prEN 18229-1: What the AI Act Logging Standard Actually Says

prEN 18229-1 is the draft European standard for logging of AI systems, intended to be harmonised under the EU AI Act and to carry a presumption of conformity with Article 12. We read the enquiry draft in full and filed nine comments on it with DIN on 5 August 2026. Five things in general circulation about this document are wrong:

1. It is titled Logging, not "logging, transparency and human oversight." Human oversight is Part 3. Citing 18229-1 for oversight requirements is a mis-cite.
2. The enquiry already happened. It did not "go to enquiry on 23 January 2026." BSI's window ran 28 May to 21 July 2026; DIN's national window ran to 5 August 2026. Both are closed. Part 3 is still open until 22 September 2026.
3. Article 12 does not apply on 2 August 2026. Standalone Annex III high-risk obligations apply from 2 December 2027, and Annex I embedded from 2 August 2028, under the Digital Omnibus in force since 27 July 2026.
4. Article 12(2) does not give you a field list. It states purposes. The only itemised log schema in the entire Act is Article 12(3), and it applies only to remote biometric identification.
5. The standard will not hand you a field list either, at least not in this draft. Every event category defers its content to another document, and the clause numbers it cites do not resolve against the edition reviewers can obtain.

If you are building AI logging to a deadline right now, points 3 and 4 change what you build and when.

What this document actually is

The draft published for enquiry is E DIN EN 18229-1:2026-07, AI trustworthiness framework, Part 1: Logging (English version of prEN 18229-1:2026). DIN's German title is simply Protokollierung.

The longer title still circulating on aggregator sites and in vendor blog posts, "Logging, transparency and human oversight," describes the original work item, not the document that went out for review. The series split:

This matters more than a naming quibble. Part 3 references Part 1 normatively at every human-oversight intervention function, so the two documents only work read together. A vendor claiming "18229-1 compliance" for oversight logging is claiming it against the wrong part.

The enquiry windows, and the trap in them

The single most useful thing we learned is procedural, and we learned it the hard way.

A CEN draft does not have one comment deadline. It has one per national standards body. The same prEN 18229-1 was open at BSI from 28 May to 21 July 2026, and at DIN until 5 August 2026. We initially recorded Part 1 as closed and unreachable after the BSI window shut, and that was wrong by two weeks and one national body.

If you want to be on the record for a draft whose British window has closed, check DIN, AFNOR, NEN, UNI and the rest before you conclude you missed it. They run their own clocks on the same text.

Two further practical notes for anyone trying this:

• The DIN online portal cannot be used at short notice. It requires a separate application for comment rights plus a signed copyright-assignment declaration (Erklärung der Einräumung der Urhebernutzungsrechte), which is a multi-day loop. The form route at din.de/de/mitwirken/entwuerfe/stellungnahme-formular has no such precondition.
• A German-language enquiry on the translation follows the English one. Closing is not final even at national level.

What Article 12 actually requires, and when

This is where most of the published commentary goes wrong, and it goes wrong in the direction that sells software.

On timing. Article 12 sits in Chapter III Section 2 of the Act. The Digital Omnibus, in force since 27 July 2026, moved the application of those obligations to 2 December 2027 for standalone Annex III systems and 2 August 2028 for Annex I embedded systems. Those dates are unconditional calendar dates. The standards-readiness trigger from the November 2025 proposal did not survive trilogue, so they cannot slip without a new amending regulation.

What does apply from 2 August 2026 is Article 50 transparency. That is a different obligation with a different scope, and conflating the two is the most common error we see.

On content. Article 12(2) tells you what logging must enable: identifying situations that may present a risk or lead to a substantial modification, supporting post-market monitoring, and monitoring deployer operation. Those are purposes, not fields. There is no schema there.

The only itemised list of what a log must contain anywhere in the AI Act is Article 12(3), and by its own terms it applies to the high-risk systems in Annex III point 1(a), remote biometric identification. Not to your recruitment screening tool, not to your credit model, not to your agent.

The draft standard mirrors this exactly: its clause 5.6, the record-keeping requirements, is entirely about RBI systems. Anyone selling you an "Article 12 log schema" for a general high-risk system is selling you an interpretation, not a requirement.

What the standard adds, and what it defers

Read structurally, Part 1's Clause 5 divides cleanly:

• 5.3 identifying events (5.3.1 e) lists human oversight functions among the factors a provider considers)
• 5.4 how to log, taking the information model from ISO/IEC 24970
• 5.5 which events must be logged: hazardous situations (5.5.2), risks missed by risk management (5.5.3), events indicating substantial modification (5.5.4), post-market monitoring including human oversight events (5.5.5), deployer monitoring (5.5.6)
• 5.6 record keeping, RBI systems only

Table 2 summarises the triggers against five logging objectives. Notably, most cells are conditional rather than required.

The catch: every event category in 5.5 defers its actual content to FprEN ISO/IEC 24970. Part 1 tells you that human oversight events shall be logged, and points at 24970 for what they are.

That reference does not currently resolve. Part 1 cites clauses 8.2.x, 8.3.x, 8.5 and 9.1 of 24970. The only edition a reviewer can obtain (ISO/IEC DIS 24970:2025, published nationally as DIN EN ISO/IEC 24970:2025-12) has a Clause 8 containing only 8.1 and 8.2. There is no 8.3 and no 8.5. The referenced subject matter sits one clause level lower: operational triggers at 7.2.x, automated-monitoring triggers at 7.3.x, and human oversight triggers at 7.4. The offset is systematic.

Either the final EN edition renumbered and the renumbering is undeclared, or the citations are wrong. Either way a reviewer cannot check the requirements against what they reference, which is the substance of what an enquiry asks reviewers to do. That was one of our two technical comments.

The gap in Annex ZA

Annex ZA is the part that matters commercially. It is the table that maps the standard's clauses to the AI Act requirements they satisfy, and it is what converts compliance with a standard into a presumption of conformity with the law.

In this draft, Table ZA.1 maps Article 12(3)(a) to (d) to clauses 5.6.1 and 5.6.2 only. But the clauses that actually implement 12(3)(b), (c) and (d) are 5.6.3, 5.6.4 and 5.6.5, which 5.6.1 itself lists.

Read literally, the presumption of conformity covers one of the four requirements and misses three. That was our second technical comment, and it is the one with real consequences: a manufacturer relying on Annex ZA as drafted would have a weaker conformity claim than the standard's own body text supports.

The remaining seven comments were editorial, and they say something about the state of the file: a running header carrying the wrong document number on every page, nine broken cross-reference fields in the definitions clause (rendering as Danish "bookmark not defined" errors inside normative text), a drafting-template placeholder left in a clause heading, a citation to a document number that does not exist, and a requirement list labelled a), a), b), c), d).

None of that is fatal, and none of it is unusual at enquiry stage. It is the reason the enquiry stage exists. But it does mean this text is a long way from being the stable thing you build a product against.

What to do if you are building AI logging now

1. Do not build to a August 2026 deadline for Article 12. It is December 2027 for Annex III, August 2028 for Annex I. Check whether Article 50 applies to you instead, because that one is live.
2. Do not treat Article 12(3) as a general schema. If you are not doing remote biometric identification, it does not bind you, and building to it is a choice you should make knowingly.
3. Do not commit to a standard at enquiry stage. No harmonised standard under the AI Act has been cited in the Official Journal yet. Until one is, there is no presumption of conformity available from any of these documents, and their clause numbering is still moving.
4. Log to your purposes, and record why. Article 12(2)'s three purposes are stable, they were not amended by the Omnibus, and a defensible record of why you log what you log survives a renumbering. Our log retention guide covers the six-month floor under Articles 19 and 26(6).
5. If you want influence rather than compliance, Part 3 is open until 22 September 2026. It is free, it is open to anyone, and participation is near zero.

How we know this

Everything above comes from reading the enquiry drafts directly, through a university licence, not from secondary coverage. We filed nine comments on prEN 18229-1 with DIN on 5 August 2026 (two technical, seven editorial) and three comments on prEN 18229-3 with BSI on 3 August 2026, in a personal capacity, as the first comments on that file.

We have not reproduced the standards' text here, and we cannot: they are copyrighted, and the value of reading them is in the structure and the cross-references, not in quotation.

If we get something wrong, we would rather be corrected than cited. Every claim above is dated and points at a clause you can check.