Hello, this is Ryuta Hamamoto from TIMEWELL.
"EU AI Act goes fully live on 2 August 2026. That's what most executives thought they knew, and right now they are in full confusion." A legal officer at a major manufacturer sent me almost exactly that message.
That reading is not accurate. The amending act, Digital Omnibus on AI (Regulation (EU) 2026/1744), was adopted on 8 July 2026, published in the Official Journal on 24 July 2026 (OJ L 2026/1744) and entered into force on 27 July 2026. Under it, the highest-impact high-risk AI obligations now land on 2 December 2027 (Annex III type) and 2 August 2028 (Annex I type).
At the same time, Article 113(2), "It shall apply from 2 August 2026," was left unamended, and two months have now passed since that date. It turned out to be neither a non-event nor the day everything started. It was the day on which what starts and what does not finally separated. Since August, the EU has been finding out what companies are actually doing, one questionnaire and one official list at a time.
What the amendment deferred is narrow: Chapter III Sections 1, 2 and 3, excluding Article 6(5). Section 5 of the same chapter (harmonised standards, conformity assessment, CE marking and registration under Articles 40 to 49) was left where it was and has been running since 2 August. The split runs at section level, which is exactly where the confusion comes from.
In the 3 October 2026 update, I added what regulators and companies did after 2 August. After that comes the confirmed schedule and the concrete work it implies.
TL;DR
- Digital Omnibus is law: Regulation (EU) 2026/1744 (adopted 8 July 2026, OJ 24 July 2026, in force 27 July 2026). Annex III type high-risk AI applies from 2 December 2027; Annex I type from 2 August 2028
- What started on 2 August 2026 was Chapter IV (Article 50, transparency), Chapter III Section 5 (harmonised standards, conformity assessment, CE marking, registration), Chapter VI, Chapters VIII–XI, and Article 101 (the Commission's power to fine GPAI providers). Not the substantive high-risk AI duties
- GPAI model provider duties and the penalty provisions (Articles 99 and 100) have applied since 2 August 2025. Prohibited practices (Article 5) and AI literacy (Article 4) have applied since 2 February 2025
- Enforcement after 2 August began with information requests to more than 30 AI companies, none of them named. As of 3 October there is no official announcement of a fine
- On the company side, about 190 organisations signed the transparency code of practice. Meanwhile the high-risk classification guidelines are still a draft, no AI Act standard has been cited in the Official Journal, and only nine Member States have a final market surveillance designation
- Japanese companies should start with extraterritorial scope assessment, Authorised Representative appointment, and internal governance on an ISO/IEC 42001 backbone, and close out Article 50 implementation before year-end
What happened after 2 August 2026
Two months on, here is the picture as of 3 October. Enforcement has started, but nobody has been fined. What moved was the Commission's information-gathering and, on the company side, signatures, labelling and disclosure. The standards and guidelines that high-risk AI depends on are still not in place. The main events, in date order:
| Date | What happened | Source |
|---|---|---|
| 20 July 2026 | Commission publishes guidelines on the Article 50 transparency obligations | European Commission1 |
| 29 July 2026 | CEN-CENELEC announces approval of EN 18286 (quality management system), the first European standard supporting the AI Act. Citation in the Official Journal expected later in 2026 | CEN-CENELEC2 |
| 31 July 2026 | Commission announces it will start enforcing on 2 August. About 190 organisations have signed the transparency code of practice | European Commission34 |
| 2 August 2026 | General application date. Article 50 and the AI Office's power to fine GPAI model providers (Article 101) start to apply | Article 113 |
| 1 September 2026 | Commission confirms its first requests for information went to more than 30 AI companies | AFP5 |
| 7 September 2026 | Official list of national market surveillance single points of contact updated. Nine designations are final | European Commission6 |
| 14 September 2026 | Bitkom, Germany's digital industry association, publishes a company survey on the AI Act | Bitkom7 |
| 16–17 September 2026 | State of the Union sets out plans to convene frontier AI labs. The AI Board holds its ninth meeting the next day | European Commission89 |
| 29 September 2026 | Targeted copyright consultation opens, covering the use of protected content in AI (open until 3 November) | European Commission10 |
Regulators: enforcement began with questionnaires
What changed on 2 August is that the Commission's AI Office can now actually request documents from general-purpose AI model providers, evaluate their models, ask for corrective measures and impose fines. The Commission's 31 July press release also set out who enforces what. The AI Office handles AI systems offered by the same provider as the underlying general-purpose model, plus AI systems built into very large online platforms and search engines designated under the Digital Services Act. National authorities cover other AI systems, and the European Data Protection Supervisor covers AI used by EU institutions3. The same release announced that the Scientific Panel of 60 independent experts had held its first meeting and that Professor Alessandro Abate of the University of Oxford had been appointed Lead Scientific Adviser. A complaint tool, a whistleblower tool and a dedicated channel for downstream providers building on general-purpose models went live at the same time.
The first visible enforcement step was the round of information requests confirmed on 1 September. According to AFP, they went to more than 30 companies, and Commission spokesperson Thomas Regnier said the questions focused mainly on model safety and security and on copyright5. The Commission has not named the recipients. Press coverage linked the timing to incidents this summer that several AI companies disclosed themselves, involving unexpected outside access during testing, but there is no official word on who received a letter. I am not going to guess. The part worth remembering is that Article 101 also allows fines when a provider fails to answer a request or supplies incorrect, incomplete or misleading information11. The quality of the reply is already on the line at the questionnaire stage.
It has also become clearer what is still missing. The high-risk classification guidelines were due under the Act by 2 February 2026; a draft came out on 19 May, and as of 3 October I could not find a final version12. On standards, EN 18286 for quality management systems has been approved and published, but a presumption of conformity only attaches once the Commission cites it in the Official Journal, which CEN-CENELEC expects later in 20262. National set-ups are still incomplete as well. Counting the official list updated on 7 September, nine Member States have a final single point of contact for market surveillance, twelve have a national designation awaiting final adoption, and six have no entry at all6. The Commission itself wrote that effective enforcement will depend on Member States properly designating their authorities and giving them adequate resources3.
September moved the conversation on. The State of the Union initiatives published on 16 September include a plan to bring together the main frontier labs and discuss how Europe can support industry efforts to "pace the frontier"8. I covered where that phrase comes from in my explainer on Amodei's "Pace the Frontier". At the AI Board's ninth meeting the next day, the agenda included the Commission's enforcement priorities, recent AI incidents, recommendations to support AI literacy, and support for the transparency rules9. On 29 September a targeted copyright consultation opened, including the use of protected content in AI10. The separate Digital Omnibus that amends the GDPR and other data rules was, as of 20 September, still at committee stage in the European Parliament, with no trilogues yet13. The question of AI training and "legitimate interest" sits in that file, so track it separately from the AI Act amendment.
On fines: as of 3 October, I could find no official announcement of the Commission imposing one under the AI Act. You will see articles online claiming huge penalties have already been handed out. The AI Office's enforcement powers over general-purpose AI only began on 2 August, and the substantive high-risk duties do not apply yet. Keep numbers you cannot trace to a source out of your internal briefings.
Companies: signatures, labels, and stated concerns
The most visible change on the company side is signing codes of practice and marking generated output.
The Code of Practice on Transparency of AI-generated Content was finalised on 10 June, and about 190 organisations had signed by the end of July. The official list updated on 24 September shows 95 signatories to Section 1 (providers) and 192 to Section 2 (deployers). The Commission's examples for Section 1 were Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, OpenAI and Synthesia; for Section 2, Bulgari, Fastweb, Getty Images, Iberdrola, Lenovo and Lufthansa. According to the Commission, about half the signatories are small and recently founded companies4. I did not expect that. This is not only a big-tech exercise.
Some of the large signatories attached concerns to their signatures. Google announced its signature in a 24 July blog post, describing its SynthID watermarking technology, its push to adopt the C2PA content-provenance standard, and partnerships with Apple, ElevenLabs, Kakao, NVIDIA and OpenAI on interoperable watermarking with SynthID. The same post said that adding regulatory complexity while technical solutions are still evolving could run against Europe's competitiveness and simplification goals, and that too many labels could confuse people14. Meta announced its signature on 28 July under the name of its VP of Public Policy for Europe, saying it wants to avoid a growing array of different labels and disclosures that overwhelm people while adding regulatory complexity for providers15. Meta does not appear on the GPAI Code of Practice signatory list, but it is listed in Section 1 of the transparency code. The lists make one thing plain: the two codes are separate processes, and companies sign them separately.
On implementation, Anthropic's help centre states that it signed the transparency code and that new models launched in the EU on or after 2 August 2026 support machine-readable marking from day one: an imperceptible watermark in generated text and signed C2PA provenance metadata on supported files, across its API and cloud channels and in all regions, not just Europe16. I went through how text watermarking works, and how providers differ, in How AI Text Watermarks Work.
On the GPAI side, the official list updated on 31 July shows 21 signatories, including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI, with xAI signed up to the Safety and Security chapter only17. At the fourth Signatory Taskforce meeting (held 17 July, published 3 August), the AI Office said that "marginal-risk" clauses in safety frameworks, under which a provider might follow a competitor that deploys unsafe models, could be invoked only in exceptional circumstances and with appropriate safeguards. On copyright, the meeting covered the commitment to publish information about web crawlers and their robots.txt handling, and to notify rightsholders automatically when it changes18.
There are numbers on how companies feel, too. Bitkom's survey published on 14 September (603 German companies with 20 or more employees, telephone interviews from mid-July to mid-August 2026) found 57% now use AI. On the AI Act, 37% expect to be affected as users (23% a year earlier), 34% are still checking, 18% see themselves as unaffected and 8% have not looked at it. Among those affected or checking, 89% expect high implementation effort and 51% very high. Sixty-six percent think the AI Act brings more disadvantages than advantages for German companies, up from 56%7. Bitkom's president said the Act was meant to build trust but is creating widespread uncertainty, and called for legal clarity and workable standards quickly. That is an industry association's view and should be read as such, but the jump in companies that see themselves as affected says a lot about the months around 2 August.
What Japanese companies should do before year-end
Japanese names do appear on the transparency code list. NTT, Inc is listed in both Section 1 and Section 2, and Dentsu in Section 24. I could not find statements from either company explaining the scope or reasons, so I will stop at the fact that they are listed. For Japanese companies in Europe more broadly, JETRO's FY2025 survey (published December 2025) found 46.5% named "AI regulation" among the digital policies they watch, up 19.6 points from the year before19. Post-August numbers will have to wait for the next round.
From all this, I would narrow the work down to three things.
First, get Article 50 to "done" before moving on. If you offer chatbots or generative AI to EU users, or your EU subsidiary publishes deepfakes or AI-generated text on matters of public interest, the obligations already apply. They apply whether or not you sign the code; if you do not sign, you will need to show equivalent, adequate means of marking and labelling AI-generated content1. For generative AI placed on the market before 2 August, the machine-readable marking deadline is 2 December. Work out whether you are the provider of the generated output or a deployer using someone else's model. Deployers should check deepfake labelling and published text; providers should check how they mark output.
Second, update what you ask your AI vendors. Have they signed the GPAI code, the transparency code, both, or neither? Do they publish a training content summary? Is output marked in a machine-readable way? All of this can be checked against the official lists and the vendors' own pages. With the AI Office's information requests aimed at model providers, downstream users will increasingly be asked what documentation they can get from those providers.
Third, revisit AI literacy. The amending act rewrote Article 4 from an obligation to ensure a sufficient level of AI literacy into one to take measures to support its development20. The wording is softer, but at the September AI Board meeting the Commission reported on its work on recommendations to support AI literacy9. Keeping role-by-role training records, so you can explain why your measures are adequate, is the practical answer. If you want a baseline first, the AI literacy check is a place to start.
On high-risk AI, my view is that you can hold off on the fine detail until the final classification guidelines and the Official Journal citations of standards arrive. Working back from 2 December 2027, though, the AI system inventory, risk classification and the search for an EU representative are things I would start this year. What the past two months showed is that the EU is not writing rules and walking away. It is checking reality bit by bit, through questionnaires and public lists. How prepared you are will eventually become visible from the outside, in the form of how you answer.
What Regulation (EU) 2026/1744 changed
Honestly, I read this amendment as an industry last-minute compromise. Annex III type high-risk duties were originally designed to fire on 2 August 2026, but CEN/CENELEC JTC 21 draft harmonised standards were still unsettled in Q2 2026 and guidance was incomplete. In practice: "we can impose duties, but no one knows what compliance means." EN 18286 for quality management was approved in July, but as of 3 October neither its Official Journal citation nor the final high-risk classification guidelines had appeared (see the section above).
The main changes to Article 113:
| Subject | Old schedule | After the amendment |
|---|---|---|
| Annex III type (stand-alone) high-risk AI — Art. 6(2) | 2 August 2026 | 2 December 2027 (Chapter III Sections 1–3, plus Arts. 22, 25, 26, 27) |
| Annex I type (product-embedded) high-risk AI — Art. 6(1) | 2 August 2027 | 2 August 2028 |
| General application date — Art. 113(2) | 2 August 2026 | 2 August 2026 (unamended) |
| Chapter IV (Art. 50, transparency) | 2 August 2026 | 2 August 2026 (unchanged) |
| Art. 101 (Commission's power to fine GPAI providers) | — | Applies from 2 August 2026 |
| New prohibited practices — Art. 5(1)(ba) non-consensual sexual deepfakes, (bb) CSAM generation, Art. 5(1a) and (1b) | New | Apply from 2 December 2026 |
| New Art. 111(4) — providers of synthetic-content AI placed on the market before 2 August 2026 | New | Must comply with Art. 50(2) by 2 December 2026 |
| Transitional rule for high-risk AI already on the market — Art. 111(2) | Fixed cut-off date | Only if significant changes in their designs are made after Chapter III applies |
The 28 April 2026 trilogue stall was not mainly about "moving the dates as such." The fight was conformity-assessment architecture for Annex I type systems: avoiding double certification against sector rules such as medical devices and machinery. As Bird & Bird noted, the deal was less about the headline high-risk dates and more about the organisational design of conformity assessment. Industry burden relief21.
Note that the amendment is already in force; this is not a "pending formal adoption" situation. On 27 July 2026, the date the amending act entered into force, Articles 102–110 of the AI Act (which amend other EU instruments) also became applicable.
AI Security training, taken seriously
A 2-day intensive course fully aligned with OWASP, NIST, ISO/IEC 42001, and METI. Take it as executives, practitioners, or both.
What did not move is the real core
Many Japanese executives misread this part. What moved later is the substantive high-risk AI duties. These five did not.
First, prohibited practices and AI literacy. Chapter I (general provisions, definitions, Article 4 AI literacy) and Chapter II (Article 5, prohibited AI practices) have applied since 2 February 2025. Emotion inference at the workplace (Article 5(1)(f)) is already banned, and it sits in the top penalty tier, 7% of worldwide turnover. Companies that filed this under "later" have in fact been under the obligation since 2 February 2025.
Second, GPAI (general-purpose AI) model provider duties. Chapter V has applied since 2 August 2025: technical documentation, sufficiently detailed training-data summaries, and EU copyright-compliance policies. Models classified as "systemic risk GPAI" also need adversarial testing, incident reporting, cybersecurity, and energy-use tracking.
Third, penalties. Within Chapter XII, Articles 99 and 100 have applied since 2 August 2025. The subtlety worth knowing: within that same chapter, Article 101 alone was excluded and has applied from 2 August 2026, the Commission's power to fine GPAI providers. It is neither "all penalties from 2025" nor "all enforcement from August 2026"; it splits at article level.
Fourth, Article 50 transparency. Chapter IV, covering machine-readable marking of AI-generated content and related duties, is not listed in any of the exceptions in Article 113(3) as amended. It applies on the general application date, 2 August 2026. What Digital Omnibus changed in Article 50 is paragraph (7) (codes of practice); the substantive duties in paragraphs (1)–(6) were left unamended. You will see this described as "brought forward"; correctly stated, it stayed where it was.
Fifth, the two things that fire on 2 December 2026. The new prohibited practices (Article 5(1)(ba) non-consensual sexual deepfakes, (bb) CSAM generation, and Article 5(1a) and (1b)), and the new Article 111(4): providers of synthetic-content generation AI placed on the market before 2 August 2026 must comply with Article 50(2) by that date. Article 50 itself does not start on 2 December.
So the AI Act is running with application dates staggered article by article. Reading it as "everything starts on 2 August" and reading it as "high-risk moved to 2027, so there is nothing to do" are both wrong in practice.
Three patterns that catch Japanese companies: self-diagnosis starting points
Article 2 sets extraterritorial reach. Japanese companies typically fall into three patterns.
Pattern 1: Direct product offering on the EU market. An automaker selling driver-assistance AI vehicles in the EU, or a medtech firm offering AI diagnostic software with CE marking. Both are Annex I product-embedded high-risk AI. CMS's extraterritorial guide is explicit: a Japanese automaker selling AI-assisted braking vehicles into a global market that includes the EU is in scope22.
Pattern 2: SaaS/API whose outputs are used in the EU. A Japanese hiring-screen AI sold to EU clients is Annex III (employment). Credit scoring and education assessment AI are the same. "EU subsidiary uses HQ AI" also fits if outputs are generated or used in the EU.
Pattern 3: Deployer role at an EU subsidiary/branch. If an EU entity uses HQ HR AI for employee evaluation, Article 26 Deployer duties apply: FRIA, log retention, and appointment of human oversight staff.
Common to all three is Article 22's Authorised Representative in the EU. Written mandate required; for non-EU providers it is the practical EU doorway. In my experience, more than half of Japanese companies still do not treat this as a "must-prepare" item.
Strategy: ISO/IEC 42001 as the common backbone
Now implementation. With high-risk dates fixed at 2 December 2027 / 2 August 2028, the most rational move is to introduce ISO/IEC 42001 (AIMS) as the common backbone for internal AI governance.
ISO/IEC 42001 was issued in December 2023 as the world's first AI management-system standard. You will find vendor material quoting a "coverage percentage" against EU AI Act high-risk requirements; there is no official mapping table behind those numbers, so this article does not quote one. Where the two overlap in practice, article by article:
| EU AI Act article | ISO/IEC 42001 mapping |
|---|---|
| Art. 9 (risk management) | Clause 6.1 |
| Art. 10 (data governance) | Annex A.7, A.8 |
| Art. 11 (technical documentation) | Clause 7.5 |
| Art. 12 (automatic logging) | Annex A.6 |
| Art. 13 (transparency / information) | Annex A.4, A.9 |
| Art. 14 (human oversight) | Annex A.5 |
| Art. 15 (accuracy / robustness) | Annex A.7, A.10 |
| Art. 17 (quality management) | Clause 8 |
For Article 17 specifically, the AI Act-specific European standard EN 18286 was approved in July. Once it is cited in the Official Journal, that is the standard that gives a presumption of conformity, not ISO/IEC 420012. Keep the two roles separate in your planning.
What ISO/IEC 42001 will never close is the EU-specific gap:
- FRIA (Art. 27): no counterpart concept in ISO 42001
- Conformity assessment (Notified Body): third-party certification for some high-risk AI
- EU database registration: EU-only
- Authorised Representative: EU-only
- CE-marking integration: product-embedded only
I describe this as "ISO 42001 as the spine, EU-specific requirements as orthotics layered on top." If a Japanese company must also cover the US (NIST AI RMF), Japan (AI Business Guidelines), and Asia (Korea AI Basic Act, Singapore AI Verify), a shared ISO 42001 language makes market overlays far cheaper.
"Only build for the EU AI Act" almost always becomes regret later.
2026–2028 roadmap
Working backwards from the confirmed dates:
- 2 August 2026 (already applies): General application date. Chapter IV (Article 50), Chapter III Section 5 (harmonised standards, conformity assessment, CE marking, registration), Chapter VI, Chapters VIII–XI, and Article 101 started applying. If the work you meant to finish before this date (extraterritorial scope across the three patterns, an internal AI inventory, and checking your Article 50 implementation) is still open, it comes first
- October–December 2026: Article 50 implementation and records, a check of your AI vendors' code-of-practice status and disclosures, and an Authorised Representative shortlist
- 2 December 2026: New prohibited practices (non-consensual sexual deepfakes, CSAM generation) start applying. Separately, synthetic-content generation AI placed on the market before 2 August 2026 must comply with Article 50(2) by this date under Article 111(4)
- First half of 2027: ISO/IEC 42001 certification path (typically 6–12 months). Watch for the final high-risk classification guidelines and Official Journal citations of standards, and fold them into your internal classification criteria
- 2 August 2027: Article 111(3), the compliance deadline for GPAI models placed on the market before 2 August 2025. It is also the date by which each Member State's national AI regulatory sandbox must be operational20
- Second half of 2027: FRIA ahead of the Annex III type application date, conformity prep, technical documentation, EU database registration
- 2 December 2027: Chapter III Sections 1–3 apply to Annex III type high-risk AI (Art. 6(2)). Articles 22, 25, 26 and 27 start the same day
- 2 August 2028: The same applies to Annex I type (product-embedded, Art. 6(1)) high-risk AI
- 2 August 2030 / 31 December 2030: Article 111(2) (use by public authorities) / Article 111(1) (Annex X large-scale IT systems)
Move on this calendar and you avoid last-minute FRIA and Notified Body panic just before the December 2027 date. Decide "we can start mid-2027" and certification-body capacity becomes a real risk.
One more point on legacy systems: Article 111(2) was amended so that high-risk AI already on the market is caught only where significant changes in their designs are made after Chapter III starts applying. The reference point is no longer a fixed 2 August 2026; it tracks the Chapter III application dates (2 December 2027 / 2 August 2028). Build your product-modification plans around that linkage.
For implementation steps at the 2 August 2026 milestone, see EU AI Act: August 2026 application and five steps. Read with this piece: strategy here, implementation there.
Where WARP SECURITY fits
TIMEWELL's WARP SECURITY covers the EU AI Act, Digital Omnibus, ISO/IEC 42001 crosswalk, and GPAI Code of Practice alongside OWASP LLM Top 10, NIST AI RMF, and METI AI Business Guidelines. A two-day, leadership-and-practitioner curriculum.
Executive track: how to use the runway until the Annex III type date of 2 December 2027; how to choose an Authorised Representative; who owns FRIA decision rights. Decision drills.
Practitioner track: map the 38 Annex A controls in ISO/IEC 42001 to EU AI Act articles and run a gap analysis on your own AI systems, including RAG log design and machine-readable marking for Article 50.
One of five tabletop incidents is "AI Office document request to a GPAI provider. First 72 hours." It forces legal, IT, comms, and leadership to set RACI in the same room. In September the Commission did in fact send information requests to more than 30 companies. You cannot learn that from reading alone.
How to use the phrase "general application date"
2 August 2026 is the general application date set by Article 113(2). It is not "full application" and not "fully in force." From that date the Commission can exercise its supervision and enforcement powers over GPAI providers (Article 101), with GPAI-related sanctions of up to €15 million or 3% of worldwide turnover, whichever is higher (Regulatory framework on AI (European Commission)).
The substantive high-risk AI duties, however, did not start on that date, and a striking amount of material, inside and outside Japan, still blurs the two. When you write an internal briefing, separate "what started on 2 August 2026," "what starts on 2 December 2026," and "what starts on 2 December 2027 / 2 August 2028."
Domestically, the 2026 Personal Information Protection Act amendment, including AI training data use, is also moving. Dual EU and Japan readiness is more necessary, not less.
Key takeaways
- Digital Omnibus is law as Regulation (EU) 2026/1744. High-risk AI applies from 2 December 2027 (Annex III type) and 2 August 2028 (Annex I type)
- What started on 2 August 2026 was Chapter IV (Article 50, transparency), Chapter III Section 5, Chapter VI, Chapters VIII–XI and Article 101, not the substantive high-risk duties
- Article 50 was not "brought forward to 2 December 2026"; it stayed on 2 August 2026. What fires on 2 December 2026 is the new prohibited practices plus the Article 111(4) deadline for existing systems
- Enforcement after 2 August opened with information requests to more than 30 companies. None were named, and as of 3 October no fine has been announced
- Companies moved by signing codes and marking output. Google and Meta signed while publicly flagging concerns about label proliferation and regulatory complexity
- For Japanese companies: finish Article 50, update vendor due diligence and document AI literacy training this year, then move on to scope assessment, an Authorised Representative and an ISO/IEC 42001 gap analysis
Treat the EU AI Act as a wall of "regulation," or as a chance to align internal AI governance with international standards. Post-2028 competitiveness will fork here. Companies that read the later dates as "grace" and those that read them as "last window before standards harden" will not close that gap in three years. Judging by the past two months, the EU will not rush, but it will keep asking more questions. If you want help working out which obligations your AI use actually triggers, you can talk to us about WARP.
References
- Regulation (EU) 2024/1689 (AI Act), Article 113 (application dates) and Article 111 (transitional provisions): text of the Regulation
- Regulation (EU) 2026/1744 (Digital Omnibus on AI, amending act): adopted 8 July 2026; OJ L 2026/1744, published 24 July 2026; in force 27 July 2026
- Regulatory framework on AI - European Commission
- AI Act Service Desk - Implementation Timeline
- AI Act Omnibus: What Just Happened and What Comes Next - IAPP
Footnotes
-
European Commission, "Guidelines on transparency obligations for providers and deployers of AI systems" (20 July 2026); see also the overview page ↩ ↩2
-
CEN-CENELEC, "First Standard Approved under the AI Act" (29 July 2026) ↩ ↩2 ↩3
-
European Commission, "Commission starts enforcing AI Act rules and new transparency requirements on 2 August," IP/26/1714 (31 July 2026) ↩ ↩2 ↩3
-
European Commission, "Strong backing for the Code of Practice on Transparency of AI-generated Content" (published 31 July 2026; signatory list updated 24 September 2026) ↩ ↩2 ↩3
-
AFP, "AI: EU seeks information from dozens of companies under its new powers" (via FashionNetwork, 1 September 2026) ↩ ↩2
-
European Commission, "Market Surveillance Authorities under the AI Act" (updated 7 September 2026; entries marked * are awaiting final adoption of the national designation) ↩ ↩2
-
Bitkom, "Erstmals nutzt die Mehrheit der Unternehmen KI" (14 September 2026; 603 companies with 20+ employees, telephone survey in calendar weeks 28–33 of 2026) ↩ ↩2
-
European Commission, "State of the Union 2026 – Quotes and initiatives" (16 September 2026) ↩ ↩2
-
European Commission, "AI Board holds its ninth meeting" (18 September 2026) ↩ ↩2 ↩3
-
European Commission, "Commission seeks feedback on challenges and way forward in the area of effect of technology on copyright" (29 September 2026) ↩ ↩2
-
AI Act Service Desk (European Commission), "Article 101: Fines for providers of general-purpose AI models" ↩
-
European Commission, "Draft Commission guidelines on the classification of high-risk AI systems" (19 May 2026) ↩
-
European Parliament, "Legislative Train: The Digital Omnibus Regulation Proposal" (updated 20 September 2026) ↩
-
Google, "Google is signing the EU AI Act Code of Practice on Transparency of AI-Generated Content" (24 July 2026) ↩
-
Meta, "Meta is Signing the EU AI Act Code of Practice on Transparency of AI-Generated Content" (28 July 2026) ↩
-
Anthropic, "How Claude marks AI-generated content" (Claude Help Center) ↩
-
European Commission, "The General-Purpose AI Code of Practice" (signatory list updated 31 July 2026) ↩
-
European Commission, "Fourth GPAI Signatory Taskforce meeting" (3 August 2026) ↩
-
JETRO, "AI規制への注目度が大きく上昇、AIの活用率も前年から倍増(日本、EU)" (December 2025; FY2025 Survey on Business Conditions of Japanese Companies Operating Overseas, Europe edition; Japanese) ↩
-
Regulation (EU) 2026/1744 (Digital Omnibus on AI), amendments to Article 4 and Article 57(1) ↩ ↩2
-
Guide to the EU AI Act for Businesses Outside the EU - CMS ↩






