AI Center › Corrections
1ArtificialIntelligence · The AI Center
Corrections
Corrections the AI Center has made to published pages, with what each used to say, what it says now, and how the error was found.
This record exists because the rest of the AI Center asks you to trust statements on the grounds that each names its source. That is only testable if you can also see what happened when a statement turned out to be wrong. Entries are listed most recent first, and withdrawn wording is quoted rather than summarised so you can judge for yourself whether you relied on it. This is not a log of every edit: it records changes to what a page asserted. Where an audit produced several findings of the same kind, they are entered as one correction and the entry says how many cells or pages were affected.
14 corrections recorded, most recent first.
We corrected two things on the page and left the machine-readable copy uncorrected for weeks
Open data exports for the Standards Directory and the Governance Register ↗
On 16 July we stopped describing voluntary standards in the vocabulary of binding law. On 1 August we withdrew 128 lifecycle claims from all 29 rows. Both corrections reached the visible page and neither reached the CC BY export, which anyone may download and republish. Until today that file still carried a next_milestone column holding the withdrawn claims, and competent_authority, max_penalty and two treaty-signatory columns for documents that have none of those things. The export tool included a column whenever the key existed, without checking whether any row had a value in it.
Both exports were rebuilt and republished today, and the superseded files were deleted and confirmed unreachable. The Standards Directory export went from 23 columns to 15. The tool now includes a column only where some row has a value, and drops any column that exactly duplicates another on every row.
How it was found: An adversarial review of the draft of THIS page. It challenged an entry that claimed the machine-readable data had been fixed; the claim was checked and was false. Exposure: 16 July to 3 August for the legal columns, 1 to 3 August for the withdrawn claims.
We published 128 claims about where standards apply, and wrote 65 of them ourselves without a source
Standards & Assurance Directory ↗
Every one of the 29 standards carried a field headed “Where it applies in the AI lifecycle”, stating which of the AI Center’s fourteen lifecycle stages the document applied to and which EU AI Act articles it bore on. We audited all 29 rows claim by claim. Of 128 lifecycle-stage claims, 63 were supported by the standard’s own scope, 63 were our own reading with no source, and 2 were wrong. Of roughly 149 article assertions, four traced to a named authority. We removed the field from all 29 rows rather than hedging it.
“Lifecycle stage 7 (risk management + FRIA / Arts 9+27) most directly, with bearing on stage 1 (problem framing / Art 6) and stage 3 (data preparation + documentation / Art 11 + Annex IV).”
The field is gone. What replaced it can hold only a direct quotation of a named authority, with that authority’s date and source. As of today not one of the 29 rows carries such a quotation — the four authority-stated links could not be attributed to the standards as they now stand. The withdrawn wording for all 29 rows is retained in full in our records.
How it was found: A claim-by-claim provenance audit, run before building a navigation feature that would have depended on those claims. Published 16 July, withdrawn 1 August.
We described a legal effect that does not exist
Standards & Assurance Directory ↗
The entry for EN 18286:2026 — Artificial Intelligence: Quality Management System for EU AI Act Regulatory Purposes, published by CEN-CENELEC JTC 21, and carried by us at the time under its draft designation prEN 18286 — told readers that applying it would place them in a particular legal position. It would not. Article 40(1) of Regulation (EU) 2024/1689 attaches a presumption of conformity only to standards whose references have been published in the Official Journal of the European Union, and no standard has been published there under the AI Act. This was the most serious error on the site: a reader could have relied on it.
“It is the first candidate to become a harmonised standard for the AI Act, which would let providers who apply it be presumed to conform with the corresponding legal requirement.”
The entry records what CEN and CENELEC state about the standard, and separately quotes Article 40(1)’s own condition, asserting no relationship between them. A check now runs before every publication that blocks this shape of sentence, and it is tested against this exact wording.
How it was found: An adversarial review of the page’s claims against their primary sources. Published 16 July, corrected 31 July — fifteen days public.
We stated the opposite of what the European Commission’s own research service had published
Standards & Assurance Directory ↗
Our entry for ISO/IEC TS 6254 said that human oversight under Article 14 of the AI Act requires an interpretable system. The Commission’s Joint Research Centre examined that exact question in EUR 31518 EN and concluded that the tools listed in that standard are not necessary to ensure human oversight as described in Article 14. We asserted the reverse of a finding by the body best placed to make it.
“Art 14 (human oversight — oversight requires an interpretable system)”
The gloss is removed. The article reference remains, with no claim about what it requires.
How it was found: The same review, checking our article pairings against the JRC’s published analysis. Published 16 July, corrected 31 July.
We made a false statement about how EU legislation was drafted
Standards & Assurance Directory ↗
Our entry for ISO/IEC 23053 claimed that Annex IV of the EU AI Act had been written against that standard’s component model. Annex IV is an annex to Regulation (EU) 2024/1689, adopted by the co-legislators. It was not drafted against an ISO document, and no source says it was.
“Provides the descriptive component model that Annex IV system-description items are written against.”
We deleted the sentence. There is no true version of it, so nothing replaced it.
How it was found: The same review. Published 16 July, corrected 31 July.
We implied the AI Act takes its conformity-assessment vocabulary from an ISO standard
Standards & Assurance Directory ↗
Our entry for ISO/IEC 17000 implied that the AI Act’s conformity-assessment framework rests on that standard’s vocabulary. It does not. The Act is built on the New Legislative Framework and aligns with Regulation (EC) No 765/2008, Decision No 768/2008/EC and Regulation (EU) 2019/1020.
The entry now states which instruments the Act’s conformity-assessment framework is built on, and does not present the standard as their source.
How it was found: The same review. This is the fourth of the four false statements corrected that day; an earlier draft of this page listed only three, and the omission was caught by a review of the draft.
We published internal editing notes in three cells instead of the facts they were meant to carry
Global AI Governance Register ↗
The Council of Europe row’s field headed “Official reference” carried a note we had written to ourselves rather than the citation. Two other rows carried a verification to-do and an instruction about how to label a bill. These were notes from our own review process and we published them by mistake. They are the last three of a larger set — see the 28 July entry below.
“Optionally refine to cite the corrected version: a certified copy of the Convention in its corrected version per the Committee of Ministers’ decision of 30 April 2025 is published on the Treaty Office page.”
The field carries the citation: CETS No. 225, opened for signature at Vilnius on 5 September 2024, with the Treaty Office’s corrected version noted. The other two cells state their facts in the reader’s voice — including, where a date rests on a ministry’s statement rather than the statute record, saying so.
How it was found: A sweep for editor-facing language across every page. Our check for this had been written on 28 July and did not catch these three, so we widened it.
We described a European standard as a draft after it had been approved and published
Six pages, including the Standards & Assurance Directory ↗
We took EN 18286:2026’s status from a European Commission page instead of from CEN-CENELEC, who issue the standard and announce its approval. That was the wrong source for the question, and the Commission page lagged. Six of our pages carried the draft designation and described the standard as under development.
The designation, status and dates are corrected across six pages. CEN-CENELEC’s newsletter states approval in June 2026; its project record gives ratification on 12 July and availability on 22 July. We record all three and say which record each comes from, because they are different events. The standard is still not cited in the Official Journal.
How it was found: Research for a separate feature happened to check the standard’s current status. Approved June 2026, corrected by us 31 July.
We left two superseded data files publicly downloadable, one holding material we had deliberately withheld
Global AI Governance Register export ↗
We published a revised export without deleting the earlier ones, and we did not check where our own public links resolved. Each earlier file stayed publicly served, and the first held the register’s most obvious address — so that address redirected anyone who used it, reader or answer engine, to a superseded 18-row file rather than to the register. That file contained the United Arab Emirates, which we had excluded from the published register because we could not verify it against a primary source, and internal reviewer notes in five of its cells.
Both files were deleted and confirmed unreachable, and the address now resolves to the register. The publishing tool retires prior versions automatically. What we cannot tell you: we do not have access to request logs for those files, so we cannot say whether anyone downloaded them, and we cannot rule out copies in third-party caches or archives.
How it was found: A post-build audit that followed the register’s own public links instead of assuming they resolved as intended. The files were reachable from 26 to 28 July.
We published internal reviewer reasoning in ten cells of the register
Global AI Governance Register ↗
Ten cells across the register carried instructions to an editor, and our own critique of the register, instead of the jurisdiction’s position. Our earlier check for this was anchored to the start of a cell, so it caught an instruction that opened a sentence and missed the same voice in the middle or at the end.
“A flat ‘non-binding’ … slightly overstates the position … Do NOT display ‘National Cabinet August 2026’ … Leave empty or state ‘none announced’”
All ten were rewritten as displayed facts. The reasoning in them was sound; it belonged in a review log, not in the record. The check now reads the whole cell rather than its opening, and was later widened again after three more cells survived it.
How it was found: A post-build audit across seven review lenses.
We applied evaluative labels to other organisations, and misdescribed who had made a set of commitments
Several pages across the AI Center ↗
We wrote framing text that ranked other bodies. Describing what an organisation publishes is neutral; ranking it is not, and it is not ours to do. In the same audit: we had described the Frontier AI Safety Commitments as intergovernmental, when they are commitments made by AI companies and announced by the UK and Korean governments; and we had converted a penalty from Korean won into US dollars ourselves, presenting our own arithmetic as though it were the source’s.
“The recognized authorities on AI safety publish their own evaluations, frameworks and records. … The AI Center attributes the specialists that own this ground. … Flagship recurring syntheses.”
The evaluative language is removed, the commitments are described as what they are, and figures appear in the currency the source states. A check now blocks our copy from conferring rank on a third party.
How it was found: The same seven-lens audit. Six of its ten confirmed findings were this one class of error.
We told readers a fining power was in force a year before it was
Three AI Lifecycle stage pages ↗
Article 113 applies Chapter XII of the AI Act from 2 August 2025 “with the exception of Article 101”. Three pages listed the exceptions and omitted that carve-out, so each stated that the Commission’s power to fine providers of general-purpose AI models had been in force since 2 August 2025. A reader planning around that enforcement could have relied on it.
All three pages carry the Article 101 carve-out.
How it was found: The same seven-lens audit, reading Article 113 against what our pages said about it.
We described the EU AI Act’s timetable as it stood before it was amended
Seventeen pages, including all fourteen lifecycle stages ↗
Regulation (EU) 2026/1744 amended the AI Act’s application dates, deferring the high-risk obligations. Our pages described the amendment’s status in terms the Official Journal had already overtaken.
“the Digital Omnibus, adopted but unpublished”
Seventeen pages cite the Act as amended, with the deferred dates. A standing watch list now records known future legal dates and warns before they arrive. It would have caught this one. It cannot catch a change we have not anticipated.
How it was found: A person re-checked EUR-Lex by hand. Nothing in our process would have caught it, which is why the watch list now exists. The amendment was published on 24 July; we corrected on 26 July.
We described 29 voluntary standards using the vocabulary of binding law
Standards & Assurance Directory ↗
We built the directory by reusing the layout made for the register of AI laws, and published it without changing the labels that came with it. Voluntary standards were presented as having a legal status, a competent authority, enforcement powers and a maximum penalty. None of those things exists for a voluntary standard.
“Does it bind yet? · Who enforces it · What the regulator can do · Maximum penalty”
Each type of page has its own vocabulary, and a check blocks legal labels from appearing on pages about standards. The machine-readable export was not corrected at the same time — see the 3 August entry at the top of this page.
How it was found: A readiness audit, before the page had been live for a full day.
How this record is kept
- Every entry gives the date, the page, what was wrong, what the page says now, and how the error was found.
- Withdrawn wording is quoted rather than described. Saying only that an error was corrected tells a reader nothing they can check. Where an error spanned many cells, a representative extract is quoted and the entry says how many were affected.
- Entries record changes to what a page asserted — a fact, a date, a status, an attribution, or a claim about what a rule requires. Typography, layout, and adding a citation to a statement that was already accurate are not listed.
- Where we can say how long an error was public, the entry says so. Where we cannot, it says that too.
- Where a correction changed an openly licensed data export, the export is rebuilt and the superseded file deleted.
- Where an error revealed a gap in our checks, the entry names the check we added or widened to close it.
- This page is maintained by the AI Center. It is not audited by anyone outside 1BusinessWorld, and it begins on 16 July 2026, when the AI Center was first published.
Cite this page
1BusinessWorld AI Center — Corrections Version as of 2026-08-03.
The AI Center is informational only. It is provided by 1BusinessWorld strictly for general informational and educational purposes. Nothing in the AI Center constitutes, or should be construed as, legal, regulatory, compliance, technical, engineering, security, investment, financial, or other professional advice, or a recommendation, endorsement, solicitation, or offer regarding any technology, product, model, provider, framework, or course of action. 1BusinessWorld is not a law firm, regulatory authority, standards body, conformity-assessment or certification body, or investment adviser, and nothing in the AI Center creates any advisory, fiduciary, attorney-client, or other professional relationship with 1BusinessWorld. Although the AI Center references official materials published by legislatures, regulators, standards bodies, research organizations, and other named authorities, 1BusinessWorld makes no representation or warranty, express or implied, as to the accuracy, completeness, timeliness, or fitness for any purpose of any content, and, to the fullest extent permitted by law, disclaims all liability for any loss or damage of any kind arising directly or indirectly from the use of, or reliance on, any information presented. Laws, regulations, standards, technical practices, and AI capabilities change frequently and differ by jurisdiction; readers must verify all information against the current official text or source and consult qualified legal, compliance, technical, and other professional advisors before acting. Any decision relating to the development, deployment, procurement, or governance of AI systems is made solely at the reader's own risk. Last reviewed: 2026-08-03.