Selling Data for AI Training: Privacy Law by Country

In almost every country outside the EU and the UK, a company may license records that identify people for AI training only with those people's consent, or after anonymising them to that country's own standard. A ground like the GDPR's legitimate interests, which lets a company go ahead after a written balancing test, is rare: Brazil's law has one, Türkiye's has a narrow one, and Canada, Japan, South Korea, India, Malaysia, the UAE, Mexico, New Zealand, Israel and China have none. Sending the records to a US buyer then runs on a contract or on consent, because none of these countries treats the United States as adequate: Brazil recognises only the EU, South Korea only the EU and the EEA, Japan does not list the US, and Canada, Australia and Malaysia keep no list at all. The hardest places to license from are South Korea, Japan, the UAE and Canada (and China); the most workable is Mexico, then the countries where a contract carries the transfer.

This guide is for the founder, CEO or counsel of a company outside the United States and Europe that holds records (support tickets, chat logs, email, CRM notes, documents, code, call recordings) and is deciding whether to license them to an AI buyer. A holder is the company that holds the records; a buyer is the AI lab or data company that trains on them. For fourteen countries it sets out what the new purpose needs, how the records may reach a US buyer, when anonymised data leaves the law, the AI and text-and-data-mining rules, what a holder does first and what is unsettled. The EU, the UK and Switzerland have their own guide, and every country has a page in the country register.

Contents

1. The answer in one table
2. Four patterns that hold almost everywhere
3. The EU, the UK and Switzerland in brief
4. Canada, with Quebec
5. Australia
6. New Zealand
7. India
8. Singapore
9. Japan
10. South Korea
11. Brazil
12. Mexico
13. Israel
14. The United Arab Emirates
15. Malaysia
16. Türkiye
17. China
18. Anonymisation standards side by side
19. Worked example: one software company, three countries
20. Questions people ask
21. What this means for you

1. The answer in one table

Every law here asks the same four questions of a licence. Does the information identify a living person, or could it? If so, may it serve a purpose the person was not told about at collection? May it leave for the United States? And does any AI or copyright rule change the answer? Records that identify no one, such as documentation, code without personal data or aggregate statistics, are outside all of these laws.

Read across a row for one country, or down a column to compare. "Difficulty" is our reading of the work a holder faces to license ordinary customer records to a US buyer, not a legal category.

2. Four patterns that hold almost everywhere

The fourteen laws differ in wording, but four patterns run through them and decide most of the work. The first is that consent is the default. These laws list the grounds on which personal data may be used, and the only one that reaches a paid licence to an unrelated company is usually the person's agreement. The exceptions that sound close are drawn narrowly: Canada's business transaction exception excludes deals whose purpose is selling personal information, and Singapore's business improvement exception covers sharing within a group, not a licence to a stranger.

The second is that no country here lists the US as adequate. The EU, the UK and Switzerland accept a US buyer certified under the Data Privacy Framework; no other country in this guide does. Some keep a list without the US (Brazil, South Korea, Japan); most keep none and put the burden on the holder. Either way the transfer runs on a contract binding the buyer to the home country's standard, or on the person's consent after being told where the data goes.

The third is that anonymisation is the universal exit, at different heights. Every law protects only information about an identifiable person, but the test runs from Israel's reasonable effort to the UAE's in any way whatsoever (compared in section 18). Removing names is never enough: Canada's regulator found in 2026 that stripping names, phone numbers and email addresses did not make data anonymous.

The fourth is that copyright exceptions do not unlock private records. Japan and Singapore let a user copy works it can already reach for analysis, and Israel's Ministry of Justice reads training as generally fair use. Private records are not reachable without the holder, so the licence is the access, and none of these rules touches privacy law. That is why buyers pay for private records at all (see how AI companies get training data).

3. The EU, the UK and Switzerland in brief

These three are the exception to the first two patterns, and their own guide covers them in full. A holder there usually relies on legitimate interests with a written balancing test and a compatibility test, tells people first and lets them object; truly anonymous data is outside the GDPR.

Transfers are easier than anywhere below: the US is adequate for a buyer certified under the EU-US Data Privacy Framework, its UK Extension or the Swiss-US framework, and for any other buyer the holder signs the standard contractual clauses or the UK equivalent.

4. Canada, with Quebec

PIPEDA, the federal private-sector law, reaches any commercial activity, and a paid licence is one. Quebec adds its own Act, rewritten in 2021, with stricter rules on consent, transfers and anonymisation. Bill C-36 would add a legitimate interest ground, but it is pending, not law.

New purpose. Consent is needed before information is used for a purpose not identified at collection or disclosed for another purpose, and the purpose must be one a reasonable person would consider appropriate. The business transaction exception excludes selling personal information, and the research exception needs consent to be impracticable and notice to the Commissioner. The regulator says new third-party disclosures need consent, opt-out works only for non-sensitive, limited, well-defined uses, and call recordings may serve only the purposes customers were told. Quebec requires consent for any secondary use and any communication to a third person, asked per specific purpose; its commercial transaction and research exceptions do not fit a licence.

To a US buyer. There is no list of approved countries and no ban: the holder stays accountable, binds the buyer by contract and tells people their data may be reachable by foreign authorities. Only a transfer to a processor working for the holder counts as a mere use; a buyer training its own model is not one. Quebec requires a privacy impact assessment and a written agreement before data leaves the province, other provinces included.

Anonymised data. Outside PIPEDA only if there is no serious possibility of re-identification. In Quebec de-identified data is still personal; anonymised means irreversible, done under a regulation requiring a qualified supervisor and a written risk analysis, and the regulator calls it almost impossible to guarantee.

AI and TDM. The regulator's generative AI principles ask for a documented legal authority for each disclosure used in training. Copyright has no text and data mining exception.

First steps.

1. Map each record set to PIPEDA, Quebec's Act or both, and treat AI licensing as a new purpose.
2. Anonymise to the applicable test, or collect specific consent (explicit for sensitive data and recordings).
3. For Quebec, finish the impact assessment and written agreement before anything leaves; bind the buyer by contract and give the foreign-access notice.

Unsettled. No regulator guidance on licensing records for training or on whether anonymising is itself a use; Bill C-36.

5. Australia

The Privacy Act 1988 and its Australian Privacy Principles (APPs) govern private companies, and the regulator, the OAIC, has published guidance on training generative AI. A reform bill is an exposure draft under consideration, not law.

New purpose. A secondary-purpose disclosure needs consent, or must be reasonably expected and related to the original purpose. The OAIC says that will be difficult to establish when service data trains a model sold outside the service, and its bank example asks first whether the data can be de-identified. Two traps catch smaller holders: the small business exemption is lost by trading in personal information, and the OAIC says any payment counts; and the employee records exemption covers only acts directly related to employment.

To a US buyer. The holder must take reasonable steps so the buyer handles the data as the APPs require, which the OAIC says generally means an enforceable contract, and stays accountable for the buyer's breaches. Alternatives are express informed consent or a buyer bound by a substantially similar law. Countries can be prescribed under a statutory test, but none has been. The disclosure needs an APP 6 basis as well.

Anonymised data. De-identified information is no longer about a reasonably identifiable person; the OAIC asks for a very low risk in the context, not zero. De-identifying for a licence is itself a use, and data no longer needed must be destroyed or de-identified.

AI and TDM. The government ruled out a text and data mining exception in October 2025. From December 2026 privacy policies must describe automated decisions.

First steps.

1. Assume the Act applies once a fee is paid for personal information.
2. For historical service data, plan consent or an informed opt-out, or de-identify to a very low risk with a basis for that step.
3. Sign an enforceable contract requiring APP-standard handling, and update the privacy policy on overseas disclosure.

Unsettled. No OAIC statement on licensing records to a data buyer beyond its bank example; the reform bill.

6. New Zealand

The Privacy Act 2020 works through thirteen information privacy principles (IPPs): IPP 10 governs use, IPP 11 disclosure and IPP 12 sending information abroad. There is no AI statute.

New purpose. Information may serve another purpose only if the new use is directly related to the original, the person authorised it, or it is used in a form that identifies no one, among listed exceptions; IPP 11 says the same of disclosure. There is no legitimate interest ground, and the Commissioner says AI training must be made clear at collection. A buyer's training is not directly related on our reading, so the holder needs authorisation or non-identifying data.

To a US buyer. IPP 12 allows disclosure abroad if the person authorises it after an express warning that the buyer may not protect the information comparably, or the holder believes on reasonable grounds the buyer must protect it comparably, for example by agreement.

Anonymised data. Information is personal if someone can link it with other information to identify the person. No certification standard exists.

AI and TDM. Commissioner guidance only. Copyright was outside the research behind this guide.

First steps. Check what people were told at collection; de-identify and test by linking, or get authorisation; for personal data left in the lot, give the IPP 12 warning or sign a comparable-safeguards agreement.

Unsettled. No guidance on licensing records for training; the current list of prescribed countries was not read.

7. India

The Digital Personal Data Protection Act 2023 and its 2025 Rules are law, but the parts that matter are not yet in force: consent, legitimate uses and transfers start in about May 2027. Until then older IT Act rules apply, and they regulate only sensitive data such as passwords, payment details, health and biometrics.

New purpose. Today ordinary records are lightly regulated. From about May 2027, personal data may be processed only with consent or for a closed list of legitimate uses, with no legitimate interest or business improvement ground, so a licence will need fresh, specific consent after a stand-alone notice. The employment ground does not cover handing staff messages to a buyer.

To a US buyer. A negative list: the government may restrict named countries and has named none. Sector rules still apply; payment system data must stay in India.

Anonymised data. The Act protects only data about a person identifiable by or in relation to it, with no anonymisation standard, so the holder must be able to show non-identifiability without an official test.

AI and TDM. No AI statute, and the Copyright Act has no text and data mining exception; a committee's proposed statutory licence for training is not law.

First steps. De-identify customer and staff data and test it; before May 2027 get permission for sensitive data, after it specific consent; keep payment data in India; bar re-identification in the licence.

Unsettled. The exact start date, what "research" covers, and any anonymisation standard.

8. Singapore

The Personal Data Protection Act 2012 (the PDPA) is consent-based with listed exceptions, and the regulator, the PDPC, issues advisory guidance, including on AI systems (2024) and generative AI (July 2026).

New purpose. Personal data may not be used or disclosed without consent, actual or deemed, or an exception. Business improvement caters for sharing within a group, not a licence. A legitimate interests exception exists, but needs a written assessment and a public statement, and the PDPC gives no example of a licence qualifying.

To a US buyer. No approved-country list. The buyer must be bound by enforceable obligations at least comparable to the PDPA: a contract naming the countries, a recognised certification, or consent after a written summary of the protection abroad. This sits on top of the basis for the disclosure.

Anonymised data. Not personal data when there is no serious possibility of re-identification, judged with what the recipient has or is likely to have. Pseudonyms are only a first step.

AI and TDM. Copyright allows copying for computational data analysis by a user with lawful access who does not supply the copy onward; it lets a buyer train, not a holder supply.

First steps. Sort customer, staff and non-personal data; anonymise to the serious-possibility test or collect specific consent; sign a comparable-protection contract; bar re-identification and onward supply.

Unsettled. No PDPC guidance on licensing records to an AI buyer.

9. Japan

The Act on the Protection of Personal Information (the APPI) is strict about third parties and foreign transfers but has a formal anonymised route no other country here matches. A July 2026 amendment adds a no-consent route for statistics, possibly including some AI development, but it is not in force and its scope waits on regulator rules.

New purpose. Giving personal data to a third party needs prior consent unless an exception applies; a buyer training its own model is not a processor, successor or joint user. The opt-out route is closed for sensitive information and foreign provision. A ticketing system or CRM is a searchable database, so the rules apply.

To a US buyer. The US is not on the regulator's equivalence list, so the holder needs consent to the foreign provision after telling the person about the US system and the buyer's measures, or a buyer with a conforming system (a contract ensuring APPI-level handling, or a privacy rules certification). Each provision is recorded.

Anonymised data. Japan's anonymised personal information is a legal category made to the regulator's standard. It may go to a buyer without consent once the holder publishes the categories and method and tells the buyer in writing; the buyer may not re-identify. Pseudonymised data may not go to a third party at all, and free-text tickets are hard to bring to the standard.

AI and TDM. Copyright Article 30-4 allows data analysis unless it would unreasonably prejudice the copyright holder's interests. It does not displace the APPI, which has no training rule.

First steps. List the personal, sensitive and voice data; read the published purpose of use; then choose formal anonymisation, each person's consent with the US notice, or aggregate statistics only.

Unsettled. Which AI development counts as statistics under the 2026 Act, and when it starts.

10. South Korea

The Personal Information Protection Act (PIPA) has two distinctive features: a consent-free route for pseudonymised data used for scientific research, which the regulator says includes AI, and a separate consent for transfers abroad that applies even to pseudonymised data. An AI special provision passed in August 2026, but it covers only AI with a public benefit and is not in force.

New purpose. Use beyond the collection purpose, or provision to a third party, needs separate consent. Pseudonymised data may be processed without consent for statistics and scientific research, which the regulator's March 2026 guideline says includes AI learning, while rating provision to a buyer-controlled environment as high risk.

To a US buyer. Cross-border provision and storage are barred unless a ground applies. Only the EU and EEA are recognised as equivalent, with the UK under review, so for a US buyer the ground is separate consent naming the data, country, buyer, purpose, retention and how to refuse, pseudonymised data included. The regulator may suspend a transfer.

Anonymised data. Information that identifies no one, reasonably considering time, cost and technology, is outside the Act: the only route that avoids both consents.

AI and TDM. The culture ministry applies fair use to training, case by case.

First steps. Test whether the lot is anonymous in the buyer's environment; if not, get separate consent for the use and the transfer; never pseudonymise a stock in advance to sell later.

Unsettled. No guidance on licensing ordinary records to a foreign buyer.

11. Brazil

Brazil's General Data Protection Law (the LGPD) is the one law here, outside Europe, with a real legitimate interest ground for a licence, and its regulator, the ANPD, has set a clear contract route for transfers. It has not said whether legitimate interest supports AI training.

New purpose. For a licence, the bases that fit are consent and legitimate interest, and the interest may be a third party's, so the buyer's can count. Legitimate interest allows only strictly necessary data, counts people's expectations and excludes sensitive data. ANPD staff treated training on data collected for other purposes as a secondary use needing clear prior notice.

To a US buyer. Transfers need an adequate country, approved guarantees such as standard contractual clauses, or specific consent. Only the EU is adequate, so the holder signs the ANPD's clauses unchanged as an addendum to the licence and publishes a Portuguese notice.

Anonymised data. Outside the law unless reversible with reasonable cost, time and technology.

AI and TDM. No AI statute is in force; the AI bill waits in the Chamber. The research behind this guide found no text and data mining exception.

First steps. Mark personal, sensitive and children's data; anonymise, or rely on specific consent or (non-sensitive data) a written legitimate interest test with an objection route; give notice; sign the clauses.

Unsettled. Legitimate interest for training, and the ANPD's planned AI and anonymisation rules.

12. Mexico

Mexico replaced its privacy law in March 2025 and moved the regulator into a government ministry. The model stays the same: a privacy notice sets the purposes, and consent can be tacit for ordinary data, which makes Mexico the most workable country here.

New purpose. A different purpose needs consent again, and consent can be tacit (the notice was available and the person did not object). Financial data needs express consent and sensitive data written consent. A fresh notice naming AI licensing with a real way to object can support tacit consent on our reading, though the route is untested.

To a US buyer. A licence is a transfer, treated the same at home or abroad: the holder passes on the notice and purposes, and the buyer assumes the holder's obligations. No approved-country list.

Anonymised data. Dissociated data, which can no longer be associated with the person, needs no consent.

AI and TDM. None in force; an AI initiative is pending.

First steps. Update the notice (purpose, recipient, transfer clause, objection route), or dissociate; bind the buyer to the notice's purposes.

Unsettled. The regulations under the 2025 law and any guidance from the new regulator.

13. Israel

Israel's Protection of Privacy Law was amended by Amendment 13, in force since August 2025; transfers abroad follow 2001 regulations. Only the Hebrew text binds.

New purpose. Database data may be processed only for the database's purpose, so a licence is a new purpose. The regulator's draft AI guidance wants a legal basis at every stage, training included, and a separate opt-in where a use goes beyond what people expect. The sections read show no legitimate interest ground.

To a US buyer. Consent or an agreement binding the buyer to Israeli standards, plus the buyer's written guarantee of adequate measures and no onward transfer.

Anonymised data. The law covers people identifiable with reasonable effort, but the regulator warns anonymisation is no guarantee, especially with AI.

AI and TDM. A 2022 opinion treats training on works as generally fair use and expressly excludes privacy.

First steps. Anonymise and document the test, or get a separate opt-in; for personal data, sign the agreement and take the guarantee.

Unsettled. The AI guidance is a draft; EEA-origin data has extra rules.

14. The United Arab Emirates

The federal Personal Data Protection Law of 2021 covers mainland companies; free zones with their own law (the DIFC, the ADGM) and health and banking data follow other rules. The Executive Regulations could not be confirmed as issued.

New purpose. Processing without consent is forbidden outside a closed list, and data may not be processed incompatibly with its purpose. A licence needs fresh, specific, provable consent; there is no legitimate interest and no opt-out.

To a US buyer. An approved country, or a contract binding the buyer to the law's measures, or explicit consent. No approval names the US.

Anonymised data. Voice, image and electronic identifiers are personal data, and anonymisation must make identification impossible in any way whatsoever, the strictest wording here.

AI and TDM. Nothing found.

First steps. Confirm the holder is outside a free zone; get consent or anonymise; use explicit consent or a contract for the transfer, after checking with local counsel for new regulations.

Unsettled. The Executive Regulations, any list of approved countries, and penalties.

15. Malaysia

The Personal Data Protection Act 2010 covers personal data in commercial transactions. A 2024 amendment rewrote the transfer rule, in force since April 2025.

New purpose. Processing needs consent and written notice before a new use, and disclosure for a new purpose is limited to listed exceptions; beyond consent, only a reasonable belief the person would have consented could fit, unsupported by the regulator. The research exemption needs data processed for no other purpose, and the regulator's AI paper says AI creates no separate legal basis.

To a US buyer. No whitelist since 2024: consent, or all reasonable precautions and due diligence, on the holder's own recorded finding. Consent needs prior written notice; precautions mean contract clauses requiring Act-level security and compliance.

Anonymised data. No standard; the regulator's paper treats pseudonymised data as still carrying re-identification risk.

AI and TDM. No AI rule; copyright has only fair dealing.

First steps. Anonymise, or notify and record consent; record the transfer finding; put the clauses and a stop-transfer right in the licence.

Unsettled. Guidance on licensing, anonymisation and staff data.

16. Türkiye

Türkiye's Personal Data Protection Law (the KVKK) was amended in 2024 to rebuild transfers around standard contracts set by its data protection board, and the regulator published a generative AI guide in November 2025.

New purpose. Explicit consent is the rule, and one listed alternative is the controller's legitimate interests where the person's fundamental rights are not harmed. The regulator's AI guide narrows it: legitimate interest is no catch-all, needs a case-by-case balancing test and should not aim only at profit, and a new use must be tested against the original purpose and people's reasonable expectations. For tickets, chat and recordings a holder should assume consent.

To a US buyer. The regulator has made no adequacy decision for any country. The routes are a safeguard such as the board's standard contract, notified to the regulator within five business days, which must be used unchanged, under Turkish law and courts; or explicit consent after being told the risks, only for a transfer that is occasional and not part of ordinary business.

Anonymised data. Data made anonymous is outside the law, but the claim must be shown by technical methods and objective criteria, and the anonymising work is itself subject to the law.

AI and TDM. No AI statute; the regulator's guide applies purpose limitation to AI development. Copyright was not researched for this guide.

First steps. Document anonymisation, or get explicit consent naming training, buyer and transfer; use the standard contract or, for a single licence, occasional consent.

Unsettled. Whether legitimate interest can ever support a paid licence.

17. China

The Personal Information Protection Law (the PIPL), the Data Security Law and the regulator's 2024 cross-border provisions make China the most procedural country here.

New purpose. The PIPL's grounds include no legitimate interest. Providing data to another handler needs notice of the recipient and its purpose and the person's separate consent, and a changed purpose needs consent again. Sensitive data (biometrics, health, financial accounts, location, children's data) carries stronger notice duties, and voice recordings may be biometric.

To a US buyer. Every export needs notice, separate consent and an impact assessment, which the volume exemptions do not remove. The filing route depends on how many people's data has gone abroad since 1 January: fewer than 100,000 people is exempt; between that and 1,000,000 needs the standard contract or certification; 1,000,000 or more, or sensitive information of 10,000 or more, needs a security assessment. Data a sector or region has listed as important data always needs the assessment.

Anonymised data. Only data that cannot identify anyone and cannot be restored leaves the PIPL; de-identified data stays personal.

AI and TDM. The generative AI rules require lawful, consented training data, but apply to services offered to the public in China, so a US lab that does not serve China is outside them.

First steps. Test irreversible anonymisation and check no record is important data; otherwise consent, the impact assessment and the volume route.

Unsettled. No guidance on licensing to a foreign AI buyer; enforcement risk is high.

18. Anonymisation standards side by side

Anonymisation works in every country, so its test is the most useful thing to compare. The words decide whether removing direct identifiers and generalising dates is enough, or whether the holder must show that no one, anywhere, could link a record back. Three questions separate the tests: whose means count, whether the standard is a risk level or an absolute, and whether the anonymising step is itself a use needing a basis.

A holder licensing across countries should meet the strictest test that applies to any of its records, or split the lot by country.

The tests fall into three groups. Reasonable-effort tests (Israel, Brazil, South Korea) ask what someone could do with realistic cost and time. Risk tests (Australia, Canada, Singapore) ask how likely re-identification is in the release's context, and are stricter in practice. Absolute tests (the UAE, Quebec, China) leave the least room, and a holder that cannot meet them should plan on consent.

Free text is the hard case everywhere: a ticket or email names places, orders and details that identify someone in combination, and no regulator here has published a free-text method. That is why buyers ask how records were de-identified, not whether. Our de-identification standard sets out the method and its measured accuracy, and the de-identification tool runs it on a sample in the holder's own browser.

19. Worked example: one software company, three countries

This is an example, not a real company. Northbeam is an 80-person software company headquartered in Montreal. It sells field-service scheduling software to small businesses across Canada, with a Sydney subsidiary that signs Australian customers and a Tokyo subsidiary for Japan. A US buyer wants eight years of support tickets, chat transcripts, recorded calls, product documentation and source code to train a service agent.

Which law governs which records. Each entity's records follow the law where it collected them. The Montreal parent's records fall under Quebec's Act and PIPEDA. The Sydney subsidiary's fall under the Privacy Act; it is small, but it loses the small business exemption once it is paid for personal information. The Tokyo subsidiary's fall under the APPI. Documentation and code hold no customer personal data, so only ownership and customer contracts matter there (see training data provenance). Staff names in tickets are personal data in all three (see employee emails and chats).

What each country requires. Quebec: consent or irreversible anonymisation, and for personal data a privacy impact assessment and written agreement. Australia: consent or an informed opt-out, since the OAIC doubts reasonable expectation covers training, or de-identification to a very low risk, plus an enforceable contract. Japan: prior consent to the provision and to the foreign transfer, or the formal anonymised route.

The order to work in, from least friction to most:

1. Code and documentation. No privacy step; license once ownership is confirmed.
2. Australian tickets and chats. De-identify to a very low risk, document the method and the basis for that step, and add an APP-standard clause for anything residual.
3. Canadian tickets and chats. Anonymise to the no-serious-possibility test, remembering the regulator rejected removing names, phone numbers and emails alone. For the Quebec share, meet the irreversible standard or run the assessment and sign the agreement.
4. Japanese tickets. Produce formal anonymised information, or leave them out of the first lot.
5. Call recordings. Voices identify people, and Canadian recordings may serve only the purposes customers were told: consent, or license de-identified transcripts instead.

Location also moves price: buyers' estimators price a company in Canada at about 0.71 times the same company in the US, and one outside the US, Europe, the UK and Canada at about 0.27 times (see what company data is worth). Splitting the lot lets Northbeam license the clean parts now instead of holding everything for the hardest country.

Questions people ask

Can I sell customer data for AI training without consent?

Outside the EU and the UK, usually not if it identifies people. Most countries here require consent or a non-identifying form for a new purpose like a licence. Brazil allows legitimate interest for non-sensitive data with a written test and notice, and Mexico accepts tacit consent after a fresh notice. Anonymised data needs no consent anywhere.

Is the United States an adequate country for data transfers?

For the EU, the UK and Switzerland, yes, for a buyer certified under the Data Privacy Framework. For every other country here, no: Brazil recognises only the EU, South Korea only the EU and EEA, and most keep no list, so the transfer runs on a contract or consent.

Which country is easiest for licensing data to AI companies?

Mexico, for ordinary records: an updated notice with a way to object can supply tacit consent, and the buyer takes on the holder's duties. Australia, New Zealand, Singapore, Malaysia, Israel and Brazil follow, each with a clear contract route abroad.

Does anonymised data fall outside privacy law?

Yes, everywhere here, if truly anonymous. The test runs from Israel's reasonable effort to the UAE's in any way whatsoever, and Australia treats the anonymising step as a use. Pseudonymised data, where a key exists, stays personal almost everywhere.

Does Japan's copyright exception let me sell data for AI training?

No. Article 30-4 lets a user analyse works it can already reach; it does not let a holder supply records or displace the privacy law. Singapore's exception likewise bars supplying the copy onward.

Do I need consent to license employee emails and chats?

Usually, or anonymisation. Australia's employee records exemption covers only acts directly related to employment, and no country here carves staff data out for a licence. See employee emails and chats.

What about data I hold for my customers?

If customers keep the records in your product and you process them on their behalf, the licence is theirs to give, not yours. Check the contracts first; the GDPR guide explains the processor rule and what a data licensing agreement contains explains how buyers check it.

What this means for you

Start with what needs no privacy step: code, documentation and records about companies rather than people. Sort the rest by the country whose law governs it, and for each set either anonymise to that country's test and document it, or collect the consent its law requires. Put the transfer terms in the licence itself, since a contract binding the buyer to your country's standard is the route almost everywhere, and leave sensitive data and voice recordings for last, or out.

The country register gives each country's privacy work, US withholding rate, buyers and price against a US company, and /countries#buyers lists every buyer's own rules. Four guides cover the rest of selling from abroad: selling data to US AI companies from outside the US, licensing under the GDPR, US withholding tax on data licence payments and which AI data buyers take holders outside the US. For the US rules read is it legal to sell customer data, and for the whole branch, the rights hub.

To test whether your records can be de-identified, run a sample through the de-identification tool in your own browser and read the standard behind it. To see what they could be worth, the valuation gives an estimate with no upload. We run sales only for holders based in the United States, the United Kingdom and Canada today; the valuation works for holders anywhere and binds no one.

This guide is research from public sources, not legal advice.