From Shubhi K | Product & Market Analysis
AI Copyright Indemnity Compared: 4 Vendors, 8 Exclusions, and the Edit Problem
On this page
Microsoft, Google, OpenAI and Anthropic all promise to defend paying customers against intellectual property claims over AI output. Then 3 of the 4 narrow that promise once a person edits the output, which is what most businesses do before publishing anything. I read the current AI copyright indemnification terms and scored them on 8 exclusions. Microsoft carried the fewest, at 4. Google carried the most, at 7.
Key takeaways
- Editing is the fault line in AI copyright indemnity. Google covers only unmodified output, while OpenAI and Anthropic exclude modified or combined output. Microsoft's commitment has no general modification exclusion.
- Trademark is outside all 4 output indemnities. Microsoft, Google, OpenAI and Anthropic each exclude trademark claims on output, so a generated logo or product name is your own risk.
- The buying route can change who defends you. Azure OpenAI output sits under Microsoft's commitment rather than OpenAI's terms, and Google's output cover is written around Google's own models.
- Anthropic is the only one of the 4 with no safety-filter condition. Microsoft requires named filters on Azure OpenAI text and code, while Google and OpenAI void cover if you bypass citations or filters.
AI copyright indemnity: the short answer
AI copyright indemnification is a vendor's contractual promise to defend you and pay the result if a third party sues over its service or output. All 4 vendors offer it to paying business customers. None offers it on free consumer plans, and the exclusions, not the headline promise, decide what you actually have.
I scored the terms on 8 tests. Each test asks one thing: does this vendor's text remove cover in this situation? A count is a blunt tool, because the exclusions are not equal in weight. It is still more useful than the marketing, where all 4 products read as protected.
This post is about the whole product line of each vendor. For the code-specific version, including Cursor and GitHub Copilot, see the clause comparison for AI coding assistants. For retention, training rights and uptime alongside indemnity, see the model provider terms comparison.
Microsoft, Google, OpenAI, Anthropic: which outputs each indemnification covers
Scope is narrower than the announcement headlines suggest, and each vendor draws the line differently. Microsoft ties its cover to a defined list of products. Google publishes a list too, and limits it to paid use. OpenAI and Anthropic write it into their business contracts, so your plan decides which contract you are on.
One structural detail deserves attention in OpenAI's paper. The general indemnity in its Services Agreement excludes claims arising from Customer Content, and commentary on the January 2026 version reads that term as covering both inputs and outputs. The promise about Output lives in the separate Service Terms. If your lawyer reviews only the master agreement, they will see the wrong answer.
| Vendor | Covered | Named as outside | Training data named? |
|---|---|---|---|
| Microsoft | Paid Azure OpenAI models in Foundry and any Copilot, under the Customer Copyright Commitment. | Free previews. Externally hosted models in Copilot Studio unless run in Azure OpenAI with the mitigations. | Not named. Cover is framed around use or distribution of output. |
| Paid, listed services using Google models, including generally available Gemini, Imagen and Veo through its agent platform API, Gemini Enterprise and Gemini in Workspace. | Free services and pre-release offerings. Output from models that are not Google's. | Yes, as a separate indemnity that excludes claims about a specific output. | |
| OpenAI | API and ChatGPT Enterprise customers, for use or distribution of Output. Trackers also list Edu and Healthcare. | Free and Plus plans. Beta services. Output from a Third Party Offering. | Not named in the Output clause I could read. |
| Anthropic | Paid use of services under the Commercial Terms, and Outputs from authorised use. | Consumer plans such as Claude.ai, which run on separate terms. | Yes, in the definition of the claim. |
The training data column matters more than it looks. The training data cases attack how a model was built, while the Disney and Midjourney output case attacks what a model produced. A customer is far more exposed to the second kind. Google and Anthropic are explicit about the first. I would still treat both as secondary to the output clause when you negotiate.
The edited output problem in AI indemnification
Most copyright indemnity comparisons list exclusions and stop. The issue that should change your buying decision is a single sentence about editing, because unedited output is rarely what a business ships. A marketing team rewrites the draft. A designer adjusts the image. A developer merges the suggestion into an existing file.
How Google, OpenAI and Anthropic treat edits
Google's generative AI terms apply to unmodified Generated Output from an indemnified service. OpenAI's exclusion is wider in its verbs: cover does not apply where Output was "modified, transformed, or used in combination with products or services" not provided by OpenAI. Anthropic excludes claims arising from "modifications made by Customer to the Services or Outputs" and from combination with outside technology.
The wording differs in a way that matters. Google's phrase is a precondition on what is covered at all. Anthropic's says the exclusion applies where the claim arises from your modification, which leaves room to argue that an infringing passage you left untouched is still covered. OpenAI's clause, read plainly, is the least forgiving of the 3.
Why Microsoft is the outlier
Microsoft's Product Terms have no general modification exclusion. The relevant condition says the customer must not "modify, use, or distribute the Output Content in a manner that it knows, or should know" is likely to infringe. That is a test of your state of mind, not of whether anyone touched the text.
My read is that this is the most useful sentence in any of the 4 contracts for a team that edits its output. It is also the sentence most likely to be tested on the facts, because "should know" has no bright line. Ask your counsel whether your editing workflow creates knowledge, and how much human editing affects copyright in the result is a related question worth settling at the same time.
Required safeguards: what you must do to keep the indemnity
Indemnity is conditional on behaviour, and the conditions range from none at all to a named configuration. This is where a contract promise turns into an engineering task. The safeguards also determine what evidence you will need, because Microsoft says a customer tendering a claim must demonstrate compliance with the relevant requirements.
| Vendor | Required | Where it bites |
|---|---|---|
| Microsoft, Azure OpenAI text | Metaprompt directing the model not to infringe. Protected material text filter on in filter mode. Prompt Shield for jailbreaks on in filter mode. | Output flagged retroactively by an asynchronous filter is not covered. |
| Microsoft, Azure OpenAI code | Metaprompt. Protected material code filter in annotate or filter mode. Prompt Shield in filter mode. | In annotate mode you must comply with any cited licence. |
| Microsoft, GitHub Copilot | No additional required mitigations from 3 April 2026. | The Duplicate Detection filter is no longer required for cover. |
| Do not disregard, disable, modify or circumvent citations, filters, instructions or other tools. | You are solely responsible for changing safety filter settings. Stop using output after an infringement notice. | |
| OpenAI | Do not disable, ignore or fail to use relevant citation, filtering or safety features. | "Did not use" reaches features you never turned on. |
| Anthropic | No filter or citation condition in the Commercial Terms. | Use you know or should know infringes is still excluded. |
Microsoft asks the most from API builders
Azure OpenAI is the only offering with a filter specification written down by name. If your application lowered a protected material setting to reduce false positives, you may have left the commitment without noticing. Microsoft does say customers have no obligation to conduct direct testing of Microsoft services to keep cover, which removes one common worry.
OpenAI's "did not use" clause is the sleeper
Exclusion (ii) in OpenAI's Service Terms covers end users who "disabled, ignored, or did not use" relevant features. A clause that voids cover for not using a feature creates a duty you may never have been told about. I would ask OpenAI to list the relevant features for your deployment in writing, and I would not sign an enterprise renewal without that list.
The buying route decides who defends AI output
Comparing 4 vendors assumes you buy from one of them directly. Many enterprises buy through a cloud marketplace instead, and that changes which contract applies. The same model can have a different defender, or none that I could confirm, depending on the invoice.
AWS does not stand behind Claude
AWS keeps a closed list of indemnified generative AI services, including Amazon Nova and Titan models, Amazon Q and Kiro outside its free tier. Third-party models are "Third-Party Content", and no Claude model is on the list. AWS's third-party model page says Anthropic services on Bedrock are sold by Anthropic. The Anthropic on Bedrock terms, version 18 August 2025, carry the same defence clause and the same exclusions as the direct contract.
That is a reasonable result for a buyer. It does mean your defender is Anthropic, regardless of how large your AWS relationship is. Your AWS account team cannot negotiate that clause.
Claude on Google Cloud is the open question
Google's output indemnity applies to output "using only Google Pre-Trained Model(s), a Fine-Tuned Google Model, or a Customer Adapter Model". A Claude model on that platform is none of those. I could not find a current Anthropic contract for that route that answers the question, so treat it as unconfirmed and get the answer from your reseller in writing.
AI indemnification caps and who controls the defence
The general liability cap at Google, OpenAI and Anthropic is 12 months of fees, and all 3 place their indemnity obligations outside that cap. Google's Section 12.3 says nothing in the agreement limits its obligations under Section 13. Anthropic's Section L.3.b says the same of Section K, and OpenAI's limitation clause carves out indemnification obligations.
OpenAI's Services Agreement goes one step further. It says the Service-Specific Terms Indemnity is "not subject to any liability cap" and that OpenAI may not materially reduce those protections without your written agreement. That second promise is rare, and it is worth more to a long-term customer than the headline.
Microsoft's commitment states neither a cap nor a carve-out. Its text operates under your volume licensing agreement, so the answer sits in a document that Microsoft serves behind a form and that I could not open for this review. Microsoft's 2023 announcement said it would "pay the amount of any adverse judgments or settlements", which is a statement of intent rather than a contract term.
The vendor runs your case
Uncapped does not mean unconditional. Google requires prompt written notice and sole control of the defence. Anthropic keeps control too, though you may refuse a settlement that requires you to admit wrongdoing or take on an ongoing obligation. Late notice can excuse Anthropic if it materially prejudices the defence. Google, OpenAI and Anthropic each make their indemnity the only remedy for covered third-party IP claims, which limits what you can recover beyond it. How caps and remedies work on agent contracts shows the same pattern outside copyright.
Where this comparison is weakest
The 8-test count is mine, and it is unweighted. A trademark exclusion and a patent exclusion do not cost the same in your business. A reader who generates product names should weigh row 1 far above the others, and a team that never edits output can ignore rows 5 and 6.
What I could not read directly
OpenAI's site refused automated access, so the OpenAI clauses here come from text returned by a search of openai.com and from public trackers, not from a page I opened. Google's generative AI section reached me through automated extraction, not a clean copy. Microsoft's master agreement was unavailable. Anthropic's Commercial Terms and the AWS terms were read directly. Check the live documents before you rely on any quoted wording, and treat this as analysis rather than legal advice.
The case that none of this matters much
I found no public report of a customer tendering an output claim under any of these indemnities and having a vendor defend it, though a search cannot prove none exists. The promises are close to untested. Most lawsuits so far target the model makers rather than their customers, which suggests the practical value of these clauses may be small. Negotiated enterprise paper can also override everything above. I would not choose a model vendor on indemnity alone, because a clause that has never been enforced is a weak reason to accept a worse product.
Frequently asked questions
Does Microsoft indemnify Copilot output for copyright claims?
Yes, for paid Copilots and Azure OpenAI models, under the Customer Copyright Commitment in Microsoft's Product Terms. Microsoft defends claims based on your use or distribution of output if you kept its safety systems on, had rights to your inputs and did not use output you knew or should have known was likely to infringe. Trademark claims are excluded. Azure OpenAI customers must also switch on specified filters.
Does OpenAI indemnify ChatGPT and API output?
OpenAI's Service Terms extend its indemnity to API and ChatGPT Enterprise customers for claims that using or distributing Output infringes third-party intellectual property. Free and Plus users get no such cover. The indemnity falls away if you modified or combined the output, ignored citation or filtering features, lacked rights to inputs, or should have known the output was likely infringing. Beta services are excluded.
Is Claude covered by Anthropic's copyright indemnity?
Paid use under Anthropic's Commercial Terms is covered, including claims about training data and outputs. Consumer Claude plans are not. Anthropic excludes output you modified or combined with other technology, inputs you supplied, knowing infringement, patented inventions contained in an output and trademark use in commerce. Claude bought through Amazon Bedrock runs under Anthropic's separate Bedrock terms, which carry the same defence clause.
Does Google indemnify Gemini generated output?
Google indemnifies unmodified output from paid, listed generative AI services using its own models, including generally available Gemini, Imagen and Veo through its agent platform API. A separate clause covers training data. Cover is lost if you knew the output was likely infringing, disregarded citations or filters, kept using it after an infringement notice, or lacked rights to customisation data. Pre-release offerings are excluded.
Is AI copyright indemnification capped?
Google, OpenAI and Anthropic each limit general liability to 12 months of fees, and each places its indemnity outside that limit. Microsoft's commitment text states no cap and no carve-out, so the answer sits in your volume licensing agreement, which I could not open for this comparison. A negotiated enterprise contract can change any of these positions, so check your signed paper.
Does AI indemnification cover edited output?
It depends on the vendor. Google limits output cover to unmodified output. OpenAI excludes output that was modified, transformed or combined with other products. Anthropic excludes claims arising from customer modifications and combinations. Microsoft's commitment has no general modification exclusion and instead asks whether you knew or should have known the use was likely to infringe. Anyone editing output should ask in writing how far an edit moves a claim.
Where to start this week
Pull your AI invoices and write down, for each model in production, the entity you pay and the contract that names it. That one page will show any route where you assumed a vendor stood behind you and it does not. The Claude on Google Cloud route is the first one to check.
Then send one question to each vendor's account team: "Does our edit and publishing workflow keep this output inside your indemnity, and which filters must stay on?" Keep the answer on file with a dated export of your filter settings. If a claim ever arrives, that file is what you tender. Clause language for the renewal itself is in the AI contract clauses a CFO should ask for.
References
- Microsoft, Product Terms, Online Services, Customer Copyright Commitment, read 9 November 2026. Used for covered products and conditions.
- Microsoft Learn, Customer Copyright Commitment required mitigations, updated 13 July 2026. Used for filter requirements and GitHub Copilot change.
- Microsoft, Copilot Copyright Commitment announcement, 7 September 2023. Used for the payment statement.
- Google Cloud, Service Specific Terms and Generative AI Indemnified Services, last modified 2 April 2026. Used for output and training data indemnity.
- Google Cloud, Platform Terms of Service, sections 12 and 13, read 9 November 2026. Used for the cap and procedure.
- OpenAI, Service Terms and Services Agreement. Wording from search of openai.com, November 2026; the pages could not be opened directly.
- Anthropic, Commercial Terms of Service, effective 17 June 2025. Used for Sections K and L.
- AWS, Service Terms, updated 5 November 2026, and third-party model terms. Used for the Bedrock route.
The weakest thing about this source base: 2 of the 4 vendors' key clauses were not read from a clean copy of the page, and Microsoft's master agreement was unavailable. Terms change without notice. All were last checked on 9 November 2026. This is analysis, not legal advice.
Related reading