The European Union's Artificial Intelligence Act is the world's first comprehensive law governing AI. The European Parliament adopted it in plenary on 13 March 2024 by 523 votes to 46, with 49 abstentions — 618 members voting in all. ² It was published in the Official Journal on 12 July 2024 as Regulation (EU) 2024/1689 and entered into force on 1 August 2024. ⁴

That original law set out a staggered timeline: different obligations switch on at different dates over several years. As those dates approached, businesses and some governments said the rules were arriving faster than anyone could comply. The response was a second piece of legislation — an "Omnibus" amendment (sometimes called the simplification or Digital Omnibus package) designed to adjust the timeline without rewriting the core protections.

That amendment is the subject of this report. It moved through the EU's standard three-body process: the European Commission proposes, and the Parliament and the Council of the EU (representing the 27 national governments) must both agree. The Council fixed its negotiating position on 13 March 2026. ⁴ After a failed negotiation and then a provisional deal, the Parliament endorsed the result on 16 June 2026, the Council approved it on 29 June 2026, and it entered into force on 27 July 2026. ⁴ What follows examines what was decided, who backed it, who resisted, and where the gaps lie.

The decisive moment came in Strasbourg on 16 June 2026, when the European Parliament voted to formally amend the AI Act. The law firm Ogletree Deakins reports the tally as 423 in favour to 57 against. ¹ The UCL researcher Oliver Patel reports the same 423 for and 57 against, and adds a figure the law firm omits: 174 abstentions. ¹ The two independent origins agree exactly on the in-favour and against counts; the abstention figure rests on Patel alone. ² That single-source gap is why the headline tally, as first reported without abstentions, is treated cautiously — but no source contradicts the 423-to-57 result.

That vote did not come out of nowhere. A first attempt to strike a deal collapsed: the second political trilogue — the closed-door negotiation between Parliament, Council and Commission — broke up on 28 April 2026 after eleven hours without agreement. ³ A provisional deal was reached nine days later, on 7 May 2026. ³ The Parliament's own preparatory committees had already cleared the ground: a joint vote of the internal market and civil liberties committees on 18 March 2026 recorded 101 in favour, 9 against and 8 abstentions, a figure two independent sources report identically. ² The full Parliament had adopted its negotiating position on 26 March 2026, though the specific tally there — reported as 569 to 45 with 23 abstentions — comes from a single low-credibility outlet. ¹

After the Parliament's June endorsement, the Council of the EU gave the package its final green light on 29 June 2026. ² The Commission then confirmed the AI Omnibus Regulation entered into force on 27 July 2026, published in the Official Journal as OJ:L_202601744. ¹ Every one of these procedural milestones — the collapsed trilogue, the provisional deal, the Council approval, the entry into force — is confirmed by multiple independent and official sources. ⁴ The legislative machinery, in short, functioned and is well documented.

Open: Whether an official European Parliament roll-call record confirms the 174 abstentions reported for the 16 June 2026 vote.; Whether the 26 March 2026 position-vote tally of 569-45-23 is corroborated by any official parliamentary record.

The core bargain of the Omnibus was time. The single most consequential change pushed back the compliance deadline for the highest-risk AI systems. For so-called Annex III systems — AI used in areas such as employment, credit, or essential services — the deadline moved from 2 August 2026 to 2 December 2027. ⁴ That is an extra 16 months, and it is the best-corroborated substantive change in the record, reported by six independent origins including two official EU sources. ⁴ A related category, high-risk AI embedded in physical products (Annex I), was delayed to 2 August 2028, though the precise starting date it was moved from rests on a single source. ²

Crucially, the amendment did not delay everything. The rules aimed at the most acute harms kept a near-term schedule. AI systems that generate non-consensual intimate imagery or child sexual abuse material are outlawed from 2 December 2026 — a date confirmed by five independent origins, including the Commission itself. ⁴ Broader enforcement begins earlier still: the Commission's AI Act Service Desk confirms that transparency rules under Article 50, along with enforcement over general-purpose AI, prohibitions and AI-literacy duties, start on 2 August 2026. ² One narrower obligation — the requirement under Article 50(2) to watermark AI-generated content — was given a short grace period, its enforcement postponed to 2 December 2026, a timing three independent origins agree on. ²

The amendment also confirmed the machinery of enforcement. From August 2026, the Commission's AI Office can conduct evaluations and inspections of large AI models — including, according to the Swiss public broadcaster RTS, before those models reach the market. ¹ That pre-market inspection power is broadly consistent with the Commission's own description of AI Office authority, but the specific "before market placement" framing appears in only that one account. ¹ A reader comparing the delayed and non-delayed items should note they are measured from different points in the original Act's staggered timeline, so the length of each "delay" is not directly comparable without specifying which baseline applies. ⁴

Open: Whether the AI Office's inspection powers explicitly extend to reviewing large models before market placement, beyond the single RTS account.; The exact original baseline date from which the Annex I embedded-product deadline was extended.

The Omnibus passed comfortably, and the balance of the Parliament makes the likely coalition easy to sketch — but the record does not let us confirm it for the vote that matters. For the original 2024 Act, the roll-call is complete: the parliamentary vote tracker HowTheyVote.eu records the centre-right European People's Party voting 153 in favour, 1 against, with 4 abstentions, and the centre-left Socialists and Democrats voting 129 in favour, 0 against, with 1 abstention. ¹ Country delegations broke similarly: Germany's members split 61 for, 5 against and 19 abstaining; France 65 for, 7 against, 1 abstaining; Italy 58 for, none against, 7 abstaining. ¹

For the 16 June 2026 Omnibus vote, no equivalent breakdown exists in the record. We know the aggregate — 423 to 57 — but not how any single group or country split within it. ¹ That is the central evidentiary hole in answering who supported the most recent amendment.

What the record does offer is an indirect signal of centrist dominance. The voting-analytics firm EUmatrix reports that across roll-call votes in the March 2026 plenary, the EPP was on the winning side 94% of the time, the liberal Renew group 90.6%, and the Socialists and Democrats 85.2%. ¹ These are the three groups that form the Parliament's governing centre. But two cautions apply. First, all five of these percentages come from a single analytics provider. ¹ Second, a "winning-side rate" measures how often a group ended up with the majority across many votes in one session — it is not a tally of any single vote, and cannot be compared to the raw counts from the June Omnibus decision. ² It suggests a centrist majority held together in that period; it does not prove how those groups voted on 16 June.

Open: The group-level and member-state breakdown of the 16 June 2026 Omnibus vote.; The stated reasons MEPs gave for supporting the Omnibus during the 16 June 2026 plenary debate.

Opposition to the EU's AI regime came from opposite ends of the chamber, for opposite reasons — and the documentary record for it is noticeably thinner than for the majority. On the original 2024 Act, the left-wing group The Left broke sharply against the grain, voting just 4 in favour, 21 against, with 4 abstentions. ¹ That is the clearest documented opposition in the entire record — but it dates from 2024, not the 2026 Omnibus, and there is no comparable data for how The Left voted in June 2026.

On the right, two sovereigntist groups were the session's biggest losers by EUmatrix's count: the Patriots for Europe on the winning side just 38.3% of the time, and the Europe of Sovereign Nations group 31.4% — both on the losing side of most divisions in March 2026. ¹ Their objection was ideological. At a 20 January 2026 plenary debate on AI-generated deepfakes, the ESN member Christine Anderson said: "We do not need hysteria. We need enforcement and criminal law must be enforced by the member states". ¹ The Patriots' Sebastian Kruis went further, calling the regulation "not safety... not protection. It's power play. It's a debate about who is allowed to participate in discourse and who is not". ¹ Both quotes come from a single unverified YouTube aggregator and would need an official plenary transcript to confirm. ¹

A third line of resistance came from civil society, aimed not at the existence of the rules but at holes in them. The advocacy group ECNL alleges the Act lets companies and public authorities unilaterally decide their high-risk-category AI poses no significant risk, sidestepping oversight; that AI in EU migration databases such as Eurodac, SIS and ETIAS need not comply before 2030; and that the duty to publish fundamental-rights impact summaries exempts law enforcement and migration authorities — the very areas where the group says the stakes are highest. ¹ Each of these three claims rests on ECNL alone, an interested advocacy source, and none has independent legal analysis in the record either confirming or rebutting it. ¹ Separately, industry did not oppose the Act outright but argued it remained too costly even after the extensions — a case examined in the next section. ¹

Open: How The Left, PfE and ESN groups actually voted on the 16 June 2026 Omnibus amendment.; Whether independent legal analysis supports ECNL's characterisation of Article 6(3) self-assessment, the migration-database timeline, or the fundamental-rights-impact publication exemptions as loopholes.

A law is only as strong as its enforcement, and here the record tells two stories at once. On paper, the penalty structure is proportionate and well-established. Article 99 sets three tiers: up to €35 million or 7% of global annual turnover for the most serious prohibited practices, up to €15 million or 3% for high-risk non-compliance, and up to €7.5 million or 1.5% for supplying misleading information. ¹ The top two tiers are confirmed by multiple independent sources consistent with the Act's text. ⁴ Because the cap is "whichever is higher," the percentage bites hardest on the largest firms while the fixed-euro ceiling governs smaller ones — so these headline numbers mean different things depending on a company's size. ⁴

The other story is administrative. Enforcement depends on each member state naming a national authority to police the rules, and here the timeline shows a gap. The European Parliament Research Service confirms that many member states missed the 2 August 2025 deadline to designate those authorities, risking uneven enforcement across the bloc. ¹ One outlet put a number on it — only 8 of the 27 member states had designated contact points as of March 2026, seven months past the deadline — but that specific figure comes from a single low-credibility source, even though the broader "many missed it" finding is solidly corroborated. ³

That gap is the crux of the critics' case. If two-thirds of member states lacked designated enforcers well after the legal deadline, then confirmed fine schedules and confirmed AI Office powers describe a system whose national machinery was, at the point of the record, incomplete. The evidence confirms the enforcement architecture was legislated; it does not confirm the architecture was operational at scale across the EU. ¹

Open: How many of the 27 member states had designated national competent authorities by August 2026, from an official source rather than a single weak outlet.; Whether the Commission opened infringement proceedings against member states that missed the designation deadline.

Beneath the confirmed dates lies a harder question the record only partly answers: why did legislators grant a 16-month reprieve on high-risk rules? Four explanations compete, and the missing roll-call data means none can be fully settled.

One explanation is broad cross-party consensus. On this reading, the Omnibus passed with overwhelming support from the centrist and centre-left groups because they genuinely agreed the timeline extensions were necessary and proportionate. The aggregate margin — 423 to 57 — and the Council's swift final approval fit this account, as do the high winning-side rates of EPP, Renew and S&D in the same period. ⁴ What this explanation cannot supply is the proof: the record holds no group-by-group breakdown of the June vote, and the documented cross-party support (EPP 153-1, S&D 129-0) is from the 2024 Act, not the 2026 amendment. ¹ The hypothesis is supported by the aggregate but unconfirmed at the level that would clinch it.

A second reading holds that opposition was real and principled, from both flanks. The Left voted heavily against the original Act, and the sovereigntist PfE and ESN groups were on the losing side of most March 2026 divisions, with their MEPs framing the rules as overreach and censorship. ³ This is plausible but rests on weak evidence — the floor quotes are single-source and unverified, and there is no data on how these groups actually voted in June. ³

A third explanation is industry influence. On this view, lobbying drove the timeline relief: the DIGITALEUROPE coalition put forward figures of up to €319,000 initial plus €150,000 a year per SME, 30-40% profit erosion, and roughly €500 billion in annual EU-wide compliance cost, and legislators responded by extending the deadlines. ⁴ The sequence is documented — the coalition published its statement in March 2026, and the extensions followed. ¹ But the case has a soft centre: every one of those cost figures comes from a single interested party, the €500 billion estimate is graded unverified with no methodology disclosed, and the record contains no impact-assessment document showing legislators actually relied on them. ¹ The record establishes that industry made the argument and that the deadlines moved; it does not establish that one caused the other.

A fourth explanation points to enforcement reality rather than lobbying. Here, the deadlines moved because they had become infeasible: if many member states had not even named their enforcers by the 2025 deadline, and only a handful had by March 2026, then an August 2026 compliance date faced authorities that did not yet exist. ³ This is a coherent reading of the same confirmed facts, and it does not depend on the contested cost figures. But it too lacks the decisive document — no trilogue record in the file shows member states requesting extensions on capacity grounds. ³ The third and fourth explanations are not mutually exclusive; both could be true, and the evidence to weigh them against each other — official impact assessments and trilogue minutes — is absent.

Open: Whether Commission or Council impact-assessment documents show legislators relied on the industry cost figures or on member-state enforcement-capacity constraints when setting the extensions.; Whether independent economic analysis validates or refutes the €319,000-per-SME and €500-billion EU-wide cost estimates.

The evidence forces a split verdict. What the EU decided is clear and well-documented. The AI Act Omnibus was endorsed by the Parliament on 16 June 2026, approved by the Council on 29 June 2026, and entered into force on 27 July 2026. ⁴ Its substance is confirmed by multiple independent and official sources: a 16-month extension of high-risk compliance to 2 December 2027, near-term prohibitions on abusive AI from 2 December 2026, transparency and general-purpose-AI enforcement from 2 August 2026, and a three-tier penalty regime. ⁴ On the question of what was adopted, the record is strong.

On who supported and opposed it, the record is weak — and this is not a minor caveat but a hole at the centre of the original question. The aggregate margin of the June vote is corroborated, but no breakdown by political group or member state exists in the record for that vote. ² The detailed roll-call the file does contain — EPP, S&D, The Left, and country delegations — belongs to the 2024 Act, not the 2026 amendment. ¹ The winning-side percentages that gesture at a centrist majority come from a single provider and measure a different thing than a single vote. ¹ The documented opposition quotes rest on an unverified aggregator. ¹ Anyone saying they know how the groups split in June 2026 is, on this evidence, guessing.

Two further conclusions are warranted but should be stated at the strength the evidence allows. First, the justification offered for the timeline relief — disproportionate SME costs — rests entirely on interested-party figures that no independent study corroborates. ¹ That does not make the figures wrong; it makes them unproven. Second, the enforcement gap is real: it is confirmed that many member states missed the deadline to name their enforcers. ¹ A regulation whose penalties and inspection powers exist on paper but whose national enforcement authorities were largely undesignated well past the legal deadline is, at minimum, a law whose implementation was still catching up to its text at the point this record closes. ³

One SPECULATIVE observation, labelled as such and offered only as reasoning: if the enforcement gap and the industry cost argument both fed the decision to extend deadlines — which the competing explanations above leave open — then the 16-month reprieve may reflect less a deliberate policy calibration than an accommodation to a system not yet ready to enforce the original schedule. The record neither confirms nor rules this out; the documents that would decide it are absent.