The decisive questions are who holds the relevant rights, what the business promised in its contracts, which statutory permissions apply to the particular acts, and whether personal or confidential information can be reused. An output-ownership clause answers only part of that inquiry. Copyright exceptions and mandatory contract rules differ substantially between countries.
This guide compares fifteen countries across seven regions, with EU rules and United Arab Emirates (UAE) federal, Dubai International Financial Centre (DIFC) and Abu Dhabi Global Market (ADGM) levels considered separately. The selection reflects material legal differences and developments. It supports a route decision once the business identifies its account, contracts, copying locations, dataset and release plan. It states the law, public terms and product documentation as at 21 November 2026.
Start with the acts the business proposes
Distillation uses a teacher's behaviour to improve a student, often through generated responses. Fine-tuning changes an existing model through training, so the same project can involve both. The Amazon Bedrock documentation describes a managed version of that process. The legal analysis should follow the copies, purposes and recipients at each stage.
- Access and generation. Question to settle before proceeding: Which entity holds the account, which service plan applies, and may it generate and retain responses for this purpose?
- Dataset construction. Question to settle before proceeding: Does collection reproduce protected text, code, images or database content? What rights and duties attach to the prompts and retrieved material?
- Training or fine-tuning. Question to settle before proceeding: Does the grant or a valid exception cover these copies, the intended student and its competitive use?
- Evaluation. Question to settle before proceeding: Is the dataset confined to an authorised benchmark, or will responses and evaluation results feed later training?
- Retrieval-augmented generation (RAG). Question to settle before proceeding: May the business store, retrieve and communicate the material? Will it also add the retrieved content or responses to a training corpus?
- Commercial release. Question to settle before proceeding: May it supply the hosted service, adapters or weights, and can it support the rights and data assurances promised to customers?
A dataset vendor adds another actor to this chain. The buyer needs evidence of the vendor's authority to generate and transfer the responses as well as its own training rights. A vendor's assurance that the data is synthetic leaves the underlying account, source-content and onward-use questions to be answered.
Choose a route with a demonstrable permission chain
The following list distinguishes public contractual conditions from documented product support. The dates of the relevant terms and documentation are part of the comparison; an earlier generation run may be governed by an older accepted contract or an enterprise amendment.
- Own teacher. Position and decisive condition: Document ownership, licences or another applicable legal basis for the relevant code, weights, input and training-data uses. Customer data and contractor contributions can have separate conditions. Required decision: Confirm the rights and data-use chain, then check the student base and release form.
- Original DeepSeek-R1, self-hosted. Position and decisive condition: The repository's licence statement expressly includes distillation, with MIT terms for the original code and weights. Distilled variants identify their upstream Qwen or Llama licences. Required decision: Freeze the exact artefact and commit; apply its licence and all relevant upstream conditions. Hosted API access requires its own terms.
- Llama 3.1 materials. Position and decisive condition: Under the 23 July 2024 licence, distributing Llama Materials, derivatives or products/services containing them carries agreement and attribution duties. A model trained or improved with the outputs and then distributed or made available needs a Llama name prefix. Section 2 requires separate Meta permission above 700 million monthly active users for licensee/affiliate products or services in the calendar month preceding release, subject to section 1(b)(ii)'s integrated end-user-product exception. Required decision: Check the actual artefact, scale, acceptable-use rules and intended distribution before training and naming the released student.
- Amazon Bedrock managed distillation. Position and decisive condition: As at 21 November 2026, the prerequisites list supported Nova and Meta combinations and regions. Examples include Nova Pro to Nova Lite or Micro in Northern Virginia, and Llama 3.1 70B to Llama 3.1 8B in Oregon. Required decision: Match the listed pair and region, account access, AWS terms and applicable model terms. Establish deployment and export rights needed by the product.
- Anthropic through Bedrock distillation. Position and decisive condition: As at 21 November 2026, the prerequisites say Anthropic distillation is unavailable. The May 2025 launch announcement described earlier support. Required decision: Choose a supported route; check any later restoration against both documentation and terms.
- Anthropic commercial service. Position and decisive condition: Commercial Terms effective 17 June 2025 restrict competing use. Service Specific Terms effective 31 August 2026, C, permit qualifying fine-tuning solely for use of those Fine-Tuning Services. Required decision: Establish the required model/account availability and exact service scope. An external competing student needs another applicable permission.
- OpenAI business/API service. Position and decisive condition: The Services Agreement effective 1 January 2026 restricts using Output to develop AI models competing with OpenAI. Exceptions cover models primarily for categorisation, classification or organisation, with no third-party distribution or commercial availability, and qualifying customisation of OpenAI models. Required decision: Test the competing-use restriction and every exception condition. A separately sold competing generative student requires another supported contractual route.
- Consumer chat account. Position and decisive condition: The applicable consumer forms contain their own competing-use and automation restrictions. Their conditions differ from business agreements. Required decision: Select a service and permission that cover the intended collection and training before starting the corpus.
- Purchased generated responses. Position and decisive condition: The seller must evidence generation authority, source rights and permission for the buyer's intended training and onward supply. Required decision: Accept the evidenced portions; require missing grants or replace unsupported data.
Sources: DeepSeek-R1 licence statement, MIT licence, Llama 3.1 licence, sections 1–2, Bedrock prerequisites, Anthropic Commercial Terms, B and D.4, Anthropic Service Specific Terms, C, effective 31 August 2026, OpenAI Services Agreement, 3.3(e) and 17, OpenAI consumer terms, OpenAI Europe consumer terms, updated 16 January 2026 and Anthropic consumer terms.
Identify the contract and the party that can approve the use
For an external teacher, collect the order, account record, accepted agreement, incorporated policies, enterprise amendments and approvals. For a cloud service, add the model-provider and marketplace terms. For a purchased corpus, add the supplier's generation agreement and its licence to the buyer. The student base model supplies a further set of conditions.
Entity and governing-law tests can differ. The OpenAI business agreement uses a European Economic Area (EEA)/Switzerland entity distinction while separately including the UK in its Irish-law branch. Anthropic's commercial Irish-entity branch includes the EEA, Switzerland and UK. AWS identifies the contracting party and law through its customer-country table. The actual order and negotiated terms must resolve which form applies. OpenAI agreement; Anthropic terms; AWS Customer Agreement, updated 29 October 2026.
A managed workflow also has a contractual boundary. AWS Service Terms, updated 19 November 2026, address third-party model terms, custom models and restrictions on extracting underlying model components. The documented operation must fit that stack. A product that requires delivery of downloadable weights needs an established export grant. AWS Service Terms, 50.11–50.12; Bedrock third-party model terms.
Keep provider use of customer data separate from customer use of generated responses. A commitment about the provider's own training addresses that provider's conduct. The customer's development, competition and transfer permissions remain governed by their relevant clauses.
Apply copyright exceptions and contract limits together
Commercial AI model distillation receives different treatment depending on the protected material, purpose, access, retained copies and applicable legal system. A statutory exception can resolve a particular copyright act. Whether an inconsistent contract term is ineffective requires a further legal rule.
North America: purpose and fairness require separate findings
In the United States, section 107 requires the four-factor fair-use assessment: purpose and character, nature of the work, amount and substantiality, and market effect. The June 2025 district order in Bartz v Anthropic distinguished training, digitisation of purchased books and the acquisition of a pirated permanent library. Its factual and procedural scope matters: it supplies an example of analysing separate acts, while a commercial API-output project still needs its own copying and contract analysis. 17 USC 107; Bartz, 23 June 2025 order.
The Third Circuit's precedential September 2026 decision in Thomson Reuters v ROSS affirmed the rejection of fair use for copying protected Westlaw headnotes to train a competing, non-generative legal-search system. The court examined the commercial purpose, alternatives, copying and original and licensing markets. Applying that reasoning to teacher responses requires identifying protected material and the student's function. The decision was given on an interlocutory appeal; it does not establish that all commercial distillation infringes. Opinion, pp. 5, 10–11, 16–20 and 25–28.
US output authorship also needs precision. Thaler concerned registration naming a machine as sole author. The D.C. Circuit required initial human authorship; the Supreme Court denied review on 2 March 2026. Assess the actual human contribution to an AI-assisted work and any protected material reproduced within the response. Ownership allocation between provider and customer remains bounded by the rights each can grant. D.C. Circuit opinion, 18 March 2025; official Supreme Court docket.
US contract questions require the governing state and forum. The distinction between a copyright licence condition and an independent covenant can change the cause of action and remedy. Federal preemption also depends on the claim's elements. An actual anti-training clause therefore needs analysis of assent, scope, applicable state law and the relevant federal treatment before a business relies on defeating it. 17 USC 301.
In Canada, fair dealing first requires a permitted purpose and then fairness. The Supreme Court's CCH decision interprets research broadly enough to include commercial research, followed by examination of purpose, character, amount, alternatives, nature and effect. A commercial objective can therefore enter the inquiry, but bulk retention and distribution still require justification on the actual facts. Section 30.71 requires a copy to be essential to a technological process, serve solely to facilitate a non-infringing use, and exist only for that process's duration. A reusable corpus retained for later training needs a supported basis covering that retention and use. Copyright Act, sections 29 and 30.71; CCH, 2004 SCC 13, paragraphs 48–60.
Canada also requires a provincial contract analysis. Quebec's Civil Code controls specified external, incomprehensible and abusive clauses in consumer or adhesion contracts. Eligibility and the particular clause determine the effect. A claim that a business contract is adhesive needs its formation facts; the federal copyright result alone cannot settle that question. Civil Code of Québec, articles 1435–1437.
Europe: the research branch, commercial branch and contract rules differ
The EU Digital Single Market (DSM) Directive separates research text and data mining (TDM) for qualifying organisations under Article 3 from the wider Article 4 exception. Article 4 requires lawful access, retention limited to necessity and the absence of an appropriate express rights reservation. For online content, the reservation mechanism includes machine-readable means. Article 7's mandatory protection against contrary contract terms covers Article 3, along with Articles 5 and 6; it omits Article 4. National implementing text is therefore essential. Directive 2019/790, Articles 2–4 and 7.
Two national implementations and the UK's own rule show the practical consequence:
- Germany. Copyright route: UrhG section 44b permits qualifying TDM of lawfully accessible works unless rights are effectively reserved. Copies must be deleted when no longer necessary. Section 60d has a distinct research-beneficiary regime. Contract and evidence consequence: Section 60g protects the listed research and education exceptions against contracting out; section 44b sits outside that list. Record access, reservations, eligibility and retention, then assess the actual restriction.
- Ireland. Copyright route: CRRA section 53B provides a wider TDM route subject to lawful access, appropriate reservations, necessary retention and the statutory three-step limit. The September 2026 amendment uses “rightholder” wording. Contract and evidence consequence: Section 53B(6) expressly makes contrary contractual provisions unenforceable. The activity must first qualify. Machine-readable website or service terms can also operate as reservations, so the same document requires analysis in both capacities.
- United Kingdom. Copyright route: CDPA section 29A permits computational-analysis copies for the sole purpose of non-commercial research, with lawful access and the applicable acknowledgement condition. Contract and evidence consequence: Conflicting terms are ineffective for the permitted act. A commercial student falls outside that sole-purpose condition; later commercial reuse or transfer of research copies also needs authority.
Sources: UrhG sections 44b, 60d and 60g; Ireland S.I. 567/2021, regulation 4(a), as amended by S.I. 469/2026, regulation 2(b); CDPA section 29A.
The Irish comparison needs careful application. A contractual prohibition and a rightholder's reservation perform different legal functions. A developer must establish lawful access, a qualifying TDM act and the absence of an effective reservation before relying on section 53B's contract protection. The relevant provider must also have authority over the rights it purports to reserve. Selecting Irish law in an agreement leaves the territorial law of copying and other mandatory rules to be determined.
The EU Data Act adds a separate business-to-business (B2B) data-contract issue. Article 13 addresses specified unfair terms on access to and use of data, liability and remedies that one enterprise unilaterally imposes on another. Scope, an unsuccessful negotiation attempt, the particular term and the main-subject-matter exclusion matter. Chapter IV applies to contracts concluded after 12 September 2025, with later treatment for specified older contracts. This creates a clause-specific inquiry where its conditions fit. Data Act, Articles 1, 13 and 50.
German litigation also shows why retained expression matters. In GEMA v OpenAI, the Munich first-instance court addressed reproducible song lyrics and distinguished training-corpus copying from protected expression retained in the models. As at 21 November 2026, the judgment was under appeal, and the separate LAION dataset appeal was pending before the Federal Court of Justice, with pronouncement listed for 17 December 2026. A developer should assess its own dataset and release behaviour against the specific legal question and procedural status of each case. Munich judgment, 42 O 14139/24, paragraphs 167–193; BGH proceeding I ZR 281/25.
The 3 November 2026 Advocate General Opinion in Like Company v Google Ireland, C-250/25, supplies a different, non-binding analysis. Advocate General Szpunar favoured applying Article 4 to large language model (LLM) training within its conditions, while proposing that the training questions in that case be inadmissible and addressing them additionally. He questioned inferring copies within model parameters solely from regurgitation, while recognising possible output infringement. Lawful access, effective reservations and mitigation of closely reproducing outputs remain material. An Advocate General's Opinion does not settle the Court's ruling or displace the national decisions. Opinion, paragraphs 45–46, 84–102, 124, 132–133, 148–149 and 240.
Asia: broader analysis permissions still have purpose and supply limits
Japan's Copyright Act Article 30-4 permits necessary uses for purposes other than enjoying, or enabling enjoyment of, the thoughts or feelings expressed in a work, subject to unreasonable prejudice to the copyright owner's interests. Information analysis is included. A process aimed at learning patterns can present different facts from one aimed at reproducing a particular work for users. Mixed purposes, the amount used, market effects and later reuse require examination. Article 49 addresses subsequent uses of copies made under exceptions. Native Copyright Act, Articles 30-4 and 49.
The Agency for Cultural Affairs explains these distinctions in its May 2024 overview, which expressly has non-binding status. Article 30-4's copyright permission leaves the service contract, confidentiality and personal-data questions to their respective rules. A project using a restricted external API should obtain an express contractual route or complete the applicable contract analysis before proceeding. ACA overview.
Japan's Civil Code supplies the separate contract tests: Articles 90–91 and 521–522 govern public-order limits and agreement, while Article 548-2 addresses qualifying standard-form terms, incorporation and terms that unfairly prejudice the other party under its good-faith test. Apply those provisions to the actual restriction and transaction before treating a clause as ineffective. Official native Civil Code, Articles 90–91, 521–522 and 548-2.
Singapore combines a defined computational-data-analysis permission with an express contract limit. Sections 243–244 cover specified analysis, including using examples to improve a program's functioning. Lawful access, the status of the first copy, purpose and restrictions on supply are material. Section 187 makes terms ineffective to the extent they exclude or restrict the protected permission. Under section 188, a foreign-law choice is void if it would exclude or restrict a permitted use and either was imposed wholly or mainly to evade that use, or a party dealt as a consumer, was then a Singapore resident, and the essential contract-making steps occurred in Singapore, including through others on that party's behalf. The interpretation of Unfair Contract Terms Act section 27(2)(b) must be considered; the person asserting non-consumer status bears the burden. Establish the applicable branch and facts. Copyright Act 2021, sections 187–188 and 243–244.
That can change a developer's training-copy analysis. A business relying on it should retain the access basis and first-copy provenance, explain the precise computational purpose and control sharing of exception-made copies. Supplying a general commercial response dataset needs a separate legal basis where it falls outside the statute's limited verification or collaboration purposes. The confidentiality of source material also remains material under the Act's preservation of confidence law.
Middle East and Africa: enactments change the available arguments
Saudi Arabia's 2026 Copyright Law creates a specific AI-development copying provision. Article 26(4) requires lawful publication, lawful acquisition of ownership of the original copy and copying within the development purpose. The final implementing regulation, published on 31 July 2026, adds conditions in Article 30 concerning necessity, commercial use, normal exploitation, records and the final product. Its Article 29(3) limits software, application and database licence terms that restrict lawful statutory exceptions. The framework took effect on 12 August 2026. Law M/169, Articles 26, 37 and 59–61; final regulation, Articles 29–30.
A business evaluating that route needs evidence about publication and acquisition of the original copy, which may be difficult to establish for private personalised API responses. Article 30(2) excludes a purely commercial context unless the use is insubstantial relative to the work or does not affect normal exploitation. The business must also address the separate restrictions on direct exploitation, market harm and unnecessary incorporation of works in final products. Keep the required source, type, purpose and date records. The interaction of the regulation's general conditions with its specific AI provision requires careful construction for a commercial competing product. The controlling Arabic text and actual corpus facts should govern that assessment. Final regulation, Articles 23 and 30.
For the UAE, the applicable legal level is an initial decision. Federal copyright law recognises software and database rights, licence conditions and specific exceptions. A developer relying on a licence should identify the granted acts, purpose, duration and territory. The replacement federal Civil Transactions Law took effect on 1 June 2026, so old civil-code citations require checking against the replacement. Federal Copyright Law, Articles 7, 9, 12–13 and 22; Civil Transactions Law, official Arabic text.
DIFC has its own intellectual property (IP), contract and data-protection framework alongside the federal-rights relationship. Establish the actual activity and court connection before selecting that analysis. DIFC IP Law, Articles 17–19, 29 and 35; DIFC Contract Law, Article 10; DIFC Data Protection Law, Articles 6 and 9–13.
ADGM receives English common law subject to its own enactments; UK statutes apply only through the relevant legal route, so UK section 29A cannot simply be assumed to govern there. ADGM also has its own data-protection rules. Its section 31(4)(c) provides a limited retention route for personal data in a dataset used to lawfully train or refine an AI system without risk to data-subject rights, with a prior impact assessment and necessity limit under section 31(5). Section 31(6) requires a policy and process for handling the data when the retention grounds end. This does not authorise unlawful collection or training. ADGM reception regulation, sections 1–2; ADGM Data Protection Regulations, sections 31 and 34.
Ordinary model distillation involves no virtual-asset activity. A Virtual Assets Regulatory Authority (VARA) licensing analysis would require an additional product fact. Dubai Law 4/2022, Articles 2–3.
Nigeria offers another concrete contract distinction. Section 20(1) of the Copyright Act uses an open fair-dealing formulation and four fairness factors. Commercial character and competitive market effects remain part of that assessment. Section 20(3) voids a contractual term that restricts or prevents an act permitted under the Act. The business must first establish the permitted act, then identify the conflicting term and the relevant territorial and contract connection. Separate access, confidentiality and data-protection duties require their own authority. Copyright Act 2022, operative 17 March 2023, section 20.
South Africa's existing section 12(1) fair-dealing rule is purpose-specific. Its research wording has no express blanket non-commercial limit, but purpose, fairness and work type still matter; computer programs have separate section 19B treatment. The Constitutional Court's 26 June 2026 decision on the Copyright Amendment Bill concerned proposed provisions. As at 21 November 2026, general fair use remained a proposal. Existing Copyright Act, sections 12 and 19B. Constitutional Court judgment, CCT 306/24, paragraphs 198–209 and order.
South Africa also defines an author for specified computer-generated works by reference to the person undertaking the necessary arrangements. That directs attention to the actual generation process, ownership arrangements and originality. The statutory wording cannot identify the provider, developer or prompt customer without those facts.
Latin America: the grant and database layer can determine the route
In Brazil, copyright transactions receive restrictive construction, modes of exploitation are distinct, and assignments have writing and scope requirements. A buyer seeking a commercial student should specify dataset generation, training, processors, model distribution and any raw-data resale. The statutory private-copy and quotation provisions have defined conditions; a bulk commercial corpus needs a supported licence or a separately justified exception analysis. Software Law Article 10's invalid-clause rule concerns specified foreign-software commercialisation contracts, so its scope must be established before using it against an API term. Lei 9.610, Articles 4, 29–31 and 46–50; Lei 9.609, Articles 9–10.
Mexico adds a distinct database issue. Copyright Law Article 108 protects non-original databases for five years for their maker. A dataset can therefore require an additional rights inquiry even where individual facts or responses lack copyright. Establish its maker, creation date, contents and proposed use. Article 102, amended effective 15 May 2026, expressly includes AI programs within software protection; the particular output or weight file still needs classification under the applicable rules. LFDA, Articles 102 and 107–109, May 2026 compilation.
Mexican copyright grants require attention to written form, scope and the distinction between a transfer and a licence. For assignments of computer-program copyright, Article 103 removes Article 33's duration limit; the other applicable requirements remain. Article 148's exceptions have specific published-work, purpose and normal-exploitation conditions. For a purchased synthetic corpus, request both the relevant database grant and proof of the seller's generation and downstream training authority. LFDA, Articles 30–33, 103 and 148.
Oceania: software-study exceptions have a precise contractual boundary
Australia and New Zealand both require their own research/fair-dealing analysis. Their software provisions also differ in a way that can matter for an appropriately licensed local teacher:
- Australia. Software study: Section 47B(3) covers specified incidental automatic copying while an owner or licensee runs a program to study its functioning or underlying ideas. The relevant conditions and later-use limits apply. Effect of conflicting terms: Section 47H expressly includes section 47B(3) among the permissions protected against contrary contracts.
- New Zealand. Software study: Section 80C permits qualifying observation, study or testing during acts the person is entitled to perform. Effect of conflicting terms: Section 80D lists sections 80A(2) and 80B(1), omitting section 80C. The contract question therefore needs separate examination for that study route.
Sources: Australian Copyright Act, sections 47B–47H; New Zealand Copyright Act, sections 80A–80D.
A hosted API customer must establish the relevant lawful program-copy use before relying on those provisions. Copying and reselling response works can involve additional acts outside the software-study permission.
Australian standard-form consumer and eligible small-business contracts also face unfair-terms controls. Under the threshold as at 21 November 2026, a business may qualify through fewer than 100 employees or turnover below A$10 million, subject to the statutory conditions. The actual term, legitimate interests, imbalance and detriment require examination. Australian Consumer Law, sections 23–27.
Section 5(2)(a) of New Zealand's Copyright Act identifies the necessary-arrangements person as author of specified computer-generated works. A response buyer should establish the relevant generation and contractual arrangements before promising ownership to its own customers. The same Act's research and temporary-copy exceptions have separate conditions for purpose, fairness, copy numbers and retained economic value.
Protect confidential information and establish a privacy basis
The dataset and student can contain information that remains protected independently of copyright. A teacher may reproduce a confidential prompt, customer record, code fragment or identifiable person's information. Identify the source, who was entitled to disclose it, the purpose of disclosure and the recipient's duties. Lawful access to an ordinary response presents different facts from access to an insider's confidential corpus.
Trade-secret regimes commonly require identified secret information, commercial value or a protected interest, protective measures and qualifying wrongful conduct. The actual means of acquisition and duties matter. For example, the EU Trade Secrets Directive separately addresses independent discovery and observation of lawfully possessed material without a valid restrictive duty, while preserving claims for unlawful acquisition or disclosure. Directive 2016/943, Articles 2–4. A permission request should therefore address confidential content and authorised processors as well as output copyright.
Privacy requires a source-to-release assessment. Establish whether prompts, responses or accessible model behaviour identify people; who determines the purposes; the original collection purpose; the basis for training; the required notices; and any onward or overseas transfer. A synthetic-data label supplies none of those factual findings.
Several differences are especially useful for the release record:
- EU and UK: document the applicable lawful basis, purpose compatibility, special-category conditions, processor arrangements and transfer route. The AI-model opinion of the European Data Protection Board (EDPB) treats anonymity as a case-specific inquiry into reasonably likely identification or extraction means. The UK's 2026 recognised-interest and purpose provisions require their own conditions. GDPR, Articles 5–6, 9, 28, 35 and 44; EDPB Opinion 28/2024, paragraphs 38–47; UK GDPR Article 8A.
- Japan: use the operative duties under the Act on the Protection of Personal Information (APPI). The principal analysis-related provisions of the amendment promulgated on 17 July 2026 have a later commencement mechanism. As at 21 November 2026, a release needs a basis under the operative APPI. Amending Act, supplementary Article 1.
- Singapore: the final 20 July 2026 generative-AI guidance of the Personal Data Protection Commission (PDPC) addresses training, publicly available data, consent notices and provider roles. Within its scope, assess accessibility and digital barriers, document an appropriate purpose, and make a training purpose sufficiently specific when relying on consent. A developer using data for its own model has different duties from a service provider acting solely for a customer. Statutory consent alternatives and other Personal Data Protection Act (PDPA) duties retain their conditions. The guidance is interpretive. PDPC guidelines, paragraphs 1.4–1.5, 3.3–3.9, 4.2–4.8 and 7.1–7.4.
- South Africa and Nigeria: purpose limitation, further use, accountability and transfers need documented treatment. Nigeria's transfer rules require recording the basis; South Africa's Protection of Personal Information Act (POPIA) can also protect information concerning existing juristic persons. Nigeria Data Protection Act, sections 24–28 and 41–43; POPIA. Nigeria's 2025 General Application and Implementation Directive (GAID), Article 43, requires controllers or processors deploying or intending to deploy AI for personal-data processing to document parameters and conduct a data protection impact assessment (DPIA), filed as part of the compliance audit return. Its conditions also address disparate outcomes, vulnerable subjects, testing, retesting and continuing monitoring. GAID, Articles 43–45.
- Brazil and Mexico: apply the privacy framework to identifiable outputs, including the applicable source, purpose, notice and transfer conditions. Brazil's research-body grounds require the relevant status. Mexico's replacement 2025 private-data law and its regulator should govern the review. LGPD; Mexico's law and reform record.
- Australia and New Zealand: assess the actual collection and secondary-use route. New Zealand's indirect-collection notification principle IPP3A has applied since 1 May 2026. An exception from that notification duty leaves IPP10's separate fair-and-reasonable-use condition for publicly sourced information to be assessed. OAIC training guidance; IPP3A; IPP10.
For US and Canadian operations, identify the relevant states, provinces and sectors. For UAE operations, establish the federal, DIFC or ADGM privacy regime from the actual connections. For Saudi Arabia and Singapore, complete the applicable statutory-basis, purpose and transfer assessment for the chosen data and recipients. These decisions should precede the supply of personal-data-bearing corpora to another trainer.
Test the accepted dataset and released student for protected-content reproduction, identifying information and confidential leakage. Record what the tests cover, their limitations and the action taken on findings. Distributed weights and an internal access-controlled model expose different practical means of querying or extracting material. The documented test results should support the specific assurance given to customers.
Use a decision tree from teacher access to release
- Map teacher and student. Next: generation authorised?
- Generation authorised? No: get permission or change route. Yes: training and sale permitted?
- Training and sale permitted? Unclear or no: exception route established? Yes: record basis and conditions.
- Exception route established? No: get permission or change route. Yes: record basis and conditions.
- Record basis and conditions. Next: content and data cleared?
- Content and data cleared? No: remove data or obtain rights. Yes: release conditions met?
- Remove data or obtain rights. Next: content and data cleared?
- Release conditions met? No: resolve conditions and retest. Yes: release within approved scope.
- Resolve conditions and retest. Next: release conditions met?
Apply both permission gates to the supplier's generation and the buyer's use when acquiring a response corpus. Identify the actors and territories before selecting a statutory exception. A service's governing-law clause and the law applicable to a local copying or data-processing act can produce different questions.
Ask for permission that covers the actual commercial product
An effective approval should come from the party with contractual authority or the relevant rights holder, through the agreement's required amendment process. The request should resolve the clause that creates the obstacle and take effect before the restricted activity.
We request written approval for the identified customer and authorised processors to generate and retain responses from the specified teacher, account and service plan, and use the approved dataset to train, fine-tune and evaluate the identified student. The attached schedule describes the task, processing locations, intended customers and delivery as a hosted service, distributed weights, or both.
Please confirm how the specified competing-product restriction applies and approve the described activity where required. The approval should identify the provisions it overrides and the applicable versions.
Please confirm permitted use and transfer of raw responses, curated datasets, adapters, checkpoints and final weights; the rights of the listed affiliates, contractors and customers; further training and replacement models; and rights surviving termination.
Please identify available provenance evidence, third-party and customer-content exclusions, data-processing conditions, and the agreed claim-handling, replacement, retraining and withdrawal obligations.
Attach a short model and data schedule. A permission for internal evaluation can leave a commercial student outside the grant; approval for a hosted service can leave weight distribution unresolved. Spell out the acts needed by the actual destination product.
- Training and competition. Point to settle: Named teacher/student, authorised task, competing-service treatment and priority over conflicting boilerplate.
- Data and resale. Point to settle: Generation, retention, annotation, processors, customer delivery, raw-corpus resale and sublicensing.
- Model evolution. Point to settle: Adapters, checkpoints, replacement students, new teacher versions and further training.
- Provenance. Point to settle: Teacher account and accepted terms, source records, lawful acquisition, third-party rights and access to supporting evidence.
- Privacy and confidence. Point to settle: Permitted inputs and purposes, roles, recipients, safeguards, notice/consent obligations and transfer arrangements.
- Downstream claims. Point to settle: Evidence-backed warranties, exclusions, claim control, indemnities, caps and remediation cooperation.
- Exit. Point to settle: Surviving use of existing students, data retention/deletion, customer continuity and a practical response to defective rights.
Check indemnity language against the actual student claim. Output protection can have exclusions for transformations, combinations, missing input rights or fine-tuning. The business needs an allocation that fits its development and release, together with the ability to provide the promised evidence. OpenAI Service Terms, updated 13 November 2026, section 1; AWS Service Terms, section 50.10.
For broader dataset grant structures, use the accompanying guide to AI data licences for training, fine-tuning and RAG. The distillation-specific addition is the linked permission for teacher access, output generation and the competing student product.
Assemble the release evidence dossier
The release approver should be able to trace the chosen student version to authorised models, data and acts. Give each record an owner and identify what would require reassessment.
- Contract and permission file. Responsible function: Legal and procurement. Evidence it should contain: Actual parties and account; accepted versions; incorporated policies; enterprise priority clauses; required written approvals.
- Teacher and student register. Responsible function: Engineering. Evidence it should contain: Exact weights/code/model identifiers; licence and commit; upstream base conditions; supported managed pair and region.
- Corpus manifest. Responsible function: Data and supplier teams. Evidence it should contain: Source IDs; generation dates and account; prompts and inputs; supplier lineage; transformations; accepted and excluded portions.
- Rights assessment. Responsible function: Legal and rights reviewer. Evidence it should contain: Licence scope or elements of the particular exception; reservations; database rights; permitted recipients and release form.
- Privacy and confidentiality record. Responsible function: Privacy and security owners. Evidence it should contain: Roles, purposes and basis; notices and required consents; processors and transfers; secret-information permissions; retention and access controls.
- Test and remediation record. Responsible function: Engineering and product. Evidence it should contain: Reproduction and leakage findings; test scope; excluded data; changed configurations; retraining or replacement decisions.
- Commercial release terms. Responsible function: Commercial and legal teams. Evidence it should contain: Required notices, attribution and model naming; accurate warranties; actual indemnity coverage; customer restrictions and exit cooperation.
- Version-specific release decision. Responsible function: Accountable release owner. Evidence it should contain: Permitted product and territories; satisfied conditions; unresolved scope excluded from release; refresh triggers.
The AI training-data supply-chain guide explains the wider upstream evidence chain. Where responses contain code, the separate guide to selling AI-generated code addresses code-output release concerns.
An EU-facing student needs an AI Act classification check. For a general-purpose AI model, Article 53 can require documentation, downstream information, a copyright policy and a training-content summary. The qualifying open-source exception concerns the first two duties; copyright-policy and summary duties remain. Chapter V generally applies from 2 August 2025, while models placed on the market before that date have the Article 111(3) transition until 2 August 2027. Assess a smaller task-specific model against the actual definition and capabilities. AI Act, Articles 3(63), 53, 111 and 113.
A separate rule can apply when the student is supplied as an in-scope AI system generating synthetic text, audio, images or video. Article 50(2) requires machine-readable output marking and detectability, subject to technical feasibility and its exceptions, including standard editing assistance or no substantial change to the input or its semantics. Specialised generators can fall within this rule; being a student classifier alone supplies no trigger. General application began on 2 August 2026; the amended Article 111(4) gives systems placed on the market before that date until 2 December 2026. AI Act, Article 50(2); Regulation 2026/1744, Article 1(39)–(40).
Make the route decision against the evidence
An own-teacher route becomes workable when upstream rights and permitted data reuse are established. An open-model route needs the exact licence and release conditions. A managed route needs a supported pair and a contractual deployment path. A restricted external API needs a use within its grant, effective permission, or a fully supported statutory and contract analysis. A purchased corpus needs evidence reaching back through the seller's generation process.
When a decisive grant is absent, seek the identified permission, change the teacher or data, or narrow the product. When a claim depends on an exception, record each condition and its factual support. When confidential or personal information lacks a valid reuse basis, remove or separately authorise the affected material before training or release. Those choices give the business a documented basis for the product it actually intends to sell.
Prepare for a scoped contract review
For a scoped contract review, assemble the teacher and cloud terms, dataset licence, model and data map, approvals and intended customer commitments. Licentium's AI supplier-contract work addresses the contractual chain, data terms and procurement evidence needed to align those documents.