Vendors are quietly rewriting who pays when an AI agent's training data turns out to infringe someone's copyright, and increasingly that's you, not them. Here's the exact clause language to demand before signing, not after a lawsuit names your company.
- Standard SaaS indemnity clauses were written for data breaches, not for AI output accused of reproducing copyrighted training material
- Many AI agent vendor contracts narrow indemnification to cases where the vendor "knowingly" used infringing data, a standard that's nearly impossible for a customer to prove
- Getty Images sued Stability AI in UK and US courts over training data, showing these disputes are already in court, not theoretical
- Microsoft's Copilot Copyright Commitment, announced in 2023, extended indemnification to customers in a way most smaller AI vendors still refuse to match
- Demand uncapped indemnification tied to IP claims specifically, separate from the general liability cap that covers everything else in the contract
You buy an AI agent to handle customer support, draft contracts, or generate marketing copy, and you sign the same vendor agreement you'd sign for any SaaS tool. Nobody flags the indemnification section as different. It should be.
Here's what's actually happening. The vendor trained its model, or fine-tuned someone else's foundation model, on a dataset it may not fully own or have licensed. If a copyright holder later proves their work ended up in that training set and the agent's output reproduces it, somebody pays. Standard vendor contracts increasingly decide that somebody is you, the customer, not the company that built and trained the model.
Indemnification clauses in software contracts were built for a different world. They typically promise the vendor will defend you if a third party sues claiming your use of the vendor's product infringes a patent or copyright the vendor controls. That's a clean, bounded risk: the vendor wrote the code, the vendor owns the IP, the vendor indemnifies.
Generative AI breaks that model because the infringement risk doesn't come from the code. It comes from the training data, which the vendor often licensed from a third party, scraped, or bought in bulk from a data broker with murky provenance. The vendor's lawyers know this, and many vendor-drafted agreements now carve training data out of the indemnity entirely, or bury a carve-out inside a definitions section that redefines "Vendor IP" to exclude the model weights and training corpus.
How a Liability Waterfall Actually Works Across Your AI Vendor Stack Liability waterfall multiple AI vendors contracts rarely specify payment order, leaving startups exposed when agents fail. Air Canada's 2024 chatbot ruling shows the customer-facing company pays first, regardless of which vendor's model caused the error. - liability waterfall in AI vendor contracts - who pays when AI agent stack fails
Read your current AI agent contract's indemnification section right now. If it indemnifies you only against claims that the "Software" or "Platform" infringes IP, and separately defines "Software" to exclude training data, outputs, or model weights, you have no coverage for the exact risk that matters. That's not an oversight. It's a drafting choice, and it shifts the entire cost of a copyright suit onto you.
The real cases this is already producing #
This isn't a hypothetical risk. Getty Images sued Stability AI in both UK and US courts, alleging Stable Diffusion was trained on millions of Getty's copyrighted photographs without a license. The UK case went to trial in 2025, with a High Court ruling in November that largely sided with Stability AI on the core copyright claims while still finding trademark infringement over watermarks appearing in generated images, according to Reuters. The New York Times sued OpenAI and Microsoft in December 2023 over training data drawn from its archives, a case still working through discovery. The New York Times is also pursuing a related case against Perplexity. These aren't fringe plaintiffs. They're organizations with the resources to litigate for years, and every one of them is suing the company that built the model, not that company's customers, for now.
That "for now" is the part founders miss. Precedent doesn't require the plaintiff to sue the reseller or the business customer first. A copyright holder can sue whoever is distributing or commercially benefiting from infringing output, and a company that resold an AI agent or built a product on top of one, embedding the vendor's model in a customer-facing tool, is a defendant with money and a public footprint. If your contract leaves you holding that risk alone, you're a more attractive target than the vendor's legal department, not less.
What Microsoft did, and why most vendors won't match it #
In September 2023, Microsoft announced its Copyright Commitment for Copilot, pledging to assume responsibility for legal risks if a customer is sued over copyright infringement from Copilot's output, provided the customer used the built-in guardrails. Microsoft extended this across its Copilot products, and it's become the reference point founders cite when negotiating with smaller vendors. It's worth citing, because it proves the clause you want is commercially viable, not a legal fantasy you're asking a startup to invent from scratch.
Most AI agent vendors, especially smaller ones building on top of OpenAI, Anthropic, or open-weight models, will tell you they can't offer what Microsoft offers because they don't control the base model's training data. That's a legitimate point, and it's also exactly why you need to negotiate the clause instead of accepting the default. A vendor that genuinely can't indemnify you against base-model training risk can often still indemnify you against their own fine-tuning data, their own RAG corpus, and their own output filtering decisions. Push for what they actually control before accepting a blanket no.
The specific language to put on the table #
Four things separate a real indemnity clause from a cosmetic one, and you should ask for all four in redline, not just one.
First, strike any "knowing infringement" standard. Some vendor drafts indemnify you only if the vendor "knowingly and intentionally" used infringing material in training. That standard is almost never met in practice, because proving what a vendor knew at the time it assembled a dataset is close to impossible for an outside customer, and the vendor controls all the evidence. Replace it with a strict standard: the vendor indemnifies regardless of the vendor's knowledge or intent, so long as the underlying IP claim relates to the vendor's model, training data, or fine-tuning corpus.
How To Negotiate An AI Agent Benchmark Pricing Clause An AI agent benchmark pricing clause ties price hikes to independently verified gains, not vendor marketing. Buyers are pushing it into renewals now, naming benchmarks like SWE-bench and labs like Epoch AI to check claims before paying more. - how to negotiate AI pricing contracts with vendors - AI agent benchmark pricing clause for SaaS tools
Second, separate the IP indemnity cap from the general liability cap. Most SaaS contracts cap total liability at the fees paid in the prior 12 months, often a modest number for an early-stage AI agent deployment. If your indemnity obligation is capped at the same figure, a losing copyright judgment can exceed that ceiling by orders of magnitude, and you absorb the rest. Ask for IP indemnification to be carved out of the general cap entirely, or capped separately at a materially higher figure.
Third, require the vendor to control and fund the defense, not just reimburse you after the fact. A clause that says the vendor will "reimburse reasonable legal costs" still leaves you fronting the money and running the litigation. The clause you want gives the vendor the right and obligation to defend the claim, with its own counsel, at its own cost, from the moment a claim is filed.
Fourth, get a warranty, not just an indemnity, that the vendor has the rights needed to train and license the model as delivered. An indemnity only pays out after you're sued. A warranty gives you a separate breach-of-contract claim against the vendor the moment it turns out the representation was false, which matters if the vendor goes bankrupt, gets acquired, or simply refuses to honor the indemnity in good faith.
What to do if the vendor won't move #
Smaller vendors, especially ones reselling access to a foundation model through an API, will often say they genuinely cannot assume unlimited IP risk for a model they didn't build. Sometimes that's true. In that case, the fallback isn't to accept silence. It's to get the vendor to flow down whatever indemnity the underlying foundation model provider offers them, and to disclose in writing which specific IP protections, if any, exist at each layer of the stack you're paying for.
If a vendor refuses even that disclosure, treat it as information. A vendor that won't tell you where the IP risk sits in its own supply chain either hasn't mapped it or doesn't want you to see it, and either way you're the one left exposed when a plaintiff's lawyer eventually does the mapping for you. Also read: How Founders Can Structure a Non-Compete That Survives an Acquisition • How a Liability Waterfall Actually Works Across Your AI Vendor Stack • How to Calculate Lead Conversion Velocity Before the Blame Game Starts
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.