Most AI contracts have an IP indemnification clause. Here is what that clause actually says at OpenAI, Microsoft, Google, and Anthropic, and the questions to bring to your attorney about the work you deliver.
Disclaimer: I am not a lawyer, and this is not legal advice. This post points you at the published terms and gives you better questions to ask. Nothing here tells you what your legal exposure is. For anything material, have an actual attorney read your actual contract.
You Delivered That Work. Now Whose Problem Is It?
You wrote a client's website copy with AI help last month. You built a course outline. You generated fifteen product images for a launch.
That work is out in the world now, with your name on the invoice.
And if someone claims a piece of it infringes their copyright, the first question anyone asks is whether your AI provider agreed to defend you.
Here is the direct answer: the four largest providers all publish an IP indemnification clause covering output, and every one of them attaches conditions to it. The conditions matter more than the headline. Most of them require a paid tier. Most of them exclude output you modified. And all four exclude trademark claims when you use output in trade or commerce, which is exactly what selling work is.
So the honest answer to "am I covered" is that nobody on the internet can tell you. The terms are public, they are short, and you can read the relevant section in about ten minutes.
That is the point of this post. Not panic. A repeatable review you run once, write down, and repeat when you change tools.
Candidly, I was not planning to write about contract language this week. Two newsletters pushed me into it, and I will credit them properly below.
Key Takeaways
- All four major providers publish an output indemnity, and none of them apply it to every tier or every use.
- Google Cloud's definition ties indemnity to paid use and excludes free tier usage in the definition itself.
- Anthropic's Commercial Terms tie the customer indemnity to "Customer's paid use of the Services."
- OpenAI, Google, and Anthropic all carve out trademark claims arising from your use of output in trade or commerce.
- Modification and combination exclusions appear in all four sets of terms, which is worth an attorney conversation if you edit AI output before delivering it.
The Real Problem Is Not the Lawsuits. It Is That Nobody Reads the Clause.
Two big legal stories landed in the last several weeks, and both of them made the same point from opposite ends.
On September 4, 2026, The Seattle Times Company and Newsday LLC filed a federal complaint against nine OpenAI entities and Microsoft, alleging their journalism was copied without permission to train and operate ChatGPT, Copilot, and related services. TechCrunch and Engadget both covered it. It joins a line of publisher suits going back to the New York Times complaint in 2023.
On July 20, 2026, a federal judge granted final approval to the 1.5 billion dollar class settlement in Bartz v. Anthropic. The Authors Guild reported the final approval. Roughly 350 authors opted out, and attorneys' fees were cut to about 101.6 million dollars.
Put those two together and you get the shape of the moment. One company's training data exposure is now a settled number. Several others are still open questions in active litigation.
Here is the thing. None of that is your fight.
You are not a defendant in any of those cases. You are a coach, a consultant, a small agency owner. Your risk is not the training data question. Your risk is much smaller and much more specific: a client sees something in the work you delivered, someone else claims it as theirs, and your contract with your provider decides who picks up the phone first.
That risk existed before September. It exists whether or not the publishers win.
And most founders I talk to have never opened the clause.
That is not a character flaw. Provider terms are spread across three or four linked documents, they get updated without announcement, and they use words like "Configurable GAI Services." Nobody reads that on a Tuesday.
So let us make it small enough to actually do.
What the Published Terms Actually Say
I read the current terms pages for four providers. Here is what is in them, quoted, with links in the Sources section so you can check my work.
OpenAI. The Service Terms, updated June 12, 2026, say in Section 1: "OpenAI's indemnification obligations to API customers under the Agreement include any third party claim that Customer's use or distribution of Output infringes a third party's intellectual property right." Six exclusions follow in the same paragraph, including where "Output was modified, transformed, or used in combination with products or services not provided by or on behalf of OpenAI," and where "the claim alleges violation of trademark or related rights based on Customer's or its End Users' use of Output in trade or commerce."
Section 3(b) grants a parallel output indemnity to "Enterprise customers." Section 3(a) defines that term: "ChatGPT Enterprise, ChatGPT Edu, and ChatGPT for Healthcare (collectively, 'Enterprise')." Section 2 says Beta Services "are excluded from any indemnification obligations OpenAI may have to you."
I will not tell you what that means for you. I am telling you the definition sits in Section 3(a), and you should go read which bucket your plan falls into.
Microsoft. The Customer Copyright Commitment is a provision in the Microsoft Product Terms. Microsoft's published required mitigations page states that for Azure OpenAI, "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." For open text generation, the page requires that "the protected material text model must be configured on in filter mode." It also requires a retained testing and evaluation report, which "must be retained by the customer and provided to Microsoft in the event of a claim."
Read that last part twice. The coverage assumes you kept records.
Google Cloud. The Service Specific Terms define a "Generative AI Indemnified Service" as a listed service or feature "where the use of such Service or feature is paid for by Customer and not subject to credits or free tier usage." The generated output indemnity applies to "unmodified Generated Output." Exclusions include use "after receiving notice of an infringement claim" and allegations "based on a trademark-related right as a result of Customer's use of such Generated Output in trade or commerce."
Anthropic. The Commercial Terms of Service, effective June 17, 2025, define a Customer Claim in Section K.1 as a third-party claim alleging that "Customer's paid use of the Services (which includes data Anthropic has used to train a model that is part of the Services) in accordance with these Terms or Outputs generated through such authorized use violates any third-party intellectual property right." Section K.3 excludes claims arising from "modifications made by Customer to the Services or Outputs," from combination with non-Anthropic technology, and from "an alleged violation of trademark based on use of an Output in trade or commerce." The same page states plainly that these terms are not for consumer use and that Claude.ai is governed by separate Consumer Terms.
Adobe. Adobe markets Firefly as commercially safe and publishes legal FAQs for enterprise customers describing IP indemnification on qualifying plans. I did not verify current enterprise contract language against a live terms page, so treat that one as unverified here and ask Adobe directly which of your entitlements carry it.
One pattern held across every set of terms I read: trademark claims from commercial use are carved out, and modified or combined output is treated differently from raw output.
That is a question worth bringing to a lawyer, not a conclusion I am qualified to draw.
Build the Review Once. Run It Forever.
Credit where it is due. The Neatprompts newsletter raised the practical version of this question this week: does your provider's IP indemnification cover output you ship to customers, or only output you use internally? Their read is that most founders assume the first while plenty of contracts say the second.
The AI Breakfast newsletter ran a separate item on a joint NSA, FBI, and CISA advisory. As reported by AI Breakfast and by outlets including DigiTimes, the advisory recommends that US labs quietly degrade responses to suspected model distillation traffic rather than block it, and avoid telling affected accounts. I have not read the advisory itself, so I am passing that along as reported by named publications, not as verified fact. AI Breakfast framed it as a buyer's warning about the API you are billing against, and that framing is fair regardless.
Two more things belong on the same page. Nvidia confirmed on September 3 that it is buying Hugging Face for 12.9 billion dollars, subject to approvals, and Hugging Face is where most open models get distributed. Neatprompts also flagged that Anthropic is backing a Massachusetts bill requiring independent evaluators while OpenAI has pointed lawmakers toward the Illinois annual-audit model, which the Boston Globe and Insurance Journal both reported. Safety posture is turning into a purchasing question.
All of it points at the same underlying thing. You have a supplier relationship you never examined.
So examine it like a supplier relationship. Not like a crisis.
I think in systems, and a system has three properties. It is written down, it runs on a schedule, and it produces an artifact you can hand to someone else.
Here is the one I use.
The artifact is a one-page vendor sheet. One row per AI tool you actually use in client work. Columns: tool name, exact plan tier, link to the governing terms document, date you last read it, and the section number where the output indemnity lives. That is it. One page.
The schedule is quarterly, plus any time you change plans. Providers update terms without a press release. OpenAI's Service Terms carry an "Updated" date at the top. Google Cloud publishes an archive of previous versions with dates. Anthropic links the previous version. Those dates are your change log.
The trigger is new client work. Before you build a delivery process around a tool, you check the sheet. If the tool is not on the sheet, it does not go into client deliverables until it is.
That is the whole system. It fits on a single page and takes twenty minutes to populate the first time.
What the sheet gives you is not legal protection. It gives you the ability to answer a question in thirty seconds instead of thirty days, and it gives your attorney something to work from that is not a shrug.
I would rather hand a lawyer one page of specifics than pay for an hour of them figuring out which plan I am on.
Seven Steps to Run This Week
1. List every AI tool that touches client deliverables. Not every tool you have ever tried. The ones whose output reaches a paying customer. Most people find between three and six and are surprised by two of them.
2. Write down your exact plan tier for each one. Not "I have ChatGPT." Which product, which plan, billed to which entity. This single line resolves more ambiguity than anything else on the sheet, because the published indemnities are tier-specific.
3. Open the governing terms and find the indemnification section. OpenAI puts it in the Service Terms. Anthropic puts it in Section K of the Commercial Terms. Google Cloud puts it in the Service Specific Terms. Microsoft puts it in the Product Terms with a separate required mitigations page. Bookmark the exact URL.
4. Have AI pull the clause text for you, not summarize it. Use this prompt. It is built on the Perfect Prompt Framework 2.0, and it is deliberately written to stop the model from playing lawyer.
[The Job]
Extract and quote the intellectual property indemnification language from the terms document I paste below. Do not interpret it, do not summarize it, and do not give me an opinion about my coverage.[The Background]
I run a small business and I deliver AI-assisted work to paying clients. I am preparing a list of questions for a licensed attorney. I need exact language, not analysis. If any part of the document is ambiguous, incomplete, or refers to another document I have not provided, say so explicitly instead of filling the gap.[The Deliverable]
Return three sections. First, the verbatim text of every indemnification provision, with its section number. Second, the verbatim text of every exclusion, condition, or carve-out attached to it, with section numbers. Third, a plain list of every place where you are uncertain what the language covers, or where the document points to terms I did not give you. Quote exactly. Do not paraphrase. Do not advise.[The Questions]
Ask me any questions you have.
5. Have an attorney review this. This is the step people skip, and it is the only one that produces an actual answer. Bring the vendor sheet, the quoted clauses, and the uncertainty list from step four. A licensed attorney in your state can tell you what those exclusions mean for your specific delivery model. I cannot, and neither can a chatbot.
6. Check your own client contracts while you are in there. The Koley Jessen commercial technology team published practical AI contracting guidance in February 2026 covering both sides of this. Their point that indemnity language written for "software" may not extend to "model outputs and predictions" cuts both ways, including in the agreements you hand your own clients.
7. Put the sheet on a calendar and close the loop. Quarterly review, thirty minutes. Note the "last updated" date on each terms page. If a date moved, reread that one section. If nothing moved, you are done in ten minutes.
Frequently Asked Questions
Does this mean I should stop using AI for client work?
No. Nothing in this post supports that conclusion. Every major provider publishes an output indemnity precisely because they expect commercial use. The point is to know which tier you bought, what conditions attach, and to get an attorney's read before you scale a delivery process on assumptions.
I am on a free or personal plan. Am I covered?
Read your own terms and ask a lawyer. What I can tell you is what the documents say. Google Cloud's definition of an indemnified service excludes free tier usage in the definition itself, and Anthropic's Commercial Terms tie the customer indemnity to paid use. Those are quotes, not conclusions about you.
I always edit the AI output before I deliver it. Does that help or hurt?
That is exactly the question to bring to your attorney. All four providers address modified or combined output somewhere in their exclusions, and the wording differs between them. Editing is good practice for quality and for copyright registration. What it does to an indemnity is a legal question I am not qualified to answer.
Can I copyright work I made with AI?
The Copyright Office's position is that human authorship is required, and applicants must disclaim AI-generated portions when registering. The Supreme Court declined to take up the question in March 2026, leaving that framework in place. Detailed prompts alone are generally not treated as authorship. Registration questions go to an attorney.
How often do these terms actually change?
Often enough to matter. OpenAI's Service Terms show an update date of June 12, 2026. Google Cloud's indemnified services list was last modified July 20, 2026, and maintains a public archive of sixteen prior versions. Microsoft's mitigations page shows requirement changes taking effect as recently as April 2026. Quarterly is a reasonable cadence.
Ten Minutes and a Bookmark
You delivered that work with your name on it. That was the right call, and I would make it again.
What I would not do is assume that the contract behind the tool says what I hope it says.
Here is the thing about this whole category of anxiety. It feels enormous because it involves lawyers and lawsuits and words like indemnification, and enormous things are easy to postpone. But the actual task in front of you this week is not enormous. It is one page, four columns, and a bookmark to a terms page you have never opened.
You will not come out of it with certainty. You will come out of it with specifics, and specifics are what let a real attorney give you a real answer in one billable hour instead of five.
That is the trade. Ten minutes of reading now buys you a much cheaper conversation later.
Do not take my word for any of it. Every clause I quoted is linked below. Go read the one that applies to you, disagree with me if you want, and bring your questions to someone licensed to answer them.
If you want the vendor sheet template and the walkthrough, that is the kind of thing we build together inside AI Insiders. The door is open when you are ready.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he teaches non-technical entrepreneurs to use AI to amplify the skills they already have. He runs a Facebook community of roughly 500,000 members and the AI Insiders membership.
Sources
Primary terms and official documents:
- OpenAI Service Terms, updated June 12, 2026: https://openai.com/policies/service-terms/
- Microsoft Customer Copyright Commitment Required Mitigations: https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/customer-copyright-commitment
- Google Cloud Generative AI Indemnified Services: https://cloud.google.com/terms/generative-ai-indemnified-services
- Google Cloud Service Specific Terms: https://cloud.google.com/terms/service-terms
- Anthropic Commercial Terms of Service, effective June 17, 2025: https://www.anthropic.com/legal/commercial-terms
- U.S. Copyright Office, Copyright and Artificial Intelligence: https://www.copyright.gov/ai/
- Adobe Firefly legal FAQs for enterprise customers (referenced, not verified against current contract language): https://www.adobe.com/cc-shared/assets/pdf/enterprise/firefly-legal-faqs-enterprise-customers-2024-06-11.pdf
Reporting and legal commentary:
- TechCrunch, Seattle Times and Newsday sue OpenAI and Microsoft: https://techcrunch.com/2026/09/05/seattle-times-and-newsday-are-the-latest-publications-to-sue-openai-and-microsoft/
- Engadget, Two more news organizations sue OpenAI and Microsoft: https://www.engadget.com/2251707/seattle-times-newsday-sue-openai-microsoft-for-copyright-infringement/
- The Seattle Times, The Seattle Times sues OpenAI, Microsoft: https://www.seattletimes.com/business/local-business/seattle-times-sues-openai-microsoft-over-copyright-infringement/
- The Authors Guild, Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement: https://authorsguild.org/news/court-grants-final-approval-anthropic-copyright-settlement/
- TechCrunch, Anthropic's landmark $1.5B copyright settlement is approved: https://techcrunch.com/2026/07/20/anthropics-landmark-1-5b-copyright-settlement-is-approved/
- Koley Jessen, AI Contracting: Practical Legal Guidance for Growing Businesses, February 24, 2026: https://www.koleyjessen.com/insights/publications/ai-contracting-practical-legal-guidance-for-growing-businesses
- Morgan Lewis, US Supreme Court Declines to Consider Whether AI Alone Can Create Copyrighted Works: https://www.morganlewis.com/pubs/2026/03/us-supreme-court-declines-to-consider-whether-ai-alone-can-create-copyrighted-works
- Proskauer, OpenAI's Copyright Shield Broadens User IP Indemnities: https://www.proskauer.com/blog/openais-copyright-shield-broadens-user-ip-indemnities-for-ai-created-content
- NVIDIA, NVIDIA to Acquire Hugging Face: https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/
- TechCrunch, Nvidia confirms it will buy Hugging Face for $12.9 billion: https://techcrunch.com/2026/09/03/nvidia-confirms-it-will-buy-hugging-face-for-12-9-billion/
- The Boston Globe, Anthropic backs new AI regulations in Massachusetts: https://www.bostonglobe.com/2026/06/26/business/anthropic-massachusetts-ai-regulations-bill/
- Insurance Journal, OpenAI Warns of Confusion as Massachusetts Readies Crackdown: https://www.insurancejournal.com/news/east/2026/08/20/882315.htm
- DigiTimes, US security agencies tell AI firms to degrade answers to suspected distillation traffic: https://www.digitimes.com/news/a20260909VL206/ai-training-claude-deepseek-geopolitics.html
Newsletters that raised these questions: