Future Proof The Authority Stack
Independent European Publication AI Act Operator Desk · Regulation 2024/1689 24 April 2026
Agent Liability AI Act Operator Desk
Agent Liability EU

Corrections Log

Every material correction issued on this publication is recorded here, with the date, the page affected, what was wrong, and what replaced it. Corrections are never made silently.

The Editors Last reviewed 17 August 2026

How corrections work

This page is the public record required by section 5 of our editorial standards. It exists so that a reader who relied on an earlier version of a page can see exactly what changed.

A material correction is a change to a fact a reader could have acted on: a date, a figure, a legal citation, an article number, a named authority, a carrier or product name, or the substance of an obligation. Material corrections are logged here and carry a dated note on the affected page.

A non-material change is a typographical fix, a broken link repair, a clarification of wording that does not alter meaning, or a routine update to a tracker whose entries are dated by design. These are not logged individually.

Log

17 August 2026: freshness sweep. The Digital Omnibus is in force and this desk had not said so.

The largest correction this publication has had to make, and the one with the clearest cost to a reader. The Digital Omnibus on AI entered into force on 27 July 2026. Until today, most of this site still described it as a proposal in trilogue whose outcome was open, and the sitewide banner still read "trilogue 28 April 2026. The 2 August 2026 high-risk deadline may shift to 2 December 2027." That banner was live on 60 pages, three weeks after the amendment became law. A page that tells a compliance officer a settled deadline "may shift" is worse than no page.

  • The banner. Now reads that the Omnibus has been in force since 27 July 2026, that Annex III high-risk obligations apply from 2 December 2027 and Annex I from 2 August 2028, and that Article 5, Article 50 and the GPAI obligations are unchanged. Normalised across every page on this site; there were thirteen different versions of it, several of them contradicting each other.
  • The editor's update on 32 pages. The block that said the Omnibus "proposes to push" the deadline and that "the original 2 August 2026 deadline remains legally binding" has been replaced with the settled position on every page that carried it.
  • The Digital Omnibus Tracker. Frozen at 25 April 2026, with a live countdown to a date that had already been superseded, a legislative stage panel reading "Omnibus is in Trilogue", and a 90 day watch calendar of meetings that had all since happened. Rebuilt: the passage record now runs to entry into force, the provision matrix is rewritten against the Commission's own implementation timeline, and the calendar now lists the dates that are actually ahead. The tracker also cited the Commission proposal as COM(2025) 529 of 26 November 2025. It is COM(2025) 836 of 19 November 2025.
  • The Article 50(2) transitional period. This site said providers of synthetic content systems already on the market had a six month grace period running to 2 February 2027. That was the Commission's proposal. The adopted text is shorter: the period ends 2 December 2026. This is the correction most likely to have cost a reader real time, because it moves a deadline two months earlier, and it is the next binding date on this site.
  • What the Omnibus added. This desk reported the Omnibus purely as a deferral. It also added a prohibition on AI systems that generate child sexual abuse material or that depict an identifiable person's intimate parts without consent, with systems already on the market to comply by 2 December 2026. That was missing entirely and is now in the tracker and on the homepage timeline.
  • Article 4 AI literacy. Our Article 4 guide said the provision was untouched by the Omnibus. The adopted text amended it: AI literacy remains an obligation on providers and deployers, but no specific or sufficient level is mandated, and the Commission and Member States take a stronger role. The training duty for staff exercising human oversight of a high-risk system is unchanged.
  • Structured data. The homepage served a schema.org Event advertising a Digital Omnibus trilogue on 28 April 2026 with eventStatus EventScheduled. An AI assistant reading this site would have been told a legislative meeting was upcoming four months after it happened. Replaced with a Legislation node describing the adopted act.
  • Two archived pages. The April 2026 Omnibus master brief and the 10 July 2026 three week checklist are kept as dated records and now carry superseded notices at the top and in their metadata. The master brief promised it would be updated within 24 hours of adoption. It was not.

17 August 2026: Member State implementation tracker, re-verified at national sources

Fifteen Member States in this tracker were re-read at their own ministry, regulator, parliament or official gazette. Four entries were wrong, four were stale in a way that misled, and one rested on a claim that could not be found anywhere.

  • Germany. Recorded as a draft in consultation. The KI-MIG has been in force since 29 July 2026, BGBl. I 2026 Nr. 223.
  • Poland. Recorded as proposed. The Act on AI Systems was published Dz.U. 2026 poz. 1003 and has been in force since 11 August 2026. KRiBSI is not yet constituted.
  • Sweden. Recorded as pending, with IMY and Digg as likely co-supervisors. Five authorities were assigned by government decision on 12 June 2026: PTS, IMY, Finansinspektionen, Lakemedelsverket and Swedac. PTS holds the main responsibility. Digg is not designated.
  • Ireland. Three errors. The fifteen competent authorities were designated on 16 September 2025, not in March 2025, when a roadmap named an initial eight. HIQA was listed as the health authority; it is not among the fifteen. And the single point of contact is now the AI Office of Ireland, established under the Regulation of Artificial Intelligence Act 2026 and operational since 2 August 2026, not the Minister for Enterprise.
  • Netherlands. The five authority model from the November 2025 advisory was superseded by the draft bill consulted from 20 April to 1 June 2026, which names ten authorities and does not include ACM.
  • Denmark. Recorded as having implementing legislation pending. LOV nr 467 of 14 May 2025 is enacted.
  • France. This tracker told deployers to treat CNIL as the contact. The published DGE scheme makes DGCCRF the coordinator and single point of contact; CNIL is proposed for defined use cases and nothing is adopted.
  • Belgium. BIPT was named as the proposed market surveillance authority, attributed to a Government Declaration of 31 January 2025. That could not be found on any Belgian official domain, including BIPT's own AI Act page. Withdrawn.
  • Portugal. AMA was named as lead. It appears in no official AI Act material. ANACOM coordinates the 14 designated fundamental rights bodies.
  • Latvia. Recorded as pending. The Consumer Rights Protection Centre appears on the Commission's own market surveillance authority list without a pending marker.
  • Counts we could not confirm. Finland's ten sectoral authorities, Sweden's four fundamental rights bodies and Austria's three advisory forums were all stated as facts and none could be confirmed at a national source. All three counts have been withdrawn rather than restated.
  • Article 5 date. The article body and the sidebar both said the prohibitions applied from 2 February 2026. They applied from 2 February 2025.
  • The source itself. This tracker described the Commission's market surveillance authority list as "updated continuously". Its own page carries a last update of 26 September 2025, and it is demonstrably behind for at least Poland and Sweden. The methodology note now says so, and entries re-read at source are dated in the table so a reader can tell them from entries that still rest on the April 2026 research.

17 August 2026: third-party claims flagged by the claims guard

  • Article 101 penalty ceiling. The Article 99 guide said Article 101(2) sets a lower ceiling of EUR 3 million or 1 per cent for GPAI model providers. Article 101 sets EUR 15 million or 3 per cent of worldwide annual turnover, whichever is higher. Read at the Commission's AI Act Service Desk.
  • Who enforces what. The penalties tracker attributed the Article 5 ceiling of EUR 35 million or 7 per cent to the AI Office. National market surveillance authorities enforce it. The AI Office holds the separate general-purpose AI regime under Article 101.
  • EIOPA's product name. Three pages called EIOPA's August 2025 output a supervisory statement, supervisory guidance, or an AI governance opinion in lower case. It is the Opinion on AI governance and risk management, published 6 August 2025. An Opinion is supervisory guidance, not binding law, and the name matters for that reason.
  • Beazley and QBE. Our professional liability article said both had introduced AI sublimits capping payouts at roughly 10 per cent of the limit. Neither the carriers nor the figure could be confirmed at either carrier's own domain. Both removed; the structural point about sublimits is kept without the number.
  • A Code of Practice on marking AI-generated content. The Article 50 guide cited a first draft of 17 December 2025 and a second draft of March 2026, and two country guides cited a Commission Article 50 guidance published in June 2026. None could be found at digital-strategy.ec.europa.eu. All withdrawn with a sourcing note rather than restated.
  • FCA "AI Update", February 2024. The document and its date could not be confirmed at fca.org.uk. The citation is dropped; the FCA's substantive position, which it has stated repeatedly, is kept without it.
  • ISO commercial general liability form numbers. The two articles that rely on these numbers now carry a dated sourcing note stating which forms have been read from a specimen document, CG 40 48 01 26 and CG 35 08 01 26, and which are reported but unread, CG 40 47 and PC 51380. The note also says plainly that a search engine confirming these numbers may be reflecting this network's own pages back at the reader.

15 August 2026: third-party claims about named insurers

A review of every claim on this site that names a real insurer, reinsurer, syndicate or standards body. Where a claim could not be verified at that party's own domain it was removed rather than softened.

  • Munich Re aiSure described as parametric. The AI insurance coverage topic index and the Article 17 deployer guide described aiSure as parametric. Munich Re does not use that word for aiSure anywhere it publishes. Munich Re and Mosaic Insurance describe it as a performance guarantee for AI systems that settles on measurable performance data. The label was ours and has been removed.
  • Munich Re Special Enterprise Risks. The Article 17 deployer guide attributed aiSure to a Munich Re Special Enterprise Risks division. No such divisional attribution is published. It has been removed.
  • Hamilton Insurance Group. Our AI exclusions article described Hamilton as applying a management liability AI exclusion across directors and officers and errors and omissions cover, and listed it that way in the carrier comparison table. Other trade reporting describes Hamilton's position instead as a sublimit inside errors and omissions or cyber cover. For a buyer those are opposite outcomes: one removes the cover, the other caps it. Hamilton publishes nothing about AI underwriting on its own site, so neither version can be checked at the carrier. The page now states neither and says why, and the table row reads as unconfirmed.
  • Testudo capacity expansion. The April 2026 briefing dated the Atrium and QBE capacity expansion to March 2026. Testudo dates it 26 February 2026. Corrected.
  • EIOPA opinion reference number. The reference EIOPA-BoS-25-360, cited across this site for the Opinion on artificial intelligence governance and risk management, was checked against the cover page of the published PDF and is correct. No change was needed. It is recorded here because an earlier review across the network could not confirm it and removed it elsewhere.

17 August 2026: Lloyd's syndicate numbers

Lloyd's runs a public market directory that needs no login. The syndicate numbers this site cited were read there. One of the two could not survive the reading.

  • Chaucer syndicates 1084 and 1176. The AI exclusions article, its reference list and the April 2026 briefing all said Armilla writes standalone AI liability through Chaucer's Lloyd's syndicates 1084 and 1176. Half of that is false. Lloyd's own market directory lists 1084 as Chaucer Syndicates 1084, active, writing classes that include cyber, professional indemnity and directors and officers. It lists 1176 as Nuclear Syndicate 1176, whose classes of business are cargo, engineering, nuclear and terrorism. A nuclear syndicate does not write AI liability. Beyond that, no source names any syndicate number for this product: Chaucer's own announcement of the Vanguard AI structure says only that it is backed by Lloyd's of London. Both numbers have been removed rather than reduced to the defensible one, because keeping 1084 alone would still be our inference rather than anyone's published statement. What is published, and is retained, is that Armilla Insurance Services is a Coverholder at Lloyd's and that the structure was launched with Chaucer on 10 February 2026. Verified at ldc.lloyds.com and chaucergroup.com.
  • Aggregate limits stated as USD 25 million or more. The same two pages gave Armilla's AI aggregate limits as USD 25 million or more per organisation. Armilla publishes a Standalone AI Liability Policy with limits of up to USD 25 million per organisation. Up to and or more are opposite claims about the same number. Both pages now use the published form. Verified at armilla.ai.

Reporting an error

If you believe a page on this site states something that is wrong, write to editors@agentliability.eu with the page URL and the specific statement. We answer every report. Where a claim cannot be verified at the source that issued it, we remove the claim rather than soften it.

Corrections practice across the wider network is published on each site: