From Ritu Raj | Product & Market Analysis

AI Copyright Indemnity Compared: 4 Vendors, 8 Output Types, 2 Clean Answers

On this page

I tested 8 output scenarios against the published terms of Microsoft, Google, OpenAI and Anthropic, and only 2 of the 32 results needed no safeguard from you. Both belong to Anthropic, and both narrow once a person edits the text. AI copyright indemnification is now standard on paid plans. What it pays for is far narrower than the announcements suggest.

Key takeaways

  • Editing is the fault line in AI copyright indemnity. Google covers only unmodified output, OpenAI excludes modified output, and Anthropic excludes claims to the extent your edits caused them. Microsoft's commitment instead asks whether you knew an edit was likely to infringe.
  • Microsoft asks for the most evidence. Azure OpenAI text and code need 4 named safeguards, including a retained testing report. You must be able to demonstrate compliance when you tender a claim.
  • 3 of 4 vendors state the indemnity sits outside the fee cap. Google, OpenAI and Anthropic cap general liability at 12 months of fees and carve indemnity out. Microsoft's commitment names no figure, and its cap sits in your licensing agreement.
  • The same model can carry different promises. Claude inside Microsoft Copilot gets Microsoft's commitment. Claude bought through the Microsoft Foundry catalogue runs on Anthropic's terms, and Google's output indemnity is limited to Google's own models.
2 of 32Output scenarios where cover depends on no safeguard. Both are Anthropic text and code. Source: Zan Digital reading of 4 vendors' published terms, November 2026.
3 of 4Vendors that remove or reduce cover for output you edited. Source: Google, Anthropic and OpenAI terms, 2026.
6 monthsTime Microsoft gives you to adopt a newly published safeguard before cover depends on it. Source: Microsoft Learn, July 2026.

AI copyright indemnity: the short answer

Which AI vendor offers the best copyright indemnity? None wins on every clause. Google, OpenAI and Anthropic state that the indemnity sits outside the 12-month fee cap. Anthropic imposes no safeguard setting. Microsoft alone does not exclude edited output as such. Your answer depends on whether you edit output, and which model you run.

This is written for the founder or director who signs the AI contract. That person inherits the gap between the marketing name and the clause when a claim letter arrives.

The product names mislead in a specific way. Microsoft calls its promise a Customer Copyright Commitment, yet the text speaks of third-party intellectual property claims. Google, OpenAI and Anthropic also write the grant around intellectual property generally. The carve-outs then pull trademark and, at Anthropic, patent claims back out.

My position is that the choice of vendor matters less than 2 operating facts. Those are whether a person changes the output before it ships, and which seller's paper governs the model. Coding tools raise extra questions of their own, which are covered in the comparison of AI coding assistant indemnity by plan. This piece stays on the general model and platform contracts.

Which AI outputs are covered: 8 scenarios, 4 vendors

The table below is the original work in this post. I read each vendor's clause against 8 situations a buyer actually meets, then graded every cell. A blank promise of "copyright protection" cannot answer any of these rows.

AI copyright indemnity by output scenario, published terms read 9 November 2026
ScenarioMicrosoftGoogleOpenAIAnthropic
1. Text from the vendor's own model, unedited.Conditional. Needs 4 safeguards on Azure OpenAI.Conditional. You must not disregard or disable its tools.Conditional. You must use relevant citation, filtering and safety features.Clear. No setting is required.
2. Code from the vendor's own model, unedited.Conditional. Needs 4 safeguards on Azure OpenAI, none on GitHub.Conditional. Same tools test.Conditional. Same features test.Clear. No setting is required.
3. Image or video output.Conditional. Needs the 2 universal safeguards.Conditional. Generally available Imagen and Veo are listed.Conditional. No image carve-out appears in the extracts read.Not applicable. No image generator appears in the terms read.
4. Output a person edited.Conditional. Fails if you knew, or should have known, the edit was likely to infringe.Not covered. The grant is limited to unmodified output.Not covered. Modified output is excluded.Not covered to the extent the claim arises from your edit.
5. Output combined with other content.Unconfirmed. The commitment is silent, and the licensing agreement could not be retrieved.Not covered to the extent the claim arises from the combination.Not covered. Combination with non-OpenAI products is excluded.Not covered to the extent the claim arises from the combination.
6. Output from a fine-tuned model.Conditional. You need sufficient rights to the data used to customise.Conditional. A Fine-Tuned Google Model is named, if you hold rights to the data.Unconfirmed. Not addressed in the primary extracts read.Unconfirmed. No carve-out is stated, but your data is excluded.
7. Another lab's model on the platform.Split. Claude in a covered Copilot is in. Claude sold in Foundry is on Anthropic's terms.Not covered on my reading. The grant names Google models.Not covered. Third Party Offerings are excluded, per a secondary summary only.Not applicable. Anthropic sells its own models.
8. Beta, preview or free use.Not covered for free Previews.Not covered. Pre-GA and free services are excluded.Not covered. Beta Services are excluded.Not covered. Beta use has no indemnity, and only paid use counts.

Grades are my reading of published terms, not legal advice. Clear means no product setting or safeguard behaviour is a condition, while general exclusions such as knowing infringement still apply in all 32 cells. Negotiated agreements can differ from every column.

Clear cover is rare: 8 output scenarios, graded by vendor. Each block is 1 scenario. Source: Zan Digital reading of published terms, 9 November 2026. Microsoft 5 1 1 1 Google 4 4 OpenAI 3 1 4 Anthropic 2 1 2 3 Clear. Conditional. Split by route. Unconfirmed. Not applicable. Not covered. Zan Digital, zero weighting: every scenario counts once, whatever its frequency in your business.
Look at the red blocks, not the blue one. Microsoft has the fewest cells marked not covered, and it also asks the most of you before cover applies.

The tally of 32 cells breaks into 2 clear, 12 conditional from Google and OpenAI's 7 plus Microsoft's 5, and 12 not covered across the group. A further 3 are unconfirmed, 1 is split by route and 2 do not apply. Treat the arithmetic as a map of where to read, not a ranking.

Edited output is where AI copyright indemnification narrows fastest

A human edit reduces cover at 3 of the 4 vendors, and the 3 do it in different ways. That matters because almost nothing a business publishes leaves a model untouched. Marketing copy is rewritten, contracts are redlined and code is merged.

Google has the strictest construction. Its grant applies to an unmodified Generated Output, so an edit takes the text outside the promise before any exclusion is reached. OpenAI reaches a similar result through an exclusion for output that was modified or combined with products it did not provide.

Anthropic's wording is narrower, and I think it is the fairest. Its obligations fall away only to the extent the Customer Claim arises from your modification. A passage you left alone should stay protected even when the sentence after it was rewritten.

Microsoft's commitment does not exclude edits at all. It asks instead that you do not modify, use or distribute the output in a manner you know, or should know, is likely to infringe. That is a state-of-mind test, and it is the most buyer-friendly clause in this comparison.

None of the documents I read defines how small an edit counts. Whether fixing a typo is a modification is exactly the kind of question that gets decided after a claim. Keep the generated version beside the published one, because the record is what lets you argue the infringing passage came from the model.

Output ownership is a separate question from indemnity. Whether edited text can be protected at all is examined in the analysis of copyright in AI-generated content, and an indemnity does nothing to answer it.

Required safeguards differ more than the promises do

The safeguards are where the 4 contracts stop resembling each other. One vendor publishes a checklist. 2 write a behavioural test. 1 asks for no setting at all.

Microsoft: a checklist you must be able to prove

Microsoft's page sets out 2 universal mitigations: a metaprompt directing the model not to infringe, and a testing and evaluation report. The report must be retained and handed over if a claim arises. Azure OpenAI text and code then add protected material and jailbreak filters, for 4 items in total. Images and transcription stop at the universal 2.

If you tender a claim, you will be required to demonstrate compliance with all relevant requirements. Output retroactively flagged as protected text under the asynchronous filter is simply not covered. GitHub went the other way: since 3 April 2026 its offerings carry no additional mitigation, and the duplicate detection filter is optional.

What Microsoft asks before its commitment applies. Named safeguards by route, counted from Microsoft Learn, page updated 13 July 2026. Azure OpenAI text4 Azure OpenAI code4 Images, transcription2 GitHub offerings0 Source: Microsoft Learn. Copilot Studio has its own rule for models hosted elsewhere.
The same commitment demands 4 safeguards on one route and none on another, so the evidence you must keep depends on how you bought the model.

Google: do not disregard the tools

Google names no checklist. Cover is lost if you disregard, disable, modify or circumvent source citations, filters, instructions or other tools Google makes available. Cover is also lost if you keep using an output after a rightsholder's notice of a claim. That exclusion turns a claim letter into an operational deadline, and it needs an owner inside your business.

OpenAI: the widest behavioural test

OpenAI excludes cases where the customer or its end users disabled, ignored, or did not use any relevant citation, filtering or safety features. The words "did not use" are the ones I would negotiate. Disabling a feature is an act you can audit. Failing to use one is a judgement a vendor can make in hindsight. Note also that the test reaches your end users, so a customer-facing product inherits their behaviour.

Anthropic: no setting, but a definition doing quiet work

Anthropic's terms list no required filter or configuration. The covered claim is still limited to use in accordance with the terms, so a usage policy breach falls out by definition. Its exclusions also cover use you know, or reasonably should know, infringes the rights of others. No configuration clause is not the same as no conditions, and buyers should not read it that way.

Scope, training data and caps in the AI copyright indemnity clauses

On the money question, 3 vendors give a written answer and 1 sends you to another document. The announcements rarely mention that. Liability caps deserve a separate read, and the wider set of remedy limits is compared in the review of AI agent liability caps and remedies.

Scope and financial terms of each vendor's indemnity, read 9 November 2026
TermMicrosoftGoogleOpenAIAnthropic
Claim type.Third-party IP claims on output.IP rights, excluding trademark use in trade.IP rights in Output.IP rights, excluding patent practice and trademark use in trade.
Training data claims.Not addressed in the commitment.Yes, in a separate indemnity.Not found in the extracts read.Yes, named in Section K.1.
Liability cap on the indemnity.No figure in the commitment. See your licensing agreement.None. Section 12.3 lifts the cap for Section 13.None, per the Services Agreement extract.None. Section L.3.b lifts the cap.
Who is protected.Customers on paid Covered Products.Customer and Affiliates using the account.API and Enterprise customers.Customer, personnel, successors and assigns.

The OpenAI column relies on search extracts of its live policy pages, which refused direct retrieval. Microsoft's cap clause sits in a licensing agreement that could not be retrieved.

Google's position reads cleanly. Section 12.2 limits liability to fees paid in the 12 months before the event, and Section 12.3 removes Section 13 indemnification from that limit. Anthropic's Section L.3.b does the same in one sentence. OpenAI's Services Agreement states the service-specific indemnity is not subject to any liability cap, and that OpenAI may not materially reduce it without written agreement.

Microsoft's commitment is silent on any amount. Its 2023 announcement said it would pay the amount of any adverse judgments or settlements, but a blog post is not the contract. Ask your Microsoft account team where the defence obligation sits against your cap, and get the answer in writing.

Training data cover sounds valuable and, in my view, is the least useful row. Claims about how a corpus was built land on the labs, as the review of the copyright cases on training data shows. Claims about what a generated image or paragraph reproduces are the ones that can reach you, and the dispute in the Disney and Midjourney output case shows what that argument looks like.

Buying another lab's model changes whose promise you hold

The contract follows the seller, and the seller is not always the company whose name is on the invoice. A model available in 3 clouds can sit under 3 different allocations of risk.

Microsoft shows the split most clearly. Its documentation says the commitment applies to Claude models used within products covered by the CCC, including Microsoft Copilot and Copilot Studio. In the Foundry catalogue, the same family is a partner model. Microsoft's page says Anthropic is the seller and that its terms of use govern.

Copilot Studio has its own trap. Output from a model hosted outside Copilot Studio is not covered unless it runs in Azure OpenAI and meets the listed mitigations. Bringing your own model is a coverage decision as well as an engineering one.

Google's grant covers Google Pre-Trained Models, and its terms disclaim indemnity for Separate Offerings under third-party terms. I read Claude on Google's platform as falling outside Google's promise, although Google's documentation does not say so directly. Anthropic's Bedrock position is explicit: Anthropic's services on Bedrock are sold by Anthropic, on its terms.

A team moving to a cheaper route should check this first. The saving on inference can be real while the indemnity counterparty quietly changes. That is the same trap as personal AI accounts inside company networks, only with an enterprise purchase order attached.

Where this comparison is weakest

My evidence is uneven, and 2 gaps could change cells in the table. You should weigh them before you rely on it.

Documents I could not read in full

OpenAI's policy pages refused direct automated retrieval. Its column rests on search extracts of the live Service Terms and Services Agreement, and the Third Party Offering and fine-tuning points rest on secondary summaries. Microsoft's cap clause sits in a licensing agreement that I could only reach as an empty shell. Both are marked in the tables, and either could move on a full read.

Untested promises and the case against caring

I found no reported case in which a customer invoked one of these output indemnities and a court construed it. Every reading here is therefore a reading of drafting, not of outcomes. The tally also weights all 8 scenarios equally, which is wrong for any particular business.

There is a fair argument that Microsoft's checklist is better for you than Anthropic's silence. A protected material filter can stop an infringing passage from shipping, and an indemnity only helps afterwards. There is also a real argument that this risk is small for most buyers, since the large cases so far have targeted model builders. I would still read the clause, because it costs 1 afternoon and the downside is a claim with no defence.

Frequently asked questions

Does ChatGPT Enterprise include copyright indemnity?

Yes, on OpenAI's published terms. The Service Terms extend an output indemnity to API and Enterprise customers, covering third-party claims that your use or distribution of Output infringes an intellectual property right. It does not apply if you knew the Output was likely to infringe, or if you disabled, ignored or did not use relevant citation, filtering or safety features. Beta Services are excluded.

Is AI copyright indemnification capped?

At Google, OpenAI and Anthropic, no. Each states that indemnity obligations sit outside the general liability cap of 12 months of fees. Microsoft's Customer Copyright Commitment states no figure, but it operates through your volume licensing agreement, and I could not retrieve that clause. Check your own agreement. Negotiated contracts can differ from all 4 published positions, and preview services carry separate limits.

Does Microsoft's copyright commitment cover Claude models?

It depends on the route. Microsoft says the commitment applies to Claude models used within products covered by it, including Microsoft Copilot and Copilot Studio. Claude bought through the Microsoft Foundry catalogue is different. Microsoft lists it as a partner model, names Anthropic as the seller, and says Anthropic's terms govern. Your claim would then run to Anthropic, not Microsoft.

Does editing AI output void the indemnity?

It can. Google's output indemnity covers only unmodified Generated Output, and OpenAI excludes Output that was modified or combined with products it did not provide. Anthropic excludes claims to the extent they arise from your modifications. Microsoft has no such exclusion in its commitment, but cover fails if you modify output in a way you know, or should know, is likely to infringe.

Does Google indemnify Claude on Vertex AI?

On my reading of the published terms, no. Google's output indemnity applies to Generated Output from services using Google Pre-Trained Models, and its terms disclaim indemnity for Separate Offerings governed by third-party terms. Google's Claude documentation requires you to accept terms before use but does not name an indemnity. Treat this as unresolved until Google or Anthropic confirms it in writing for your account.

Do free AI plans include copyright indemnity?

No, not under the terms reviewed here. Google's general indemnity excludes services provided free of charge. Microsoft's Covered Product definition requires a fee or paid subscription and excludes free Previews. Anthropic's covered claim is limited to paid use of its services. OpenAI's output indemnity is written for API and Enterprise customers. If your staff use personal accounts, assume no vendor defence applies.

Where to start

Pick the 1 AI workflow that publishes the most outside your company. Write down the seller of record, the exact model, and whether a person edits the output before release. Then find that row in the first table.

If the row says conditional, name the person who holds the evidence, and add a step that pulls an output once a claim letter arrives. If the answer is unconfirmed, send your vendor the scenario in 1 sentence and ask for a written reply. That question belongs in the same round as the AI agent procurement checklist, before renewal gives you less room.

Read next

For the rest of the contract, including retention, training rights and uptime, see the four terms that decide your risk with an AI provider.

References

  1. Microsoft Learn, Customer Copyright Commitment Required Mitigations, updated 13 July 2026. Used for every safeguard, date and evidence requirement.
  2. Microsoft, Product Terms, Universal License Terms for Online Services, read 9 November 2026. Used for conditions and the Covered Product definition.
  3. Microsoft Learn, Claude models from Anthropic in Microsoft Online Services, updated 4 November 2026. Used for Claude inside Copilot.
  4. Google Cloud, Service Specific Terms and Generative AI Indemnified Services, read 9 November 2026. Used for grants, exclusions and the service list.
  5. Google Cloud, Master Agreement, read 9 November 2026. Used for Sections 12 and 13.
  6. OpenAI, Service Terms, seen in search extracts on 9 November 2026. Direct retrieval was refused.
  7. OpenAI, Services Agreement, seen in search extracts on 9 November 2026. Used for the cap carve-out.
  8. Anthropic, Commercial Terms of Service, effective 17 June 2025, and Service Specific Terms, effective 31 August 2026. Used for Sections K and L and beta services.

The weakest part of this source base is the OpenAI column, which rests on search extracts rather than a full read, followed by Microsoft's unretrieved cap clause. All terms change without notice. Last verified 9 November 2026. This is commercial analysis, not legal advice.

ST
Ritu Raj
Writes for Zan Digital on AI contracts, vendor terms and what published clauses actually commit a seller to do.

Related reading