2026.10.04 2026.10.04

Update the Law Before You Memorise It: Japan AI Act, Copyright, and the EU AI Act

swiftwand

On the technical side, definitions rarely shift within a few years. The law and ethics domain is different. Since Syllabus 2024 was written, Japan promulgated and brought into force an AI Act and adopted an Artificial Intelligence Basic Plan by cabinet decision twice. In the EU, a regulation moved the application dates of the AI Act, and the transparency obligations started to bite. Materials cannot keep up with that pace. Nor can the AI: asked when the obligations for high-risk AI apply in the EU, it answered 2 August 2026. That date was pushed to 2 December 2027 by a regulation adopted in June 2026 and published in the Official Journal on 24 July.

This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.

This article covers the way to handle the law and ethics domain of the G-test: update to the current rules before memorising anything. Hand the AI the syllabus mid-level items and the primary sources, have it build a correspondence table, and check the dates and article numbers yourself. The substance below rests on official material from the Cabinet Office, the Agency for Cultural Affairs and the EU as of 21 September 2026, and every date carries the day it was taken. This is not legal advice; it is an exam-preparation summary, and individual cases belong with a professional.

What the law and ethics domain asks
Why update before memorising
The method: hand over the sources, get a correspondence table
Japan as it stands: the AI Act and the Basic Plan
Copyright: the Agency for Cultural Affairs paper and article 30-4
The EU AI Act: what was postponed and what was not
With and without primary sources, side by side
How to memorise what you updated
Traps
Summary
Sources

忍者AdMax

What the law and ethics domain asks

The law and ethics domain of Syllabus 2024 revision 1.4 has two major items across 18 mid-level items. Law and contracts around AI holds seven: personal information protection law, copyright, patents, unfair competition, antitrust, AI development outsourcing contracts and AI service provision contracts. AI ethics and AI governance holds eleven: domestic and international guidelines, privacy, fairness, safety and security, misuse, transparency, democracy, environmental protection, labour policy, other important values, and AI governance.

Look at the keywords and you find few specific statute or regulation names. Under domestic and international guidelines sit AI ethics, AI governance, value principles, hard law, soft law and the risk-based approach; neither the EU AI Act nor the Japanese AI Act appears as a proper noun. Under copyright sits the purpose of article 30-4; under contracts, the contract guideline on the use of AI and data; under privacy, the Camera Image Utilisation Guidebook.

So what is examined is not whether you have memorised an article number or a date, but whether you understand the shapes of thinking: the difference between hard law and soft law, the idea of varying the strength of regulation with risk, the split between the training stage and the generation stage in copyright. Those shapes do not move. The examples that illustrate them do.

Why update before memorising

Three reasons.

First, the knowledge in the AI has not caught up. This site asked Claude (claude-sonnet-5, bare, with no tools and no material supplied) six questions about dates as of September 2026. It answered the promulgation of the AI Act (4 June 2025) and the publication of the Agency for Cultural Affairs paper (15 March 2024) correctly. But it gave 2 August 2026 for the application of the Annex III high-risk obligations, and on the EU Digital Omnibus simplifying the AI regulation it said that as of its knowledge cutoff the proposal had not been confirmed as adopted. On the cabinet decision date of the Basic Plan it said it was not certain. It annotated the answers itself with needs checking. Skip the annotation and you memorise the pre-postponement date.

Second, the materials share the constraint. The third edition of the official textbook was published in May 2024, before the AI Act was promulgated. Problem-set and course explanations are written against the rules as they stood. The point is not that the materials are old; it is that the rules are fast.

Third, cutting distractors requires knowing what is current. The exam stays inside the syllabus, but four options come with plausible wrong answers. When an old formulation and a new one both appear among the choices, you cannot eliminate without knowing which is in force.

That AI goes wrong on proper nouns in the legal domain also showed up in our measurement of generated questions: of 30 written without the syllabus, the two with wrong answers or facts were both on guideline proper nouns, a method name and an owning ministry.

The method: hand over the sources, get a correspondence table

Four steps.

Step one: list the mid-level items. From the 18 in the law and ethics domain, pick the ones where changes in the rules are likely to matter. We picked five: copyright (law 2), domestic and international guidelines (ethics 8), misuse (ethics 12), transparency (ethics 13) and AI governance (ethics 18). Personal information protection and patent law have not moved in their skeleton, so they rank low for updating.

Step two: gather the primary sources. For Japan: the Cabinet Office (the AI Act, the Basic Plan), the Agency for Cultural Affairs (AI and copyright), the Personal Information Protection Commission, and METI and MIC (guidelines). For the EU: EUR-Lex for the text of regulations, press releases from the Council of the European Union and the European Commission, and the AI Act Explorer for reading article by article (the Explorer is a viewer; the primary source is the text on EUR-Lex). Commentary and law-firm summaries are useful as an entry point but are not the basis for the table.

Step three: have the AI build the table. Hand it the gist of what you collected, including dates, regulation numbers and article numbers, and instruct it to produce, on the basis of this material alone, a correspondence table of what to know under the current rules for each mid-level item, dated, adding nothing not in the material. The table we produced this way carried the dates, regulation numbers and article numbers exactly as the material had them for all five items, with nothing added from outside. Same model as the bare answers above. The difference is whether there is material to hand.

Step four: verify the dates and article numbers yourself. Go back to the primary source for every cell. Even when the AI wrote what the material said, you wrote the summary of the material, so check whether a branch was lost at the summarising stage. EU regulations especially differ article by article on what was postponed and what was not.

Handing over primary sources and asking for answers grounded in them is the core idea of retrieval-augmented generation. This hands-on book explains how that works under the hood.

Hands-On Large Language Models

USD 37.68 on Amazon.com (as of 2026/09/22)

Amazon

楽天市場

ポチップ

Japan as it stands: the AI Act and the Basic Plan

Three primary sources matter in Japan.

The AI Act. Formally the Act on the Promotion of Research, Development and Utilisation of Artificial Intelligence-Related Technologies. According to the Cabinet Office page it was promulgated on 4 June 2025, partially in force the same day, and fully in force on 1 September 2025. It places an Artificial Intelligence Strategy Headquarters in the Cabinet and gives statutory footing to an Artificial Intelligence Basic Plan. It runs to four chapters, general provisions, basic measures, the basic plan and the headquarters, and carries no penal provisions: a promotion statute. Being a statute, it nonetheless falls under hard law in the syllabus taxonomy. Soft law is the guidelines below.

The Artificial Intelligence Basic Plan. A statutory plan under the AI Act. The first was adopted by cabinet decision on 23 December 2025, and the second-phase plan on 14 July 2026. The second phase is subtitled Japan AX, stronger and richer, and sets out four basic directions: accelerating the utilisation of AI, strategically strengthening development capability, leading on AI governance, and continuous transformation towards an AI society. The summary states explicitly that the plan will be revised annually for the time being, so read it on the assumption that it moves every year. Leading on AI governance includes continuous review of the framework including the AI Act, a substantial strengthening of the AI Safety Institute, and holding an AI summit in Japan at an early date.

The guidelines. The AI Business Operator Guidelines, published jointly by the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry, are the consolidated guidance for businesses. The current version is revision 1.2, published 31 March 2026; revision 1.1 of 28 March 2025 is superseded. The text describes itself as non-binding soft law. The three documents it consolidated were the MIC draft AI development guidelines for international discussion, the AI Utilisation Guideline published on 9 August 2019, and the METI Governance Guidelines for Implementation of AI Principles version 1.1.

If the syllabus item on domestic and international guidelines asks you about the difference between hard law and soft law, you can answer with the structure: in Japan, a statute (the AI Act) is promotion-oriented hard law without penalties, while the concrete code of conduct is carried by guidelines, which are soft law.

Copyright: the Agency for Cultural Affairs paper and article 30-4

The copyright item asks about the purpose and outline of article 30-4, the thinking on copyright in AI outputs, and the need for licences. The primary sources are the Agency for Cultural Affairs paper Thinking on AI and Copyright (15 March 2024, Legal System Subcommittee of the Copyright Subdivision of the Council for Cultural Affairs) and the Checklist and Guidance on AI and Copyright (31 July 2024).

The spine of the paper is the separation of the training stage from the generation and use stage. Training is in principle permissible without permission under article 30-4, which covers uses not aimed at enjoying the thoughts or feelings expressed in a work, with exceptions such as training aimed at making a particular work output. The generation and use stage is judged by ordinary infringement analysis: reliance and similarity. Article 30-4 was consolidated in the 2018 amendment, and the paper sets out an interpretation without changing the text.

On updating: copyright is the area where the 2024 paper remains the current reference, and what is moving is the litigation. Cases over generative AI have been filed in Japan and are pending. A judgment would affect how the paper is applied, so look at the Agency page once before the exam to see whether a new document has appeared.

The EU AI Act: what was postponed and what was not

The EU AI Act, Regulation (EU) 2024/1689, is the clearest illustration of the syllabus keywords risk-based approach and hard law: it classifies AI systems by risk, imposes heavy obligations on the high-risk tier, and is legally binding.

What moved in 2026 was the timing. The Council of the European Union (the Council of ministers, distinct from the European Council of heads of state) gave final approval to the Digital Omnibus on AI on 29 June 2026, and it was published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744. That regulation pushed the obligations for high-risk AI systems listed in Annex III from 2 August 2026 to 2 December 2027, and for high-risk AI embedded in products under Annex I from 2 August 2027 to 2 August 2028.

Here is the branch you must not lose. What was postponed is the high-risk obligations. The transparency obligations in article 50 apply as planned from 2 August 2026. Article 50 requires providers of AI systems generating synthetic content to mark the outputs in a machine-readable format and make them detectable as artificially generated or manipulated, and requires deployers of deepfakes to disclose that the content has been artificially generated or manipulated. According to the Commission, providers of generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with the marking and detection duties of article 50(2), under the transitional provision in article 111(4). Alongside this, the Code of Practice on transparency for AI-generated content went through a first draft in December 2025 and a second in March 2026 before the final version was published on 10 June 2026.

Remember only that the AI Act was postponed and you lose the fact that the transparency rules are live. Split it three ways: high-risk to 2 December 2027, transparency from 2 August 2026, existing generative systems given until 2 December 2026.

With and without primary sources, side by side

To see the difference the material makes, we ran the same six questions bare and then as a correspondence table with the sources attached, on the same model and settings.

Item	Without primary sources	With primary sources
AI Act promulgation and entry into force	Promulgation and full entry correct. Partial entry not mentioned. Annotated needs checking	As in the material
Cabinet decision dates of the Basic Plan	Not certain	2025-12-23, 2026-07-14
Application of the EU high-risk obligations	2026-08-02 (the pre-postponement date)	2027-12-02
Application of EU article 50	2026-08-02	2026-08-02, existing systems given until 12-02
Adoption of the Omnibus	At the knowledge cutoff it was a proposal; cannot confirm	Regulation 2026/1744, Official Journal 2026-07-24
Agency for Cultural Affairs paper	2024-03-15	2024-03-15

What stands out in the bare answers is not that they were wrong but that the model wrote needs checking and not certain against the places where it was wrong. It declared its own knowledge cutoff as January 2026 and flagged that it might not know changes after that. That honesty only means something if the reader performs the check. That is the reason to build updating into the study routine as a step rather than an intention.

In the legal domain, asking the AI for grounding is less reliable than handing grounding to the AI.

Comparing answers with and without primary sources is a small evaluation. For a broader approach to evaluating model outputs, see this book on building applications with foundation models.

AI Engineering: Building Applications with Foundation Models

USD 52.40 on Amazon.com (as of 2026/09/22)

Amazon

楽天市場

ポチップ

How to memorise what you updated

Once the table exists, do not memorise the right-hand column wholesale. Memorise in pairs: postponed against not postponed (high-risk obligations against transparency), training stage against generation and use stage (article 30-4 against reliance and similarity), statute against guideline (the AI Act against the AI Business Operator Guidelines). Pairs decide which option to cut when an old formulation and a new one appear side by side.

Use the paired items as material for mutual explanation, and send what you get wrong to the mistake log and on to Anki. Legal cards get the rule and the exception on separate lines. For article 50 that is three lines: rule, applies from 2026-08-02; exception, systems on the market before 2 August have until 12-02; confused with, the high-risk obligations moved to 2027-12-02.

Traps

First, the granularity of the summary. Where a regulation differs article by article on what was postponed, a coarse summary drops the branch. Memorise application of the AI Act was postponed as one sentence and you lose the start of the transparency obligations. Any regime with branches gets summarised article by article.

Second, confusing owners and revisions. Guideline names resemble each other, and both the owning ministry and the revision change. The AI Utilisation Guideline (MIC, August 2019) and the AI Business Operator Guidelines (MIC and METI, currently revision 1.2 of 31 March 2026) are different documents, the latter being the current consolidated one. Our own measurement of generated questions produced exactly this error, marking the AI Utilisation Guideline as a METI document.

Third, over-updating. The exam asks about the shapes of thinking, so there is no point memorising every article of a regulation. Narrow to the mid-level items where the change affects the example, which is what step one was for.

Fourth, mistaking this for legal advice. What is written here is an arrangement for exam preparation. Real-world judgement on a specific matter belongs with a professional.

Summary

The law and ethics domain is where materials and AI both fall behind the rules, so update before memorising. Pick the mid-level items affected, gather primary sources, have the AI build a dated correspondence table on that material alone, and verify the dates and article numbers yourself. In Japan the anchors are the AI Act (promulgated 4 June 2025, fully in force 1 September 2025, no penalties), the Artificial Intelligence Basic Plan (cabinet decisions on 23 December 2025 and 14 July 2026, revised annually) and the AI Business Operator Guidelines (revision 1.2, 31 March 2026, self-described soft law). In the EU, high-risk obligations moved to 2 December 2027 while the article 50 transparency obligations have applied since 2 August 2026, with existing generative systems given until 2 December 2026.

That completes the six methods across 40 days. Fixing the scope, explaining, generating, sourcing, repeating and updating: the sixth is the only one that has to be redone every time the calendar moves, which is why it goes last and closest to the exam.