Can You Warrant IP on AI-Generated Work?

Most services agreements contain some version of this sentence: "Vendor represents and warrants that the Deliverables are original and that Vendor owns all rights necessary to assign them to Client."

Five years ago, that was boilerplate. Today, if your team uses AI tools to produce deliverables, it may be a promise you can't keep.

The problem

US copyright law requires human authorship. The Copyright Office has said so in its guidance on AI, and the D.C. Circuit affirmed the human authorship requirement in Thaler v. Perlmutter. Content generated by AI with no meaningful human creative contribution may not be protected by copyright at all. Prompting alone generally doesn't clear the bar. Human-authored expression, and human selection, arrangement, and modification of AI output, can.

So take a vendor who delivers code, copy, or designs that were AI-assisted. Ask what it means to "own all rights" in the parts a model generated. In some cases, there may be nothing to own. And a warranty that the work is original or fully protectable can be breached the day you hand it over, with nobody doing anything wrong.

What vendors should do about it

You don't need to stop using AI, you just need to stop over-promising. In vendor-side drafting, I focus on five things.

  1. Assign what you have, warrant what you can verify. Assign "all right, title, and interest, if any" in the Deliverables. Don't warrant that any particular element is copyrightable or exclusive to the client.

  2. Warrant process, not outcome. Commit to what you control: you use commercially reasonable efforts, you use enterprise-grade tools with no-training terms, and a human reviews the work. That is a promise you can keep and prove.

  3. Qualify non-infringement to knowledge. No vendor can guarantee what a model's training data did or didn't include. "To Vendor's knowledge, without independent investigation" is doing real work here.

  4. Define AI Tools and disclose them. Put a clause in the agreement that says you may use AI tools in delivering the services, subject to the tool providers' terms. Disclosure up front is protection, but silence is a dispute waiting to happen.

  5. Watch the indemnity. If the IP indemnity is uncapped and covers all third-party claims, AI-generated output is where it will bite. Cap it, or carve AI-related claims into a separate, capped bucket. And have the client warrant the inputs, since their prompts and materials are theirs.

A note for the other side of the table

If you're a client asking for a clean IP warranty on AI-assisted work, expect pushback, and expect it to be reasonable. The better ask is transparency: what tools, what review, what terms.

The bottom line

AI in delivery isn't the risk. Contracts written for a world before it are. Fixing this takes a handful of drafting changes, and it's much cheaper to make them before a dispute than after.

—James

If you're a vendor and your form agreement hasn't been updated for AI, that's worth a conversation. This post is general information, not legal advice.

Previous
Previous

Your Personal Email Is Not Confidential

Next
Next

The Meta Settlement Fixes One Privacy Problem by Creating Another