Microsoft will defend you over AI copyright, if you configured it correctly
The Customer Copyright Commitment is real, and it is conditional. Coverage for Azure OpenAI depends on the customer having implemented a specific list of mitigations, and Microsoft states that a customer tendering a claim "will be required to demonstrate compliance with all relevant requirements". Most buyers believe they are indemnified because they bought the product. They are indemnified because of how they configured it, and a large number have configured nothing.
What the commitment actually is
The Customer Copyright Commitment, usually shortened to CCC, is a provision in the Microsoft Product Terms describing Microsoft's obligation to defend customers against certain third-party intellectual property claims relating to output content. That is a meaningful promise, and it is why the CCC is quoted so often in procurement conversations.
The part that rarely survives into those conversations is the sentence that follows it. For Azure OpenAI and any configurable generative AI service, the customer also must have implemented all mitigations required by the Azure OpenAI documentation in the offering that delivered the output content that is the subject of the claim.
So the commitment is not a warranty that attaches on purchase. It is closer to an insurance policy with conditions precedent: the cover exists, and it is contingent on things you were supposed to have done before the claim arrived. Discovering the conditions at claim time is the worst possible moment, because compliance is assessed retrospectively against what you actually had in place.
The two universal requirements
Two mitigations apply to every offering delivering output from Azure OpenAI and configurable generative AI services, with GitHub Offerings and Copilot Studio handled separately. Both have been in force since 1 December 2023.
A metaprompt. The customer offering must include a metaprompt directing the model to prevent copyright infringement in its output. Microsoft points to its own "Protected Material" safety system message component as the worked example. This is a system-level instruction in your application, not a setting in a portal, which means an application built before anyone read this page almost certainly does not have one.
A testing and evaluation report. The offering must have been subjected to evaluations using tests designed to detect the output of third-party content, through guided red teaming, systematic measurement or an equivalent approach. Significant ongoing reproduction of third-party content found in evaluation must be addressed. Then the operative clause: the report of results and mitigations must be retained by the customer and provided to Microsoft in the event of a claim.
That second requirement is where most organisations quietly fall out of cover. It is not enough to have been careful. There has to be a document, it has to be kept, and it has to be produceable years later when a claim lands. Microsoft adds one clarification in the customer's favour: you are under no obligation to test Microsoft's own services to maintain coverage. The evaluation is of your application.
Code and text are not treated the same
Beyond the universal pair, Azure OpenAI adds requirements by use case, and they are cumulative: an application spanning several use cases must satisfy the mitigations for all of them.
For code generation, the protected material code model must be on in either annotate or filter mode, and Prompt Shield for jailbreak attacks must be on in filter mode. Annotate mode is permitted, with a string attached: if you choose it, you must comply with any cited licence provided for the output content that becomes the subject of the claim. In practice that converts a compliance question into an engineering one, because someone has to actually act on those annotations.
For text generation, including journalistic content and writing assistance, the protected material text model must be on in filter mode, and Prompt Shield in filter mode. There is no annotate option that preserves cover here. Text is the stricter of the two, which surprises teams who assume code is the sensitive case because code carries licences.
Image generation models, transcription models and all other use cases carry no additional requirements.
The exclusion that is easiest to miss
One clause deserves separating out, because it removes cover rather than conditioning it, and because the feature it concerns is chosen for performance reasons by people who are not thinking about indemnity at all.
If you use the asynchronous filter, output content retroactively flagged as protected material is not covered by the commitment. For text that is absolute. For code it holds unless the customer complies with the cited licence. This has applied since 21 May 2024.
The asynchronous filter exists because synchronous filtering adds latency, and streaming responses feel better without it. So the trade is explicit once you see it: a faster-feeling product, and no indemnity on exactly the outputs the filter would have caught. That is a decision worth making deliberately at the architecture stage rather than discovering in a legal review.
GitHub Copilot and Copilot Studio follow different rules
The universal mitigations do not apply to GitHub Offerings or Microsoft Copilot Studio. These are configurable services with their own service-specific requirements, and conflating them with Azure OpenAI is a common source of wrong advice.
For GitHub Offerings, Microsoft states that as of 3 April 2026 there are no additional required mitigations, and that use of the Duplicate Detection filter is no longer required for coverage. The feature remains available to use, and organisations that enabled it for its own sake can keep it, but it is no longer the price of admission to the CCC.
For Copilot Studio, the single stated requirement concerns bringing your own model. If you connect or incorporate a model hosted outside Copilot Studio, output from that model is not covered unless the model runs in Azure OpenAI and meets the Azure OpenAI mitigations. That has applied since 1 June 2025, and it is a meaningful constraint on architectures that route to a third-party model for cost or capability reasons.
Requirements can change, and you get six months
Microsoft states that for new configurable services, features, models or use cases, new CCC requirements will be posted and take effect at or following launch. Otherwise, customers have six months from the date of publication to implement any new mitigation required to maintain coverage.
The practical consequence is that CCC compliance is not a one-time project. It is a page someone has to re-read on a schedule, because a requirement added quietly starts a clock that nobody is watching. Six months is generous by the standards of contractual notice, and it is short by the standards of an enterprise change queue.
What to actually check
Five questions answer whether an organisation is genuinely covered, and all five can be answered from the application and its records rather than from a contract.
Does the application include a copyright-prevention metaprompt? Is there a written evaluation report showing tests designed to detect third-party content output, and is it retained somewhere retrievable? Are the protected-material filters on, and for text specifically in filter mode? Is the asynchronous filter in use, and if so, has anyone accepted that the outputs it flags are outside the commitment? And if a model is brought from outside Copilot Studio, does it run in Azure OpenAI under the same mitigations?
An organisation that cannot answer those from evidence is not necessarily going to face a claim. It is holding a commitment it has not yet earned the right to invoke, which is a different and more recoverable problem than having no cover at all.
What this page does not cover
This page describes Microsoft's published requirements only. Other vendors publish their own indemnity positions with their own conditions and exclusions, and those positions differ materially from one another.
We attempted to include Google Cloud's generative AI indemnified services terms alongside this and could not retrieve the page in a form we could quote accurately. Rather than describe another company's contractual position from memory, we have left it out and said so. If you need a vendor-by-vendor comparison for creative tools specifically, our creative-AI indemnification matrix covers Adobe, Midjourney, Google Veo and Getty, and our guide to AI output rights and ownership covers the underlying question of who owns generated work.
Nothing here is legal advice. It is a reading of a published requirements page, quoted and linked so you can check it against your own deployment and take the question to counsel with the specifics already assembled.
Frequently asked questions
Does Microsoft indemnify me for AI copyright claims automatically?
No. The Customer Copyright Commitment is a provision in the Microsoft Product Terms, and for Azure OpenAI and other configurable services it applies only where the customer has implemented all required mitigations. Microsoft states that a customer tendering a claim "will be required to demonstrate compliance with all relevant requirements". Buying the service is not what earns the cover; configuring it is.
What are the universal required mitigations?
Two, both effective from 1 December 2023. First, the customer offering must include a metaprompt directing the model to prevent copyright infringement in its output. Second, the offering must have been evaluated using tests designed to detect output of third-party content, through guided red teaming, systematic measurement or an equivalent approach, and the report of results and mitigations must be retained by the customer and provided to Microsoft in the event of a claim.
What extra is required for code generation?
For code generation and transformation scenarios, the protected material code model must be configured on in either annotate or filter mode, and the jailbreak model, meaning Prompt Shield for jailbreak attacks, must be on in filter mode. If annotate mode is chosen, the customer must comply with any cited licence provided for the output content that is the subject of the claim.
What extra is required for text generation?
For journalistic content, writing assistance and other open text generation, the protected material text model must be configured on in filter mode, not annotate, and Prompt Shield must be on in filter mode. The difference matters: for text there is no annotate option that preserves cover.
Is output caught by the asynchronous filter covered?
No. Microsoft states that where the asynchronous filter feature is used, output content retroactively flagged as protected material is not covered by the commitment, for text outright and for code unless the customer complies with its cited licence. That has applied since 21 May 2024 and it is the single most easily missed exclusion.
Do these conditions apply to GitHub Copilot?
Differently. GitHub Offerings and Microsoft Copilot Studio are configurable services to which the universal mitigations do not apply; service-specific mitigations apply instead. As of 3 April 2026 there are no additional required mitigations for GitHub Offerings, and use of the Duplicate Detection filter is no longer required for coverage, though it remains available.
What about a model I bring myself into Copilot Studio?
Output from a model connected to or incorporated into Copilot Studio but hosted outside it is not covered, unless that model runs in Azure OpenAI and meets the Azure OpenAI required mitigations. That has applied since 1 June 2025.
Source
Every requirement, date and exclusion above is taken from a single published Microsoft page, quoted rather than paraphrased where the wording carries the obligation.
- Customer Copyright Commitment Required Mitigations, Microsoft Learn. Source of the universal mitigations and their 1 December 2023 effective date, the per-use-case requirements for code and text generation, the 21 May 2024 asynchronous filter exclusion, the 3 April 2026 position on GitHub Offerings, the 1 June 2025 Copilot Studio bring-your-own-model rule, and the six-month implementation window. Retrieved 30 August 2026.
Microsoft revises this page, and requirements added to it start a six-month clock. The version above was read on 30 August 2026; check the source before relying on any requirement here.