Ahmed Shaami
A client hires you to build a brand identity. You generate forty logo directions with an image model, refine three by hand, and deliver the final files under a contract promising the client owns full copyright in everything you handed over. Under current US Copyright Office guidance, parts of that deliverable may not be copyrightable by anyone, which means you cannot assign ownership you never had. An AI clause is the contract language that closes that gap by stating where AI was used, who owns which parts, and who carries the risk if a third party challenges it.
This is general information for freelancers rather than legal advice. If a contract is large enough to hurt you, have a lawyer in your jurisdiction read it.
The ownership promise is the real exposure
Almost every freelance contract contains an assignment clause. You transfer all rights in the deliverables to the client on final payment, and you warrant that the work is original and does not infringe anyone else. Both halves of that sentence get complicated once a generative model touches the file.
The US Copyright Office addressed this directly in Part 2 of its report on copyright and artificial intelligence. Human authorship remains a requirement, so work generated entirely by AI is not copyrightable, and prompting alone does not create authorship no matter how detailed the prompt. Where a work mixes human and machine contributions, only the human contributions are protectable, and applicants must identify and disclaim the AI-generated portions when registering. You can read the Copyright Office report on copyrightability in full, and Jones Day published a readable summary of the human authorship analysis.
The Office has registered hundreds of works containing AI-generated material, with registration covering the human author contribution. So the outcome is rarely “you own nothing”. It is that the boundary of what the client owns is narrower and blurrier than your contract claims, and you are the one who warranted otherwise.
The Authors Guild summary of what the report means for creators is a useful reference to send a client who pushes back on the language.
Four clauses worth adding
1. Disclosure and consent
State that you may use AI-assisted tools in producing the deliverables, and name the categories rather than specific products so the clause does not go stale every quarter: text generation, image generation, code assistance, transcription, translation. Give the client a mechanism to restrict specific categories in writing before work starts. This turns a possible later accusation into a documented agreement.
2. An IP carve-out for AI-generated portions
Replace a blanket ownership warranty with something you can actually stand behind. A workable structure: you assign all rights you hold in the deliverables, and you warrant originality and non-infringement to the extent of your own authorship, while noting that AI-generated components may not attract copyright protection under applicable law and are delivered on that basis. The client still gets everything you can give. They no longer get a promise that legally cannot be kept.
3. Training-data exclusion for client materials
Clients increasingly care more about this than about ownership. Commit that client-supplied materials, brand assets, internal documents, and confidential inputs will not be submitted to tools that train on user data, and that you will use enterprise or opt-out settings where the work requires a model. This is a promise you can genuinely keep, and it is the one that wins trust with corporate clients. It also forces you to actually check the settings on the tools you use, which most freelancers have never done. Our roundup of AI tools that actually save freelancers time is a reasonable place to start that audit.
4. A liability cap tied to the fee
Cap your total liability under the agreement at the fees paid for the affected deliverable, and exclude indirect and consequential damages. Without this, a claim over an AI-generated asset in a national campaign is theoretically unbounded against a freelancer who was paid a few hundred dollars. A cap is standard commercial practice and clients rarely fight it.
Disclosure stopped being purely a courtesy
For freelancers with European clients, transparency requirements now sit in law rather than in etiquette. If you build a client-facing chatbot or produce synthetic audio and video, the disclosure obligations apply to the deployed system, and your client will look to their contractor to have handled it. We covered the enforcement position in our report on EU AI Act transparency rules for chatbots and deepfakes.
Platform terms are a second layer. Upwork publishes its own position on intellectual property rights on the platform, and where a platform contract governs, its terms sit over whatever you and the client agreed privately. Read them before you assume your custom clause controls.
When a client bans AI outright
Some will, and the ban is usually broader than they intend. A client who says “no AI” is often objecting to generated final assets while having no issue with transcription, grammar checking, or code completion. Ask them to specify which outputs they want free of AI generation rather than which tools you may open.
If they hold the line on a total ban, price it. Fully manual production of assets that would otherwise be generated takes longer, and the rate should reflect that. Accepting a no-AI clause at your AI-assisted rate is how freelancers end up working at a fraction of their effective hourly.
Also, keep evidence. If you agreed to restrictions, retain the drafts, layered files, and version history that demonstrate human production. That record is worth more than the clause itself if the question ever comes up.
Raising it without losing the deal
Do not send a client four dense paragraphs of new legal text with no framing. Lead with what it protects them from. Something close to: “I have added a short section covering how AI tools are used on this project, how your materials are kept out of training data, and how ownership works for any generated components. It is there so neither of us gets surprised later.”
Framed that way, the training-data commitment reads as a benefit, which it is, and the ownership carve-out reads as precision rather than as you withdrawing something. Most clients sign it without comment. The ones who ask questions are usually the ones already thinking about this, and they will respect that you were first.
Put the clause in your template once and it protects every engagement afterwards. Leaving your contract silent on AI does not make the risk disappear. It just means the default answers are set by whoever reads the agreement after something goes wrong.
