Last materially reviewed: September 6, 2026
Intellectual Property → AI-Generated Works → AI Training & Enforcement
Direct Answer
Permission is safest when AI training involves copyrighted works and no clear exception or other legal basis covers the intended use. A license can authorize collection, copying, preprocessing, training, fine-tuning, evaluation and sometimes output uses that would otherwise remain uncertain.
Not every dataset requires a paid license. Public-domain works, properly licensed open content, factual data and uses that fall within a statutory exception can be different. The point is to identify the legal basis rather than assume all data is either “free” or “restricted.”
Legal Status
Established contract tools; developing AI application. Philippine copyright owners can license and assign rights. AI-specific clauses are newer, but ordinary licensing principles remain useful.
What an AI Training License Should Define
- Exactly which works or datasets are covered.
- Whether use is allowed for pre-training, fine-tuning, retrieval, evaluation or embeddings.
- Whether the license is exclusive or non-exclusive.
- Territory, duration and model versions covered.
- Whether derivative datasets may be created.
- Retention and deletion obligations.
- Restrictions on model outputs that reproduce source material.
- Attribution, metadata and provenance obligations.
- Compensation, royalties or usage fees.
- Audit, reporting and compliance rights.
- Indemnities, warranties and liability caps.
Licensing Is Not the Same as Buying the File
Possession of a digital copy does not automatically transfer copyright. A dataset vendor may have technical access to content without holding the rights needed to sublicense model training. Buyers should verify the chain of rights.
Dataset Vendor Due Diligence
Ask where the data came from, whether it was scraped, licensed, contributed by users, generated internally or obtained from public-domain sources. Require the vendor to explain what rights it actually grants and what happens if a rights holder objects.
See Dataset Provenance and AI Copyright.
Rights-Holder Licensing Models
Creators and publishers can consider flat fees, per-work fees, revenue shares, model-specific licenses, research-only licenses, enterprise licenses or collective arrangements. The commercial model should match the value of the repertoire and the level of control the rights holder needs.
When a License May Not Be Necessary
A license may not be required where the material is not protected, the copyright has expired, the use is already authorized by an open license, the owner has clearly granted the relevant rights, or a statutory exception applies. Because AI training is still developing, businesses should document the specific basis they rely on.
Verify the Rights Before Negotiating the Price
A dataset vendor must have authority to grant the relevant permissions. Identify the original owners, assignments, contributor agreements and sublicensing rights. Ownership of a copy and copyright are distinct; a transfer of the copy alone does not necessarily transfer copyright. IP Code, Sections 178 and 181.
Section 177 identifies economic rights, including reproduction and transformation. A grant should address the acts the project actually requires. Copyright assignments and licenses also need to be considered under the amended Code. RA 10372, amendments to Section 180. These statutory references do not replace review of the particular agreement.
A Dataset and Use Schedule
| Schedule field | Question to answer |
|---|---|
| Covered material | Which works, versions, dates and excluded assets? |
| Permitted acts | Collection, storage, preprocessing, training, evaluation or other specified uses? |
| Recipients | Which entities, vendors and subcontractors may receive copies? |
| Commercial scope | Which products, models, territories and business uses? |
| Duration | When may collection and training begin and end? |
| Post-termination treatment | What happens to datasets, backups and already-developed models? |
| Evidence | What records and reports demonstrate compliance? |
This is a negotiation aid, not a mandatory government form. Spell out terms whose everyday meanings may differ between the parties. Permission to query an information service should not automatically be treated as permission to retrain a model on the entire archive.
Three Hypothetical Licensing Deals
A Publisher Licenses Only Its Own Articles
The archive also contains freelance photographs and syndicated stories. List exclusions or secure the necessary additional permissions. A promise covering the whole archive is unreliable if the publisher’s contributor agreements do not support it. Verify a representative sample and investigate discrepancies before expanding access.
A Research License Is Reused for a Paid Product
A team first receives material for a defined study and later proposes commercial deployment. Compare the new purpose with the grant. If it exceeds the agreed scope, obtain an amendment before relying on that permission. Do not treat the earlier delivery of files as consent to every future use.
A Vendor Promises Deletion at Termination
The buyer and seller need to define which records are deleted and whether any model-related action is required. Deleting source files and changing an already-trained model are different commitments. Ask technical staff to assess feasibility before signing; do not promise an outcome merely because a contract draft uses the word deletion.
Documents to Request From a Dataset Vendor
Request the material inventory, provenance explanation, rights-granting documents or appropriate evidence of them, relevant restrictions and the list of excluded works. Ask whether contributors permitted sublicensing and whether third-party assets remain embedded. Keep the exact license version and correspondence clarifying scope.
Use a risk-appropriate review. Sensitive contracts may need controlled disclosure, but confidentiality should not be replaced by an unsupported assurance of complete rights. If evidence is insufficient, narrow the dataset or defer use of the uncertain subset. See dataset provenance for traceability and the business due-diligence checklist.
Allocate Problems the Parties Can Actually Manage
Define who investigates a complaint, preserves records, communicates with claimants and decides whether to suspend a source. Warranties and indemnities should match the grant, with review of exclusions, caps, notice requirements and control of the defense. An indemnity does not erase a third party’s rights or guarantee recovery from the vendor.
Compensation can use different commercial structures, but specify how usage is measured and audited. Do not invent a standard Philippine training royalty. A negotiated price depends on the repertoire, permitted scope and parties’ agreement.
Before Signing and Using the Data
- Confirm who can grant the rights.
- List included and excluded works.
- Match the grant to the actual pipeline and recipients.
- Resolve retention, outputs and post-termination obligations.
- Review confidentiality and personal-data duties separately.
- Have legal and technical reviewers examine material uncertainties.
- Retain the signed version and link it to the approved dataset.
If the grant does not cover the use, do not assume either infringement or fair use without analysis. Use the fair-use guide for that separate inquiry. The training-material overview explains other bases for use. A lawful dataset license still does not guarantee that every output can be published without review.
Frequently Asked Questions
Can a website’s terms grant an AI training license?
Potentially, yes, if the terms form a valid agreement and clearly grant the relevant rights. Scope and assent still matter.
Can an AI company license only part of a publisher’s archive?
Yes. The agreement can define works, dates, formats, territories, uses and model families precisely.
Should a license address model memorization?
Yes. Rights holders may want testing, output safeguards and remedies where models reproduce protected material.
Official Sources
- IPOPHL — Copyright, transfers and licensing
- Republic Act No. 8293 (our IP Code explainer)
- WIPO — Generative AI: Navigating Intellectual Property
- WIPO — transparency, consent and compensation for AI uses
Important: This article provides general educational information about Philippine law and technology. It is not legal advice and does not create an attorney-client relationship. Laws, agency procedures, platform terms and the facts of each situation may change the result. Verify current requirements through the cited official sources and seek qualified professional advice when your rights, deadlines, money or legal exposure may be affected. Contract wording should be tailored to the actual dataset and model use.
Featured image: Photo by Vitaly Gariev via Unsplash.