If an AI reads a copyrighted contract template and produces something 'inspired by' it, is that a derivative work? A framework for thinking through the IP and ownership questions that law firms, insurers, and compliance teams need to be asking right now.
Founder of Grayhaven
Founder of Grayhaven
I'm an AI engineer, not a lawyer. Keep that in mind as you read this. Nothing here is legal advice. What I can offer is a framework for thinking through questions that I watch legal and compliance professionals navigate when they start actually deploying AI systems — and a set of questions worth taking to your own legal team before you're forced to.
Here's the question that started me down this path: if an AI model reads a copyrighted contract template and produces something "inspired by" it, is that a derivative work?
It sounds academic. It isn't.
Software developers will recognize a technique called clean room implementation. The classic example: you want to implement functionality similar to a competitor's proprietary system without exposing yourself to copyright claims. So you have Team A read the original code and write detailed functional specifications — what it does, without capturing how it does it. Team B then implements from those specs, never having seen the original source. The argument is that Team B's work isn't a derivative work of the original, because Team B never had access to the protected expression.
It's never been a perfect defense, but courts have found it persuasive in some cases.
Now apply that to an AI model. The model has been trained on essentially everything — legal briefs, contract templates, policy forms, court filings, regulatory guidance documents, and vast quantities of copyrighted text that nobody explicitly licensed for this purpose. The model has "read" it all. There is no Team B that hasn't seen the source material.
When your associate uses an AI to draft a motion, or your underwriting team uses AI to generate endorsement language, or your compliance function uses AI to draft a regulatory filing — that output is coming from a model that absorbed an enormous amount of material whose ownership and licensing status is contested and, in some cases, actively litigated.
I want to be factual here, because the landscape is genuinely uncertain and anyone who tells you otherwise is guessing.
Copyright in AI outputs. The US Copyright Office has been clear on one point: works generated by AI without meaningful human creative input are not eligible for copyright protection. You cannot own what a machine generated autonomously. This matters for your firm's work product: if AI drafts something and a human rubber-stamps it, that output may not be protectable.
The gray zone most firms actually operate in. The more interesting question isn't pure AI output. It's AI-assisted work, where a human directed the AI, reviewed the output, made substantive edits, and finalized the product. That work almost certainly does have copyright protection, proportional to the human creative contribution. The US Copyright Office has started issuing guidance on this — requiring applicants to disclose AI-generated portions and disclaiming protection for those portions specifically. This area is evolving quickly.
Training data litigation. Multiple major cases are working through the courts right now — the New York Times v. OpenAI case being the most prominent — that challenge whether using copyrighted works to train AI models constitutes infringement. The outcome of these cases will significantly shape what liability the model providers carry, and potentially what downstream liability flows to users of those models. Nobody knows how this resolves yet.
The reproduction question. There's a distinct and underexamined risk: AI models sometimes reproduce content verbatim or near-verbatim from their training data, particularly for well-represented content like common legal phrases or standard contract boilerplate. If your AI generates language that is substantially similar to a copyrighted source in its training data, your firm is the one holding that output. Whether that creates infringement exposure depends on facts nobody has litigated yet.
If your associates use AI to draft briefs, motions, or contract language, a few questions are worth having clear answers to:
Who owns the AI-assisted work product? Your engagement agreements almost certainly assign work product to the client. But if the copyright in that work product is uncertain because of ambiguous human contribution, what are you actually assigning? This is a question for your partnership agreement and your client contracts, not just an abstract IP question.
What if the model reproduced something? Courts have started requiring attorneys to certify that AI-generated content has been reviewed for accuracy. Fewer have addressed what happens if that content contains elements reproduced from a source the attorney never saw. Your professional responsibility exposure here is unsettled.
Disclosure obligations. Some courts are now requiring disclosure of AI-assisted drafting. Some clients will want to know. Some might object. Do you have a policy on when and how you disclose?
Your AI vendor's terms. The major AI providers have different positions on output ownership. OpenAI's terms generally say you own your outputs. But terms change, and what "you own it" means when the output might incorporate third-party content is a different question than what the terms say.
Policy language is heavily standardized for good reason — consistency in coverage interpretation. But that standardization means policy forms, endorsements, and coverage language are extensively represented in AI training data. ISO forms, NCCI filings, carrier-specific endorsements: all of it.
When an underwriter or a policy drafting team uses AI to generate or modify policy language, they may be producing text that is structurally and substantively similar to language with contested ownership. The practical exposure here is probably low for standard industry language — but it's not zero, and it's a question worth putting to your reinsurers and your E&O carrier.
Speaking of E&O: does your professional liability coverage extend to claims arising from AI-assisted work product? Not just errors in the output (which are a separate conversation), but IP-related claims related to the output itself. Most E&O policies predate widespread AI use. This is worth a specific conversation with your broker — not an assumption.
If AI generates your compliance documentation, regulatory filings, or responses to examinations, the IP questions are somewhat different — regulatory filings generally aren't protected expression in the way creative works are — but the audit trail questions are significant.
If a regulator asks how you produced a particular document or analysis, "we used AI" is now a meaningful statement that warrants a follow-up. Which AI? What was the human review process? Can you demonstrate that a qualified person verified the output? The process of generating AI-assisted compliance documentation needs to be documented as carefully as the document itself.
I work at LexisNexis Risk Solutions, which means I live with the reality that legal research platforms have been AI-powered for years. Westlaw's AI and LexisNexis's AI tools are pulling from vast corpora of legal materials to surface research, suggest language, and synthesize findings.
The line between "AI-assisted research" and "AI-generated content" has been blurring for longer than the current AI conversation suggests. What's changed is the explicitness: attorneys are now using general-purpose models that don't have the same licensed-content architecture that legal research platforms are built around. That's a materially different risk profile, and most firms haven't thought through the distinction.
I want to be direct about what I'm offering here: not answers, but a better set of questions to ask people whose job is to answer them.
On work product and ownership: Review your AI vendor's terms of service specifically on output ownership. Ask whether your engagement agreements need language addressing AI-assisted work product. Determine whether your copyright registration practice needs to change to account for AI contribution disclosures.
On reproduction risk: Ask whether you have or need a policy on scanning AI outputs for potential verbatim reproduction. Some organizations are beginning to treat this the way they treat plagiarism screening for academic work.
On E&O coverage: Have an explicit conversation with your broker about whether AI-generated or AI-assisted work product creates coverage gaps. Ask whether any exclusions for intellectual property claims would apply to AI output.
On professional responsibility: If you're a law firm, check your state bar guidance. Several state bars have issued formal or informal guidance on AI use in practice. ABA Formal Opinion 512 (2024) addresses this at the national level. These obligations are real and enforceable.
On disclosure: Decide your firm's policy on AI disclosure to clients and to courts before a situation forces the decision. A proactive policy is better than a reactive one.
The one concrete thing I'll say with some confidence: regardless of how the legal questions resolve, firms that can demonstrate a clear human-in-the-loop workflow for AI-assisted work product are in a better position than firms that can't.
If you can show that a qualified professional directed the AI, reviewed the output, made substantive judgments about what to keep and what to change, and exercised professional judgment in finalizing the product — that documentation supports both copyright ownership claims and professional responsibility compliance. It's also the kind of audit trail that regulators and courts are going to start expecting.
I build AI workflows for legal and insurance teams, and in every implementation I work on, the process documentation for AI-assisted output is as important as the output itself. Not as an afterthought. As a design requirement.
The IP questions raised by AI-generated work product are going to be resolved by courts and regulators over the next few years, in ways that are hard to predict right now. The firms that will navigate it best are the ones that built good practices before they were forced to — not because they saw the specific ruling coming, but because they treated the uncertainty seriously.
If you're trying to figure out what that looks like for your team's specific AI workflows, I'm happy to think through it with you. Reach out through the contact form.
More insights on AI engineering and production systems
We help legal, insurance, and compliance teams implement AI that saves time and reduces risk. Let's talk about your needs.
