Japan: Rules for Licensing Data for AI Training

Japanese law lets AI developers copy works for data analysis without permission under Article 30-4 of the Copyright Act, but a proviso excludes uses that unreasonably prejudice the copyright owner, and that limit, together with dataset protection and privacy rules, is why data licensing exists in Japan.

This page covers copyright, dataset protection, personal data and the AI statute as of 2026-11-09, each with its citation and date, within AI training data licensing as a whole. The 2026 amendment to the Act on the Protection of Personal Information is enacted but not in force, and no licensing duties were found in Japan's AI statute.

What Japanese Law Decides When a Company Licenses Data for AI Training

Japanese law decides four things in a training data deal: whether works may be copied without a license, whether a licensed dataset stays protected after delivery, whether personal data in it may be provided, and whether any AI statute adds duties.

• Copyright: Article 30-4 of the Copyright Act and its proviso.
• Dataset protection: Article 2(6) (trade secrets) and Article 2(7) (shared data with limited access) of the Unfair Competition Prevention Act.
• Personal data: the Act on the Protection of Personal Information (Act No. 57 of 2003) and its 2026 amendment.
• AI statute: the AI Promotion Act, with no licensing duties found.

Japanese rules on employee data, call recording consent and database rights are not covered on this page yet.

Japanese Instruments That Apply to a Training Data Deal

Eight Japanese instruments and guidance documents shape a training data license; each row gives the citation, the date it took effect and whether it binds. The General Understanding is guidance, not law, and the APPI amendment is enacted but not in force.

The table lists the instruments in deal order, as of 2026-11-09.

Article 30-4 of the Japanese Copyright Act, Clause by Clause

Article 30-4 has a main clause, an item on information analysis and a proviso; each decides a different part of a deal.

The Non-Enjoyment Rule and Item (ii) Information Analysis

Article 30-4 permits exploiting a work "in any way and to the extent considered necessary" when the purpose is not to enjoy, or cause others to enjoy, the thoughts or sentiments expressed in it. The Copyright Act is Act No. 48 of 1970 (e-Gov 345AC0000000048); its current Article 30-4 was put in place by Act No. 30 of 25 May 2018 and has been in force since 2019-01-01 (supplementary provisions, Article 1). The English text commonly cited is the unofficial CRIC translation; the Japanese text on e-Gov was read on 2026-11-09.

Item (ii) covers information analysis: extracting, comparing, classifying or otherwise analysing language, sound, image and other elements from many works or large amounts of information. That is the item AI training relies on. The statute contains no opt-out or reservation mechanism, which sets Japan apart from the EU model of a machine-readable reservation and from the US fair use model; the comparison sits in copyright and AI training across jurisdictions.

The Proviso: Uses That Unreasonably Prejudice the Copyright Owner

The exception does not apply where the use would unreasonably prejudice the interests of the copyright owner in light of the nature or purpose of the work or the circumstances of its exploitation. The General Understanding on AI and Copyright in Japan, prepared by the Legal Subcommittee of the Copyright Subdivision of the Cultural Council, reads the proviso to cover reproducing works from paid, analysis-ready databases and circumventing robots.txt or ID-password measures. The Agency for Cultural Affairs published an English overview in May 2024; the document is non-binding guidance. On page 11 of the overview, collecting training data from websites known to distribute pirated content "should be strictly avoided".

Japan therefore does not "allow everything", and statements that copyright does not apply to AI training contradict the proviso. Robots.txt matters in Japan through that reading of the proviso, not as a statutory opt-out; the technical side is explained in AI crawlers and how to block them.

Where Article 30-4 Stops: Enjoyment Purposes, Output-Oriented Uses and Article 47-5

Article 30-4 stops where a purpose of enjoyment coexists with analysis, and the official reading names two training-adjacent uses that fall outside it.

• Fine-tuning aimed at intentional overfitting, that is, reproducing a creator's expression.
• Retrieval-augmented generation aimed at outputting expression from the source works.

Article 47-5(1) permits minor exploitation of works made public, incidental to computerised data processing that creates new information, and the guidance names it as a possible basis for some output uses where Article 30-4 does not apply.

What a License Still Adds When Japan Permits Analysis Without Permission

In Japan a license rarely buys the right to analyze a work; it buys access, delivery and assurances the exception does not give. Three points carry the licensing case:

1. Analysis-ready databases offered for payment sit inside the proviso, so copying them for training needs the rightholder, on the official reading.
2. Datasets behind ID-password or technical measures cannot be collected by circumvention without proviso risk.
3. A dataset provided as a business to specific persons, in substantial volume and managed electronically, is "shared data with limited access" under Article 2(7) of the Unfair Competition Prevention Act (Act No. 47 of 1993) even when it is not secret, and a confidential one may be a trade secret under Article 2(6).

Whether a contract can exclude Article 30-4, which UCPA remedies apply and when Article 2(7) took effect are not covered here. The clauses that carry access, delivery and confidentiality are set out in AI data licensing agreements, clause by clause.

Personal Data in a Japanese Training Set: The 2026 APPI Amendment

Japan has enacted, but not yet brought into force, an amendment to the Act on the Protection of Personal Information that will let personal data be provided without consent when it is used only to create statistical information, which the amendment treats as including AI development. The Act (Act No. 57 of 2003) is administered by the Personal Information Protection Commission (PPC). The amendment went to Cabinet and to the 221st Diet on 2026-04-07, passed on 2026-07-10 and was promulgated on 2026-07-17; it applies in principle within two years of promulgation, and the PPC states no exact dates. Its main points:

• Statistics exception: new Articles 30-2 and 31-3 allow consent-free provision of personal data, and acquisition of public sensitive data, used only for creating statistical information.
• Conditions: publication, a written agreement between provider and recipient, and use limits.
• Surcharges: Articles 148-3 to 148-17 set surcharges equal to illegal gains, limited to large-scale cases with a threshold of 1,000 data subjects.
• Facial feature data: the opt-out provision of facial feature data is banned.

Until the amendment applies, the current rules apply; their consent, third-party provision and transfer details are not covered here. Basic approach papers of the PPC followed on 2026-09-16, 2026-10-01 and 2026-10-14; the comparison is in privacy law and AI training data compared.

Japan's AI Promotion Act: A Framework Without Licensing Duties

Japan's AI statute, the Act on Promotion of Research, Development and Utilization of AI-Related Technologies, sets up government planning and imposes no data-licensing obligations or penalties found in the sources used here. It was promulgated and partly in force on 2025-06-04 and fully in force on 2025-09-01 (e-Gov ID 507AC0000000053). It creates the AI Strategy Headquarters and the AI Basic Plan, which the Cabinet adopted on 2025-12-23. Unlike the EU AI Act, it contains no duty to publish a training content summary.

Final, Enacted and Not Yet in Force: Japan as of November 2026

Three Japanese texts are in force, one is enacted but not yet applicable, and one is non-binding guidance; the table dates each step.

The table lists the dated steps, newest first, as of 2026-11-09, each with its official source.

As of 2026-11-09, the Tokyo District Court is hearing suits by Yomiuri, Nikkei and Asahi against Perplexity over AI search answers, a dispute about outputs rather than training licenses. No court case number is published on a court source; see the newspaper suits and their status.

Licensing From Japan to Domestic and Foreign Buyers

A Japanese holder can license through domestic channels or to foreign buyers, and each route carries its own contract law and tax point.

On the domestic side, PIXTA AI is run by PIXTA Inc., which is listed on the Tokyo Stock Exchange Standard market (code 3416) and was established on 2005-08-25. PIXTA AI states that third-party providers list datasets on its marketplace and contract directly with buyers. Its terms, last updated 2025-03-03, state that Japanese law and the Tokyo District Court apply (Art. 16), that consent records are required where personal information is included (Art. 6.6), and that sellers pay fees with no published rate (Art. 6.7), checked 2026-11-09. PIXTA is the marketplace, not the buyer of the listed data; see the PIXTA AI profile.

For foreign routes, AWS Data Exchange states that providers resident or incorporated in Japan are eligible, with non-US providers giving a W-8 form and US bank or virtual account details, checked 2026-11-09. Mercor's call-center listing states Japan among its location preferences, checked 2026-11-09.

For tax, IRS Table 1 (Rev. May 2023) lists a 0% US treaty withholding rate for Japan on copyright royalties (income code 12) and on know-how and other industrial royalties (income code 10), subject to the treaty's limitation on benefits. The full route is in selling data to AI companies from outside the US.

Benefits, Risks, Downsides and Alternatives of Licensing Data Under Japanese Law, Including Not Selling

Japanese law gives a holder fewer exclusive rights over analysis than EU or UK law, so the case for licensing rests on access, quality and protection, and not licensing remains a valid choice.

Benefits

• Paid analysis-ready databases sit inside the proviso on the official reading.
• UCPA Article 2(7) protects a qualifying dataset provided as a business after delivery.
• IRS Table 1 lists 0% US treaty withholding on royalties for Japan.

Risks

• The General Understanding is non-binding, so the reading of the proviso can change.
• Personal data stays under the current APPI until the amendment applies.
• Surcharges apply once the amendment is in force.

Downsides

• A developer may rely on Article 30-4 instead of paying, so open web content has little license value.
• Marketplace terms such as PIXTA's demand consent records where personal information is included.

Alternatives

• Keep data behind access controls and license only analysis-ready sets.
• Wait for the APPI rules before licensing personal data.
• License non-personal data only.
• Do not sell or license at all; see alternatives, including not selling.

Beyond Japan: Comparisons, Neighboring Jurisdictions and the Seller Process

Japanese law stops at the copyright exception's border and at the deal; the pages below cover comparisons, neighbors and the sale.

• AI training data law by jurisdiction: the parent page.
• Sell or license company data to AI companies: the steps of a sale or license.
• Rights audit before licensing: whether the holder owns what it licenses.
• Web scraping for AI training: scraping rules compared.
• Employee data and AI training rules: employee data rules by jurisdiction.
• European Union rules for AI training data: the EU model with a machine-readable reservation (DSM Directive Art. 4(3)).
• Trade secrets in AI data licensing: trade-secret and UCPA rules compared.
• Tax on data licensing income: withholding and treaty relief.
• Gengo Shigen Kyokai (GSK): a Japanese language-resource distributor.
• South Korea, China and Singapore: neighboring jurisdictions.

How This Page Is Sourced and Dated

Every rule here cites e-Gov law data, a ministry or regulator page, or a buyer's own terms, with the date it was checked. The rules for sources and dates are set out in how we source, date and check every figure. This page is general information about Japanese law, not legal advice for any specific deal.

Frequently Asked Questions

Does Article 30-4 let AI developers train on any work in Japan?

No, the proviso excludes uses that unreasonably prejudice the copyright owner. Paid analysis-ready databases and circumvention of robots.txt or ID-password measures fall inside the proviso on the official reading.

Can a rightholder opt out of AI training in Japan?

No statutory reservation exists in Article 30-4. The official reading treats circumventing robots.txt or ID-password measures as falling within the proviso, so technical controls matter through that reading.

Is the 2026 APPI amendment in force?

No, it was promulgated on 2026-07-17 and applies in principle within two years. The rules are still in preparation as of 2026-11-09.

Does Japan's AI Promotion Act regulate training data licenses?

No licensing duties or penalties were found. The Act is a framework statute in force since 2025-09-01.