Contracts with AI vendors: output ownership, indemnities and data

Contracts with AI vendors: output ownership, indemnities and data

by

An AI vendor contract decides three IP questions before anyone writes a prompt: who owns what the system produces, who pays if that output infringes someone else’s rights, and what the vendor may do with your data. Standard terms answer all three, usually in the vendor’s favour and with exclusions that matter in practice. This guide covers the AI vendor contract IP clauses that in-house counsel, procurement teams and founders in Europe and Latin America should check before signing.

Key takeaways

  • “You own the output” clauses transfer only what the vendor has. If the law grants no copyright in purely machine-made content, there may be nothing to own.
  • Output indemnities usually exclude modified outputs, disabled safety filters and, in some terms, trademark claims.
  • Check whether the vendor may use your inputs and outputs to train or improve its models, and write the answer into the contract.
  • The EU’s own model AI procurement clauses leave intellectual property and liability to the rest of the contract, so you must draft them.
  • For general-purpose AI models, the EU AI Act requires a copyright policy and a public training-data summary: ask for both.

Who owns what an AI tool generates?

Most business terms assign output to the customer. OpenAI’s Services Agreement (effective 1 January 2026) is a typical example: as between the parties, and “to the extent permitted by applicable law”, the customer owns the output, and OpenAI assigns its rights in it, “if any”. The same agreement warns that output may not be unique and that similar content generated for other users is not the customer’s output.

That wording is honest about a legal gap. Under Article 5 of the Spanish Intellectual Property Act, an author is a natural person. In Germany, the Munich Local Court held on 13 February 2026 (142 C 9786/25) that three logos generated through detailed, iterative prompts were not protected, because the decisive creative choices were made by the software. A contract can allocate whatever rights exist between the parties; it cannot create copyright that the law does not recognise, and it does not bind third parties.

In practice, that means two things. First, if output is commercially important (a logo, product copy, code), plan for human creative input and keep records of it. Second, protect key assets through rights that do not depend on authorship, such as trademark registration.

Which clauses matter most in an AI vendor contract?

Clause What to look for What to ask for
Output ownership Assignment “if any” and “as between the parties”; similar outputs for other users Assignment of all rights the vendor holds, plus a licence to anything it cannot assign
Inputs and confidentiality Whether prompts, files and outputs are treated as your confidential information Confidentiality covering inputs and outputs, and deletion on exit
Use of your data for training Opt-in or opt-out; differences between consumer and business plans No training or service improvement without your written consent
IP indemnity Whether it covers the service only or also the output; exclusions; liability cap Output cover, defence of claims, and a cap that reflects the real exposure
Fine-tuned models and data sets Who owns custom models, adapters and the data sets built with your content Ownership or an exclusive licence, and handover at termination
Changes to terms Vendors that can update online terms unilaterally Notice and a right to terminate if protections are reduced

How far do AI output indemnities really go?

Several vendors now promise to defend customers against IP claims over outputs, but the conditions matter more than the headline. OpenAI’s Service Terms (version of 12 June 2026) extend its indemnity for API and enterprise customers to claims that the use or distribution of output infringes third-party IP, except where:

  • the customer knew or should have known that the output was likely to infringe;
  • relevant citation, filtering or safety features were disabled, ignored or not used;
  • the output was modified, transformed or combined with other products or services;
  • the customer had no right to use the input or fine-tuning files;
  • the claim concerns trademark or related rights arising from use of the output in trade; or
  • the output comes from third-party content offered through the service.

The general indemnity in the Services Agreement, by contrast, excludes claims arising from “Customer Content”, defined as input and output. Microsoft’s Copilot Copyright Commitment (7 September 2023) covers paid commercial Copilot services on condition that customers use the built-in guardrails and content filters and do not try to generate infringing material.

Two consequences follow. Edited output, which is what most marketing and product teams publish, may fall outside the cover. And a logo or brand name generated with AI may face a trademark claim that an indemnity of this kind does not cover, which is why clearance searches remain essential.

What happens to your data and trade secrets?

Business terms increasingly state that customer content will not be used to improve the service unless the customer agrees; OpenAI’s Services Agreement does so in section 4.2. Consumer plans and free tiers often work differently, so check which terms your teams actually accept.

Data clauses also protect your IP indirectly. Under Article 1 of Spain’s Trade Secrets Act (Law 1/2019), information is only a trade secret if its holder has taken reasonable measures to keep it secret. Pasting confidential material into a tool without contractual confidentiality weakens that position. Where personal data is processed, Article 28 of the GDPR requires a processor contract with specific terms, including documented instructions and prior authorisation for sub-processors.

What do the EU model AI clauses add?

The EU model contractual clauses for AI procurement (MCC-AI), updated on 5 March 2025, are useful for private buyers too, mainly for what they leave out. Their commentary states that they contain no conditions on intellectual property, payment, applicable law or liability, and are meant to be annexed to a contract that covers those points. What they do include is relevant: an annex that allocates rights to each data set, an option to require handover of data sets to reduce vendor lock-in, and a supplier guarantee that use of its data sets will not infringe third-party rights. The commentary also accepts that buyers of general-purpose systems will rarely obtain rights to the vendor’s training data, and suggests asking for a description of the data types used instead.

That request now has a legal anchor. Under Article 53(1)(c) and (d) of the EU AI Act, providers of general-purpose AI models must have a copyright policy and publish a summary of their training content.

What this means for your business

  1. Inventory the AI tools in use, including free and trial accounts, and the terms behind each.
  2. Classify use cases. Output that becomes a brand, a product or published content needs stronger terms than internal drafting help.
  3. Negotiate the core clauses in the table above for any tool that touches confidential data or customer-facing output.
  4. Align your internal rules with the contract: if the indemnity requires safety filters, your teams must keep them on.
  5. Register what can be registered. A trademark gives protection that an AI-generated logo alone may not.
  6. Check governing law and formalities. In Spain, Article 45 of the Intellectual Property Act requires transfers of exploitation rights to be in writing; a Latin American subsidiary contracting locally may face different rules.

If you are rolling out AI tools across several countries, our team for AI and digital asset contracts can review vendor terms and draft the IP clauses before signature.

Where companies get AI vendor contracts wrong

  • Reading “you own the output” as a guarantee. It is an allocation between the parties, not proof of protection.
  • Assuming the indemnity covers everything. Modified outputs, disabled filters and trademark claims are common exclusions.
  • Letting each team click through its own terms, so that no one knows which tools may train on company data.
  • Forgetting the exit: fine-tuned models, prompt libraries and data sets left with the vendor.
  • Treating the contract in isolation from trademark clearance and registration for AI-generated brand assets.

Frequently asked questions

Do I own the content an AI tool creates for my company?

Contractually, most business terms assign the output to you. Legally, copyright usually requires human creative input: Spanish law defines the author as a natural person, and a Munich court refused protection to logos generated by prompting. Plan for human contribution where it matters and use registered rights such as trademarks for key assets.

Does an AI vendor’s IP indemnity cover edited outputs?

Often not. OpenAI’s Service Terms, for example, exclude output that was modified, transformed or combined with other products, and claims where safety or filtering features were disabled. Read the exclusions against how your teams actually use the output, and negotiate where your main use case falls outside the cover.

Can an AI vendor train its models on our data?

It depends on the plan and the contract. Many business terms exclude training on customer content unless the customer agrees, while consumer and free tiers may differ. Put the rule in writing, cover inputs, outputs and fine-tuning data, and make sure staff use only the accounts governed by those terms.

Can IP Global Guard review our AI vendor contracts?

Yes. We review vendor terms for ownership, indemnity and data clauses, draft the IP provisions, and align them with your trademark and copyright strategy across Europe, Latin America and Africa. Where local law requires it, we coordinate qualified local counsel, all through one point of contact.

How IP Global Guard helps you contract for AI safely

IP Global Guard, the IP services line of META Channel Corporation Limited, advises on AI contracts, copyright, trademarks and licensing in more than 25 jurisdictions across Europe, Latin America and Africa, with one strategy and one billing relationship. Within the same group, our regulatory colleagues cover the EU AI Act and the GDPR, so the IP and compliance clauses are drafted together.

Send us the AI tools you use or plan to buy, the vendor terms and what you use the output for. We will flag the clauses that need negotiating and propose wording. Contact our AI contracts team.

This article is general information, not legal advice, and does not replace a review of your specific contracts.

Sources

Share

IP Global Guard

A specialist industrial and intellectual property service line of META Channel Corporation.

IP Global Guard is a service line of META Channel Corporation Limited, a Private Company Limited by Shares incorporated in Ireland. CRO 801240. Registered office: 34 Holywell Avenue, Feltrim Hall, Swords, Dublin, K67 KF22, Ireland.

© 2026 META Channel Corporation Limited. All rights reserved.