A stamp born for lifts
The mark is made of two letters. It sits on the plate inside the lift, on the bottom of the pressure cooker, on the phone charger. Its proportions are not free: a European annex fixes its grid — the C and the E drawn on circles of the same diameter — and provides that it must never be less than five millimetres high.
That mark does not say the product is good. It says that someone has declared that it is. In most cases that someone is the manufacturer. It is the device invented in 1985 to let goods circulate: the law writes the requirements in general form, the technical standards fill them in, whoever follows the standards is presumed compliant, and the mark seals the presumption. It worked well for boilers.
Article 48 of the 2024 regulation prescribes the same mark for high-risk artificial-intelligence systems. Same grid, same minimum height. On the goods lift and on the program that ranks CVs. On the pressure cooker and on the algorithm that decides whether a person deserves a loan.
From 2 August 2026 that mark was due to start appearing.
The appointment with 2 August
Regulation (EU) 2024/1689 has been in force since 1 August 2024, but it applies in stages. From February 2025 the prohibitions: social scoring, manipulation, emotion recognition in the workplace and at school. From August 2025 the obligations on general-purpose models. From August 2026 the rest, that is, the heart: the high-risk systems of Annex III — work, credit, education, essential services, justice, migration — with substantive requirements, conformity assessment, registration, marking.
2 August 2026 arrived. The heart did not.
On 24 July, Regulation (EU) 2026/1744 was published in the Official Journal, adopted on 8 July and in force from the 27th — the third day after publication, under an urgent procedure. Six days’ margin. It postpones the high-risk obligations to 2 December 2027 for Annex III and to 2 August 2028 for systems embedded in products.
The reasoning is in the text, and it deserves to be read without comment. Recital 40 points to the delay in the availability of the technical standards, the common specifications and the guidance, and the delay in setting up the competent national authorities. Recital 2 adds that the late preparation of the standards and of the national governance frameworks has already made the compliance burden heavier than expected.
Translated: the law cannot apply because its instructions for use, and the offices that were meant to enforce it, are missing. The legislator writes this, in a legislative act, about itself.
What stays on
It is not that nothing happened on 2 August.
The prohibitions remain, in force for a year and a half. The transparency obligation of Article 50 remains — whoever talks to a machine must be able to know it, generated content must be recognisable — with four months of transition for those already on the market, and therefore with the real appointment falling on 2 December 2026. And above all there remain, from 2 August, the Commission’s powers over the providers of the largest models: requests for information, evaluation of the models including with independent experts, corrective measures up to restriction, penalties. The ceiling, for these providers, is three per cent of worldwide turnover or fifteen million — not the seven that circulates in the summaries.
The July regulation then adds two new prohibitions: systems that generate non-consensual intimate images and child sexual abuse material. Recital 11 qualifies the first as a form of sexual violence and abuse, in particular against women. The operative part prohibits it from 2 December 2026. Between the qualification and its taking effect run four months, and in the public reporting the prohibition was given as immediate.
There is more, and it is not found in the summaries. For whoever manufactures the system the prohibition applies in only two cases: if generating that material is the intended purpose, or if the outcome is reasonably foreseeable and reproducible and reasonable and adequate technical measures are lacking. Which measures? Recital 12 lists them: data cleaning, training to refuse, prompts designed for safety, run-time barriers, filters, restrictions of use, abuse detection, reporting procedures. These are all things that can be documented. The harshest prohibition in the package is defended, like everything else, with a folder of papers.
The changes nobody talked about
The postponement took the headlines. The rest of the July regulation did not — and it is the rest that remains, because a date passes and an amendment does not.
Literacy. Article 4 required those who provide and those who deploy these systems to ensure a sufficient level of competence in their staff. It now requires them to adopt measures to support its development, and specifies that the obligation does not entail guaranteeing a specific level for anyone. It was the first obligation to enter into application, in February 2025. It is also the first to have been watered down, eighteen months later, without any public verification on record of how it had been implemented.
The scope. The notion of a safety component has been rewritten and narrowed; and the new Article 6 clarifies that it does not cover systems used only for user assistance, performance optimisation, service efficiency, automation, convenience or quality control. Recital 7 declares the purpose: to prevent the high-risk classification from extending beyond what is justified.
The requirements. A new paragraph of Article 2 allows the application of Articles 9 to 15 and 17 to 25 — risk management, data governance, documentation, traceability, human oversight, robustness, quality system, providers’ obligations — to be limited where sectoral legislation offers equivalent protection. Where, when and how much will be decided by the Commission, with delegated acts, by 2 August 2027.
The verification. The new third paragraph of Article 43 puts in writing that the manufacturer is not required to choose a procedure with a third-party body merely because the product incorporates a high-risk system, and that it may continue to rely on the harmonised standards. That is: independent checking remains the exception, and the regulation that simplifies says so.
The codes of conduct. Recital 41 clarifies that they have limited legal effects and that, in particular, they do not confer a presumption of conformity. Useful to remember the next time the discussion turns to who signed the code on general-purpose models and who refused: the signature does not produce the effect attributed to it.
Bias. A new article builds the legal basis for processing sensitive data where necessary to detect and correct bias, with six cumulative conditions. Then it closes with a sentence worth the whole article: the provision creates no obligation to detect or correct anything. It authorises. It does not require. Worth keeping in mind, because on the other side of the Atlantic a declaration by the trade authority is out for consultation which starts from the opposite premise: correcting bias would make the model less faithful to the data, and therefore deceptive.
On all this, for the record, an opinion existed. The European Data Protection Board and the European Data Protection Supervisor, on 20 January 2026, had asked that the calendar be kept and, if that really was not possible, that the delay be reduced to a minimum; they had recalled that the processing of sensitive data remains in principle prohibited; they had observed that promoting literacy should be added to the obligations, not substituted for them. The final text took up the point on regulatory sandboxes, where the data protection authorities are now involved. On the calendar, no.
The standards that do not exist
Here it is worth slowing down, because this is the point on which everything rests.
The presumption of conformity — the device that makes self-declaration workable — applies if the provider follows the harmonised technical standards. The regulation does not contain them. It commissioned them from two European standardisation organisations, which are writing them in a joint committee, with experts largely put forward by the regulated companies.
At the date of this article, no reference to any of those standards appears published in the Official Journal of the Union. The drafts exist and are at different stages: the one on the quality management system is at the formal vote, others are in public enquiry, the one on bias is still being drafted. A draft, however, produces no legal effects. In October 2025, to speed things up, the standardisation organisations adopted exceptional measures, including the possibility of publishing texts without the separate formal vote. The standardisation request runs to 28 February 2027.
Meanwhile, the July regulation adds to Article 40 a paragraph requiring the Commission to request “without undue delay” the drafting of the standardisation deliverables, specifying that they must be based on the standards already published in the Official Journal which confer the presumption of conformity. Standards which, for this subject matter, do not appear to exist.
One detail, to close the paragraph. Those technical standards are private documents, covered by copyright and sold for a price. To establish that, when they ground conformity with a law, they form part of Union law and must be freely accessible, a judgment of the Court of Justice was needed. In March 2024.
Rome, meanwhile
Italy moved earlier than almost everyone. Law 132 of September 2025, in force since 10 October, states principles — among them, reserving the decision to the judge — regulates health care, work, the professions and public administration, entrusts supervision to the Agency for Digital Italy and to the cybersecurity agency, and opens the criminal chapter: the new Article 612-quater punishes with imprisonment from one to five years whoever causes unjust harm by disseminating generated or altered content capable of misleading, and a new general aggravating circumstance targets the use of these systems as an insidious means.
The enabling delegation expires on 10 October 2026. Two draft decrees reached Parliament in July, after the preliminary examination of 10 June; they are not final texts. The first redesigns the governance and performs a course correction: the Data Protection Authority, kept out last year, comes back in as supervisory authority for high-risk systems in justice, law enforcement, immigration, borders and democratic processes; on 14 July it gave a favourable opinion with four conditions. The second regulates artificial intelligence in police activity, biometrics and responsibilities, and introduces into the Criminal Code Article 437-bis: failure to adopt safety measures in artificial-intelligence systems, where a concrete danger to life, public safety or the security of the State results.
It is a tracing of Article 437, the one that punishes whoever omits to install equipment intended to prevent disasters or accidents. Born for the factory, where safety measures are things: a guard, a switch, a signal. Transferred to the model, the tracing poses a question that for now remains without a written answer: what are the safety measures of an artificial-intelligence system, and who establishes them. The decree does not list them. The European technical standards are not there yet. We will come back to this, because it deserves a part of its own.
On the relationship between the two levels, the European regulation now says something that is worth bringing to Italy: recital 15 recognises that the same conduct may fall under national criminal law, that the prohibitions do not preclude criminal prosecution, and that the States must ensure respect for ne bis in idem. It is the first time that European artificial-intelligence law admits that, behind the administrative apparatus of conformity, the criminal law is already in the field.
One last note, and the chapter closes. While in Rome the criminal law advances, in Brussels the civil law retreats: the proposed directive on civil liability for artificial intelligence, presented in September 2022, was listed among the withdrawals in the work programme of 11 February 2025 and the withdrawal was formalised in the Official Journal on 6 October. What remains is the new directive on defective products, to be transposed by 9 December 2026, which now includes software: it covers death, injury, property destroyed. Not the job denied, not the credit refused.
The world, while Europe waits
Outside the Union things are moving in different and faster directions.
In Washington there is no federal law, and the executive is working to make sure the States do not have one either: an executive order of December 2025 set up at the Department of Justice a task force to challenge State laws; the Colorado law, cited by name, was suspended in April and replaced in May. Beijing has regulated since 2023 and since September 2025 requires the labelling of synthetic content: an effective rule, born for the control of content more than for the protection of rights. For a year the United Nations has had a scientific panel of forty experts and a global dialogue, whose first session was held in Geneva in July. The atomic energy agency, constantly invoked as a model, has inspectors. This structure has an annual report.
And then there is the episode which, read through a lawyer’s eyes, says more than all the others. On 9 June 2026 an American company placed two new models on the market. On 12 June the Secretary of Commerce ordered it, with an export-control directive, to suspend access for any foreign national anywhere in the world, including its own employees; unable to verify nationality in real time, the company switched everything off. Nineteen days. The return was negotiated: a new safety filter, tested by the researchers of the government standards centre, and for the most capable model access reserved to authorised organisations.
In 2026 the only act that really stopped a frontier model came neither from a law on artificial intelligence, nor from a supervisory authority, nor from a judge. It came from export controls: three days’ notice, planetary reach, no written reasons, no adversarial hearing. Effectiveness and arbitrariness in the same gesture.
Three questions, for now merely noted down
This part set out to count, not to judge. But three questions remain on the table, and we may as well write them down now, so the reader keeps an eye on them with us.
The first: who really writes the rule. If the content of the obligations sits in the technical standards, and the technical standards are written by private committees with the regulated industry inside them, the normative substance no longer lives where the legislator’s signature lives.
The second: who verifies. If for almost all of Annex III conformity is declared by the producer, the assessment is not a check: it is a compliance exercise. And compliance exercises have a market, which has in fact already been born — conferences, certifications, dashboards, heads of algorithmic governance — well ahead of the obligations that were supposed to make it necessary.
The third, which is the simplest and the least asked: how will we know whether it worked. The regulation prescribes documents to be produced, records to be kept, assessments to be preserved. There does not appear to be a single provision requiring the collection and publication of any data on outcomes: how much harm avoided, how many complaints upheld, how many decisions corrected after a challenge. Whoever has followed in these pages the story of the organisational models recognises the void, and knows where it leads.
The three questions have an answer, and it is not short. It will come in the next parts, with the texts in hand and with the strongest objections put in writing before our own.
For now the mark stays where it is: in an annex, with its grid and its minimum height. Five millimetres, waiting for a system to rest on. Machina scribit, homo signat: the machine writes, the man signs.
Note on sources
Regulation (EU) 2024/1689 of 13 June 2024. Regulation (EU) 2026/1744 of 8 July 2026, in the Official Journal of the European Union, L series, of 24 July 2026 (ELI: data.europa.eu/eli/reg/2026/1744/oj), read in the official Italian version: recitals 1 to 47 and Article 1, points 1 to 27; Article 1 from point 28, Articles 2 and 3 and the annexes were not reviewed. Provisions cited: new Articles 2(13), 4, 4a, 5(1), (1a) and (1b), 6, 40(2), 43(3), 48, 50 and 57; recitals 2, 7, 8, 11, 12, 15, 40, 41 and 46. State of technical standardisation: as at 12 August 2026 no reference to harmonised standards conferring the presumption of conformity under Article 40 appears published in the Official Journal; the state of the work of the CEN-CENELEC JTC 21 joint technical committee and the acceleration measures of October 2025 are reported according to the communications of those organisations and the publicly available reviews. Joint Opinion 1/2026 of the European Data Protection Board and the European Data Protection Supervisor, 20 January 2026, referred to in recital 47 and reported according to publicly available reviews. Court of Justice, 5 March 2024, Case C-588/21 P. Directive (EU) 2024/2853, to be transposed by 9 December 2026; proposal COM(2022) 496 of 28 September 2022, withdrawal announced with COM(2025) 45 of 11 February 2025 and formalised in the Official Journal, C series, of 6 October 2025. Italian law: Law no. 132 of 23 September 2025; draft legislative decrees approved in preliminary examination on 10 June 2026 and transmitted to Parliament, texts not final; opinion of the Data Protection Authority of 14 July 2026; Arts. 437 and 612-quater of the Criminal Code. The extra-European picture: United States executive order of 11 December 2025 and implementing acts of 2026; the Chinese rules on the labelling of synthetic content in force since 1 September 2025; United Nations acts of August 2025 and July 2026; the story of the suspension and restoration of the two models of June 2026 is reported according to the public statements of those involved. Online checks conducted up to 12 August 2026. The text distinguishes what is established, what is merely declared and what is inferred; the critical assessments here are stated and not argued, and will be developed in the following parts.
Get each part by email as well, so it does not depend on any social network: whitecollarcrimes.it
whitecollarcrimes
a collective byline · guarantor: Avv. Luca Santa Maria · 2 September 2026