AI Procurement Copyright Compliance Checklist: 24 Questions to Ask Before Buying Generative AI in 2026

Buying generative AI in 2026 is no longer just a security review, a pricing negotiation, or a product experiment. It is a copyright risk decision. The model may have been trained on books, music, images, code, news archives, customer data, stock libraries, or scraped websites. The outputs may be useful, but they may also be too close to protected works, incompatible with open-source licenses, or impossible to defend if a customer, artist, publisher, record label, or regulator asks hard questions.

The mistake many companies still make is treating AI procurement like SaaS procurement with a few extra privacy questions. That is too thin. Copyright risk sits in three places at once: the training pipeline, the output pipeline, and the contract allocation between vendor and customer. If procurement only asks, “Do you indemnify us?” or “Is your model trained legally?” the company learns almost nothing. A serious review has to force specifics: what data categories were used, what opt-outs were honored, what retrieval sources are connected, what filters exist, what records are kept, and what happens when an output is challenged.

This guide is a practical checklist for legal, procurement, product, marketing, engineering, and compliance teams reviewing generative AI vendors in 2026. It is not a generic “ask your lawyer” explainer. It is built around the litigation and regulatory signals that now matter: the December 27, 2023 complaint in The New York Times Co. v. Microsoft Corp. and OpenAI, the June 24, 2024 music-label suits against Suno and Udio, the January 1, 2026 effective date of California AB 2013 training-data transparency duties, the EU AI Act’s general-purpose AI transparency regime, and the U.S. Copyright Office’s continuing insistence that copyright protects human authorship, not machine output.

If you already have an AI vendor contract on the table, use this as a redline worksheet. If your company has not yet bought an AI tool, use it as a gate before pilot access turns into production dependency.

Why AI procurement is a copyright issue, not only an IT issue

Traditional software procurement asks whether the tool works, whether the vendor protects data, whether uptime is acceptable, and whether the customer can exit. Generative AI procurement adds a different kind of question: whose creative works made the system valuable, and who bears the risk if the system produces something too close to those works?

That risk is not hypothetical. In The New York Times v. Microsoft and OpenAI, filed in the Southern District of New York on December 27, 2023, the Times alleged that OpenAI and Microsoft copied millions of Times articles to train large language models and that the models could output near-verbatim passages from Times content. The defendants contest those claims, but for procurement teams the lesson is immediate: if a model vendor cannot explain training sources, mitigation measures, and output controls, the customer inherits uncertainty.

The same pattern appears in music. On June 24, 2024, major record companies sued Suno and Udio in federal courts in Massachusetts and New York, alleging that the AI music services copied sound recordings at massive scale to train music-generation systems. The defendants argued in public statements that training is transformative and protected by fair use. The labels framed the systems as commercial substitutes built from unlicensed recordings. Again, a buyer does not need the final judgment to learn the procurement lesson: if a vendor’s legal position depends on aggressive fair-use arguments, customers need to know exactly how that risk is allocated.

For images and visual works, cases such as Andersen v. Stability AI Ltd., filed in January 2023 in the Northern District of California, keep pressure on model developers over training data and output similarity. For software, the GitHub Copilot litigation showed how open-source license concerns can become procurement concerns even when the output is short code snippets rather than a whole copied repository. For AI-assisted writing, the Copyright Office’s February 21, 2023 cancellation decision involving Zarya of the Dawn and its March 16, 2023 policy statement on works containing AI-generated material made one point unavoidable: companies must distinguish human-authored material from machine-generated material if they want clean ownership and registration strategy.

That is why procurement has to become evidence-based. A vendor’s marketing page saying “enterprise-safe” is not evidence. A warranty without a usable indemnity process is not enough. A general statement that “we respect copyright” does not answer whether the model was trained on licensed data, web-scraped data, user uploads, synthetic data, public-domain materials, or a mixture of all five.

The 24-question procurement checklist

1. What exact AI functions are we buying?

Start with scope. Is the vendor providing text generation, image generation, code completion, music generation, voice cloning, video generation, search summarization, retrieval-augmented generation, or all of the above? Copyright risk changes by modality.

A text summarization product connected to internal documents creates different issues from a music generator trained on sound recordings. A code assistant raises license contamination and repository provenance questions. A marketing-image generator raises stock-photo, artist-style, and publicity-right concerns. A voice product can trigger copyright-adjacent issues plus right-of-publicity and biometric-law exposure.

Procurement should force the business owner to describe the intended use case in writing. “We want AI” is not a reviewable use case. “We want to generate first-draft product descriptions from our own catalog data” is reviewable. “We want to create music beds that sound like current chart hits” is a red flag.

2. Which model or models power the product?

Many vendors wrap third-party foundation models. The company selling the interface may not control the training data, model weights, or output filters. Ask whether the product uses proprietary models, open-weight models, third-party APIs, fine-tuned models, or a routing layer that sends prompts to different model providers.

This matters because indemnity and evidence may sit with different parties. A procurement contract with a small AI workflow vendor is weak if the real model provider disclaims training-data warranties. Require a model inventory: model names, versions, providers, hosting location, and whether model selection can change without notice.

3. What data categories were used for training?

Do not accept “publicly available data” as a complete answer. Public availability is not the same as copyright permission. A copyrighted newspaper article, song lyric, photograph, or software file can be publicly accessible and still protected.

Ask for training-data categories: licensed datasets, public-domain works, user-contributed data, web-crawled data, synthetic data, government materials, academic datasets, stock media, code repositories, books, news, music, images, video, and social-media content. The goal is not always to get a perfect list of every work. The goal is to understand whether the vendor has mapped risk categories and whether your intended use collides with high-risk categories.

For a deeper framework on records, see our guide to building an AI training data audit trail.

4. Which data sources are licensed, and what do the licenses allow?

A vendor may say it uses “licensed data,” but procurement needs the next layer: licensed for what? A content license may allow indexing but not model training. It may allow internal research but not commercial deployment. It may allow text training but not image generation. It may expire or prohibit sublicensing benefits to customers.

Ask for a summary of material licenses by category. You do not need every confidential contract, but you do need enough detail to evaluate whether the vendor’s rights align with your use. If the vendor relies on publisher deals, stock-media agreements, music catalogs, or code datasets, ask whether those licenses cover training, fine-tuning, evaluation, retrieval, output commercialization, and customer indemnity.

5. Does the vendor rely on fair use, text-and-data-mining exceptions, or another legal theory?

Some vendors have licenses. Some rely partly on licenses and partly on fair use. Some rely heavily on fair use for web-scale training. That is not automatically disqualifying, but it changes the risk profile.

In U.S. law, fair use is assessed under 17 U.S.C. § 107 using four factors: purpose and character, nature of the work, amount used, and market effect. AI developers often argue that training is transformative because the model learns patterns rather than republishing works. Rightsholders respond that copying entire works to build commercial substitutes damages existing and emerging licensing markets. Courts have not yet delivered a single clean rule for all AI training.

If the vendor relies on fair use, require a written explanation of the theory and ask whether any court has tested that theory for the vendor’s model or modality. Our analysis of what courts actually look for in the AI fair use defense can help your legal team evaluate the answer.

6. Has the vendor been sued, threatened, or asked to remove training data?

Ask directly about copyright claims, demand letters, DMCA notices, dataset-removal requests, artist opt-outs, publisher complaints, and settlement obligations. Vendors may resist disclosing confidential disputes, but they can usually provide a representation that identifies material litigation and unresolved claims.

The question is not only “Have you lost a case?” In AI, pending litigation itself can affect continuity. A court order, settlement, or licensing change could require model changes, output restrictions, or price increases.

7. Does the vendor honor opt-outs and robots controls?

For web-trained models, ask whether the vendor honors robots.txt, noai, noimageai, TDM reservations, publisher opt-outs, or other machine-readable signals. Technical controls are not a complete legal defense, but they show whether the vendor has a repeatable governance process.

This matters especially in Europe. The EU Copyright Directive includes text-and-data-mining provisions, and rightsholders may reserve rights in machine-readable ways. The EU AI Act also pushes general-purpose AI providers toward documentation and copyright-policy obligations. If your company operates globally, a U.S.-only answer is not enough.

For implementation background, see our guide to EU AI Act copyright transparency requirements.

8. Can the vendor remove or suppress specific works from future training or retrieval?

A serious enterprise vendor should have a process for takedown, suppression, or exclusion. That process may not erase all influence from an already-trained model, but the vendor should be able to stop using a source in retrieval, remove it from future training runs, or apply output filters for known protected materials.

Ask what happens if a rightsholder complains that its works are in the vendor’s dataset. Who investigates? How quickly? Does the customer get notice if the complaint relates to the customer’s outputs? Is there a repeatable escalation path?

9. Are customer prompts and outputs used for training?

This is both a privacy and copyright question. If your employees upload copyrighted customer materials, licensed stock assets, unpublished manuscripts, source code, or confidential product documentation, the vendor’s right to train on that material matters.

The contract should state whether prompts, uploaded files, embeddings, fine-tuning data, feedback, and outputs are used to train or improve models. If training is optional, the default should be off for enterprise customers. If the vendor claims no training occurs, ask whether that applies to all subprocessors and model providers.

10. What output similarity controls exist?

A vendor should be able to describe controls that reduce near-copy outputs: memorization testing, deduplication, similarity filters, regurgitation detection, watermarking or provenance metadata where available, and restrictions on prompts that request living artists, commercial characters, famous songs, or verbatim text.

The controls should match the modality. Text products need quotation and long-excerpt controls. Image products need reference-image and style controls. Music products need melody, lyric, and sound-recording similarity review. Code products need license and snippet matching.

11. Can we configure blocked uses?

Enterprise buyers should not rely only on vendor defaults. Ask whether administrators can block high-risk categories: generating song lyrics in the style of named artists, recreating stock images, producing celebrity voices, generating code under incompatible licenses, or summarizing paywalled articles without authorization.

A good procurement outcome is not simply “approved” or “rejected.” It may be “approved for internal brainstorming, prohibited for final customer-facing assets unless cleared.” Procurement should make that distinction explicit.

12. Who owns the outputs?

Many vendors promise that customers “own” outputs. That promise is often narrower than it sounds. A vendor can assign whatever rights it has, but it cannot magically create copyright in material that lacks human authorship, nor can it eliminate third-party infringement claims.

The contract should separate three ideas: the vendor’s claim to the output, the customer’s right to use the output, and whether the output is copyrightable. Under U.S. Copyright Office guidance, purely machine-generated material is not protected by copyright, while human-authored selection, arrangement, editing, or contribution may be protectable. That means companies should document human contribution when outputs become important assets.

For a practical documentation process, see The Creator's Guide to Proving Human Authorship in AI-Assisted Works.

13. Does the vendor provide copyright indemnity?

Indemnity is important, but the details matter more than the headline. Ask whether the vendor indemnifies against third-party copyright claims arising from training data, outputs, or both. Many AI indemnities cover only outputs, only paid enterprise plans, only use within policy, and only after the customer enabled specific filters.

The contract should define covered claims, excluded claims, defense control, settlement approval, notice deadlines, damages caps, and whether injunctive relief or replacement services are covered. A $50,000 cap may be useless if the AI output becomes part of a national campaign.

We have a deeper clause-by-clause worksheet here: AI Vendor Contract Copyright Indemnity Checklist.

14. What conduct voids indemnity?

Vendors often exclude claims caused by customer prompts, modifications, use of uploaded reference materials, disabling filters, using outputs after receiving notice, or combining outputs with third-party materials. Some exclusions are reasonable. Others swallow the protection.

Procurement should map exclusions to real workflows. If your designers routinely upload mood boards, reference images, or competitor examples, an exclusion for “customer-provided materials” may mean the indemnity rarely applies. If your developers accept AI code suggestions into proprietary repositories, an exclusion for “modified outputs” may create uncertainty.

15. What records will the vendor provide if a claim arises?

When a copyright claim arrives, you need evidence quickly: prompts, output history, model version, filters active at the time, retrieval sources, uploaded files, user identity, timestamps, and policy settings. Without logs, the company may be unable to prove whether the accused material came from the vendor, an employee upload, a third-party plugin, or later human editing.

Ask how long logs are retained, who can access them, whether they are exportable, and whether legal holds are supported. If logs are deleted after 30 days but marketing campaigns run for two years, the retention period is too short.

16. Does the product use retrieval-augmented generation?

Retrieval-augmented generation, or RAG, can reduce some risks by grounding outputs in approved sources. It can also create new risks if the retrieval corpus includes paywalled articles, licensed databases, third-party PDFs, or scraped websites.

Ask what sources the system retrieves from, whether it quotes or paraphrases retrieved materials, whether citations are preserved, and whether the customer can restrict retrieval to owned or licensed content. If the vendor provides web search or “answer engine” features, ask how it handles publisher content and robots restrictions. News publishers have been especially aggressive in challenging AI products that summarize or substitute for articles.

17. How does the vendor handle open-source software?

For code tools, require a software-specific review. Ask whether outputs are scanned against public repositories, whether license notices are preserved, whether copyleft code can be suggested, and whether the tool can block suggestions matching GPL, AGPL, or other restricted licenses.

A code assistant that is acceptable for internal prototypes may be unacceptable for production code in a closed-source product. If the vendor says matches are rare, ask for match thresholds and reporting. If the vendor offers a “public code matching” feature, make it mandatory for production environments.

18. Are there jurisdiction-specific compliance features?

AI copyright compliance is not the same everywhere. The U.S. focuses heavily on fair use and human authorship. The EU combines copyright law with AI Act transparency obligations. The United Kingdom has debated text-and-data-mining reform while maintaining narrower exceptions. Japan has a more permissive data-analysis exception, but it is not unlimited. China has platform and algorithm governance rules plus copyright enforcement.

If your company sells globally, ask whether the vendor supports regional controls, disclosures, data processing locations, and documentation required by local law. Our country-by-country analysis of AI training and copyright in 10 jurisdictions is a useful starting point.

19. Does the vendor support output review before publication?

Copyright risk often becomes real at publication. Internal brainstorming is lower risk than public advertising, product packaging, app-store screenshots, training manuals, or paid content. Ask whether the tool has approval workflows, role-based access, audit logs, and policy gates for external publication.

Marketing teams in particular should not treat AI output as automatically cleared. They need a clearance workflow for images, slogans, scripts, music, and long-form content. If the vendor cannot support review, the customer should create the workflow internally. See our AI output copyright clearance workflow for marketing teams.

24. What is our internal go/no-go decision?

The final procurement decision should be written as a risk classification, not a vague approval. For example:

• Approved for internal research only.
• Approved for customer-facing text after human review.
• Approved for code suggestions only with public-code matching enabled.
• Approved for marketing images only when no living artist, brand, celebrity, or third-party reference image is used.
• Not approved for music, voice cloning, or video generation.
• Not approved until vendor provides training-data documentation and output indemnity.

This classification becomes the bridge between legal review and day-to-day operations. Without it, employees will assume that procurement approval means all uses are safe.

A scoring model for procurement teams

A simple scoring system can make reviews consistent. Score each vendor from 0 to 3 in five categories:

1. Training-data transparency.

2. License and legal basis.

3. Output controls.

4. Contract protection.

5. Operational evidence and logs.

A score of 0 means the vendor gives no meaningful answer. A score of 1 means the vendor gives general assurances. A score of 2 means the vendor provides specific but incomplete documentation. A score of 3 means the vendor provides specific documentation, contractual commitments, configurable controls, and evidence support.

For low-risk internal brainstorming, you might accept a total score of 8 with restrictions. For customer-facing marketing assets, production code, music, video, or large-scale publishing, a weak score should block approval. The point is not to create false precision. The point is to prevent a charismatic demo from overwhelming legal reality.

What to put in the contract

At minimum, an enterprise AI procurement contract should include:

• A clear description of approved use cases.
• A warranty that the vendor has sufficient rights to provide the service.
• A warranty that customer data will not be used for training unless expressly agreed.
• A copyright indemnity covering outputs, with narrow and understandable exclusions.
• Cooperation duties for claims, takedowns, and evidence preservation.
• Audit-log retention commitments.
• Notice duties for model, data, subprocessor, policy, and indemnity changes.
• Configurable controls for high-risk outputs.
• Deletion and export rights.
• A right to suspend or terminate if legal risk materially changes.

Do not let the contract live separately from the product settings. If indemnity requires filters to be enabled, the contract should say who enables them and how that setting is verified. If the vendor excludes reference-image uploads, the admin console should restrict reference-image uploads for users who do not need them.

The bottom line

In 2026, AI procurement is copyright governance. The company that buys generative AI without asking training-data, output, indemnity, and logging questions is not moving fast; it is accepting unknown legal debt.

The best procurement process does not try to answer every unresolved question in AI copyright law. Courts and regulators will keep refining the rules. Instead, the process forces the vendor and the business team to make risk visible before deployment. What data was used? What rights support that use? What outputs are blocked? What records exist? Who pays if the output is challenged? What uses are actually approved?

If a vendor can answer those questions with documents, controls, and contract commitments, it may be ready for production. If it can only answer with slogans, keep the pilot small, keep the outputs internal, and keep looking.