation.
Already a Mercury Research subscriber?
[Read the full paper here → ]
Not yet subscribed?
Join Mercury Research to access the full paper, its request-to-outcome evidence appendix, and the complete FairByDesign research library.
[Subscribe to Mercury Research →]
Leave a Reply
You must be logged in to post a comment.