Transparency

What we have corrected

We publish legal information that people base decisions on. An error of ours should not be quietly overwritten. This page records every substantive correction we made to a published object or public page: what it said, what is correct now, why it moved, and which objects and pages it touched.

Register kept since: 8 August 2026
Recorded corrections: 8
Latest correction: 5 October 2026
Report an error: info@embedai.nl

What we have corrected

Material5 October 2026praxikon:sys:registry:correction:2026-10-05-code-of-practice-paragraphs
What it said
Signatories of the code of practice can rely on it for Article 50(2), (3) and (5).
What it says now
After a positive adequacy assessment by the Commission, signatories can rely on it for Article 50(2), (4) and (5).
Why
Article 50(7) names paragraphs 2 and 4; the guidelines (points 146 to 148) and the Commission page name (2), (4) and (5). Paragraph 3 (emotion recognition) is outside the code.
Pages touched
Open the primary source
Material5 October 2026praxikon:sys:registry:correction:2026-10-05-code-of-practice-status
What it said
The code of practice on AI content transparency is still a draft or forthcoming.
What it says now
The final, voluntary code of practice was published on 10 June 2026; the final Article 50 guidelines on 20 July 2026.
Why
The posts were written before publication and had not been updated.
Pages touched
Open the primary source
Material5 October 2026praxikon:sys:registry:correction:2026-10-05-deployer-labelling-scope
What it said
Deployers must visibly label AI-generated or manipulated content; anyone using AI images or video in campaigns or course material has marking and disclosure duties.
What it says now
A deployer only has a disclosure duty for deep fakes and for AI text published to inform the public on matters of public interest without human review and editorial responsibility. Machine-readable marking is a provider duty.
Why
Article 50(2) and (4) and the guidelines of 20 July 2026, points 113 to 116 and 131 to 138.
Pages touched
Open the primary source
Material5 October 2026praxikon:sys:registry:correction:2026-10-05-emotion-recognition-scope
What it said
Emotion recognition at work and in education is absolutely prohibited, including text analysis; transparency follows from Articles 86 and 87.
What it says now
The ban has a medical or safety exception and requires biometric data; text analysis falls outside. Outside the ban, emotion recognition is high-risk (Annex III point 1(c), from 2 December 2027) and the Article 50(3) information duty applies.
Why
Article 3(39), Article 5(1)(f) and Article 50(3).
Pages touched
Open the primary source
Material5 October 2026praxikon:sys:registry:correction:2026-10-05-fines-smaller-undertakings
What it said
Fine of up to EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher, without exception.
What it says now
The higher amount applies, except for SMEs, start-ups and small mid-caps, for whom the lower applies.
Why
Article 99(6) and (6a), inserted by Regulation (EU) 2026/1744.
Pages touched
Open the primary source
Material5 October 2026praxikon:sys:registry:correction:2026-10-05-nl-supervisors-designation
What it said
In the Netherlands the supervisory authorities have been designated, with the Dutch DPA as coordinator.
What it says now
The Netherlands has not yet formally designated any supervisor by law; the bill intends the Dutch DPA and the RDI as coordinators.
Why
The bill for the Dutch AI Act implementation law has not yet been submitted.
Pages touched
Material5 October 2026praxikon:sys:registry:correction:2026-10-05-omnibus-status-and-transition
What it said
The Digital Omnibus agreement still has to be formally adopted; the 2 December 2026 date was not granted by the Omnibus; Article 4 is unchanged.
What it says now
Regulation (EU) 2026/1744 has been in force since 27 July 2026. It added the transition to 2 December 2026 for the paragraph 2 marking (Article 111(4)) and reworded Article 4 as a duty to take measures.
Why
Status claims dated from before publication in the Official Journal.
Pages touched
- /en/posts/article-50-transparency-deadline-2-august-2026
- /nl/posts/artikel-50-transparantie-deadline-2-augustus-2026
Open the primary source
Material5 October 2026praxikon:sys:registry:correction:2026-10-05-provider-test-and-obvious-exception
What it said
Whether you become the provider follows from Article 25; a label such as digital assistant may meet the obviousness exception.
What it says now
Outside high-risk, provider status follows from Article 3(3) (guidelines point 11). The exception is narrow: almost no doubt, and a generic label such as assistant is not enough.
Why
Article 25 governs high-risk systems; guidelines points 11, 38, 44 and 45.
Pages touched
Open the primary source

What belongs here and what does not

Yes: our own substantive errors: A wrong date, a misassigned duty holder, a source locator pointing at the wrong paragraph, an interpretation the source does not carry.
No: a change in the law: When the rules themselves move, that is not a correction but a new version of the object. Those live in the changes timeline and in the version history of the obligation itself.
No: editing that does not move the meaning: A rewritten sentence that says the same thing belongs in the version history, not here. Otherwise the real correction disappears into the noise.
Severity is stated: Material means the meaning moved. Editorial means the substance held while a wording, reference or locator was repaired.

The dataset is append-only: a corrected statement does not replace the old one but sits beside it with a later knowledge date. Ask for an earlier snapshot and you get back what we published then, error included. This register supplies the missing part: that it was our error, and why.

How a correction reaches this register

we compare the disputed passage with the primary source and with the version that applied at the time;
if the passage is wrong, we correct the current publication;
the correction receives a stable identifier, a date and an explanation, and names the objects it touched by identifier;
the earlier object version stays retrievable, so a citation already made remains verifiable.

For AI agents and integrations

The register is available as data as well, with the same content and the same identifiers as this page. It is always returned in full and never truncated.

Open the register as JSON Read the methodology