Skip to content
FixMyContract
All guides
Photograph of four blank sheets of paper laid out in a row like steps, with small tabs at their edges

Contracts

Your Contract Has an AI Clause. What Does It Mean? (2026)

By FixMyContractLast updated Sep 16, 202610 min read

Short answer: an AI clause is a contract term about how you are allowed to use AI tools while doing the work — and there are four common kinds, which do very different things. One asks you to tell the other side you used AI. One bans you from using it at all. One is really a confidentiality rule wearing an AI costume. And one asks you to promise a human made the work, which is the version most likely to cause you a problem later. None of these are the same obligation, and the word "AI" in all four is what makes them easy to skim past.

Why this clause suddenly appeared in your contract

Two transparency rules — one European, one Californian — both became applicable on 2 August 2026. Neither one gives you, personally, a duty to disclose AI use to your client. Article 50 and SB 942 regulate the providers and deployers of AI systems — typically the companies building or professionally operating the tools — and the duties they create run toward end users, the public, or regulators, not toward a contract counterparty. But once large organisations start building disclosure machinery for their own compliance, that language flows downhill into the ordinary contracts they hand to their suppliers.

The EU side. Article 50 of the EU AI Act applies from 2 August 2026. In the European Commission's own words, it requires that "individuals concerned are informed that they are interacting with an AI system, unless this is obvious," and that AI-generated audio, image, video or text carries machine-readable marks. Deployers must disclose deepfakes on first exposure, and must label AI-generated text published to inform the public on matters like politics, public health or the environment — unless that text went through human review or editorial control. There is one transitional carve-out: content generated before 2 August 2026 does not have to be retroactively marked, and the marking obligation for systems already on the market runs to 2 December 2026.

The California side. SB 942, the California AI Transparency Act, is aimed at generative AI providers above a user threshold: detection tools, provenance marking, disclosures. Its requirements became operative on 2 August 2026 — not 1 January 2026.

A common mistake worth correcting. A lot of writing online still says SB 942 took effect on 1 January 2026. It did not. AB 853, signed in October 2025, moved the operative date to August 2, 2026 — the bill text sets that date in Section 6 — and added later phases for large online platforms and generative-AI hosting platforms (1 January 2027) and for capture device manufacturers (1 January 2028). If a contract or a policy you are reading cites the January date, it was drafted from stale information, which is usually a sign the rest of the AI language was too.

The public-sector side, still in draft. In the United States, the General Services Administration has proposed a clause, GSAR 552.239-7001, that would require federal contractors to disclose every AI system used in performing the contract to the contracting officer within 30 days of award — and to flow that duty down to subcontractors. A first draft appeared in March 2026; a revised version was published on 17 June 2026 with comments due 3 August 2026. As of 16 September 2026 we found no final rule, so this one is a signal about where standard contract language is heading, not a rule anyone is bound by yet.

The practical point for you is simpler than any of the above: your obligation comes from the contract, not the statute. If your client's agreement says you must disclose AI use, breaching it is a breach of contract whether or not any law required that clause.

The four AI clauses you will actually see

1. The disclosure clause — "tell us what you used"

Typical wording: "Contractor shall disclose in writing any artificial intelligence or machine learning tools used in the creation of the Deliverables, including the tool, the version, and the purpose of use."

What it really does. It creates a record. Not a ban — a paper trail. The risk is not that you used a tool; it is that a disclosure duty is easy to forget six months into an engagement, and "I didn't think spell-check counted" is not much of a defence once the clause says any.

The word that decides everything is the scope word. "Any AI tool" can be read to include the grammar checker in your writing app, the autocomplete in your code editor, and the transcription service that turned the kickoff call into notes. That is almost certainly not what the client means, but the clause is what governs, not what they meant.

What to ask for: a materiality line and a timing line. Something like: disclosure applies to generative AI tools used to produce substantive portions of the Deliverables, not to routine assistive features such as spelling, grammar, autocomplete or transcription; disclosure is made once at delivery rather than continuously. You are not resisting transparency. You are making transparency a task you can actually complete.

2. The prohibition clause — "do not use AI on this work"

Typical wording: "Contractor shall not use generative artificial intelligence in the performance of the Services."

What it really does. It turns a tool choice into a breach condition. This one is increasingly common where the client's own customers demand it, in regulated work, and in creative fields where the client is selling human authorship as the product.

The trap is scope again, in the other direction. A flat ban on "generative AI in the performance of the Services" may technically cover research you did before drafting, the tools inside software you rely on, and features your client asked you to build with AI. Read it against your actual workflow, honestly, before you agree to it.

What to ask for: carve the ban down to what the client actually cares about. Contractor shall not use generative AI to produce the final Deliverables or any material portion thereof; this does not restrict use of AI for research, ideation, internal drafts, or standard features embedded in general-purpose software. And if you cannot deliver on the client's timeline or price without the tools they are banning, that is a scope and rate conversation to have before signing, not a rule to accept and quietly break.

3. The confidentiality-extension clause — an NDA in disguise

Typical wording: "Contractor shall not input, upload, or otherwise submit any Confidential Information into any third-party artificial intelligence service."

What it really does. This is not really an AI clause. It is your confidentiality clause extended to cover a specific channel, and it is the version of the four that we think is most often reasonable on its face. The client is saying: our documents do not leave our perimeter, and a chat window is a way out of the perimeter.

Why it deserves attention anyway. "Any third-party AI service" is broad. It may cover tools that are contractually committed not to train on your inputs, and enterprise tools your client uses themselves. And it sits next to a separate question worth understanding on its own: what happens to text you type into an AI tool, legally. We wrote about that in Is anything you type into an AI confidential? — in US courts, those chats are not privileged, which is a good reason to take this clause seriously rather than treat it as boilerplate.

What to ask for: an approved-tools path instead of a blanket ban. Contractor may use AI services that the Client has approved in writing, or that are configured so that Client data is not used to train third-party models. That gives the client control and gives you a way to work.

4. The human-authorship warranty — the one to read twice

Typical wording: "Contractor represents and warrants that the Deliverables are the original work of human authors, are not generated in whole or in part by artificial intelligence, and that Contractor owns all rights therein, which are hereby assigned to Client."

What it really does. It bundles three different promises into one sentence: originality, no-AI, and a full IP assignment. Breaking any one of the three breaks the warranty — and a warranty breach usually sits next to an indemnification clause, meaning you may have agreed to cover the client's costs if the promise turns out to be untrue.

The part most people miss. In US copyright practice, material generated wholly by an AI system, without human creative contribution, is not registrable as a copyrighted work. The Copyright Office has taken that position, the D.C. Circuit affirmed it on 18 March 2025 in Thaler v. Perlmutter, and the Supreme Court declined to review the case on 2 March 2026 (docket 25-449, "Petition DENIED"). The Copyright Office's January 2025 report drew the line between assistive use of AI, which can still leave a protectable human-authored work, and output produced by prompting alone, which it concluded is not protected.

Put those two facts side by side and the clause gets uncomfortable. If part of a deliverable was produced by prompting alone, there may be no copyright in that part for anyone to own — which means a clause promising to assign "all rights" in it is promising something that may not exist, while the no-AI warranty next to it has already been broken. The client is not necessarily getting what they think they bought either. This is a drafting problem, not a trick, and it is usually fixable by asking.

What to ask for: separate the three promises, and make the AI one match reality. Contractor warrants that the Deliverables were created under human direction and review, and that Contractor has the rights necessary to assign the Deliverables to Client. Contractor's use of AI-assisted tools in accordance with Section [X] does not breach this warranty. If you genuinely produced the work without generative AI, you can still say so — just say it as its own sentence, so you know exactly what you signed.

The clause pointing the other way: your work as training data

There is a fifth version, and it runs in the opposite direction — the client wanting the right to use your delivered work to train AI models. The Authors Guild, for instance, publicly recommends that authors add a clause to publishing and distribution agreements prohibiting AI-training uses of their work, which tells you how contested this has become in creative industries.

If you are delivering creative work, look for a training-rights grant hiding inside a broad licence. Client may use the Deliverables for any purpose, in any medium, now known or later developed is broad enough to cover training a model on them. If that is not what you intended to sell, ask for a line that says so: Client shall not use the Deliverables to train, fine-tune, or otherwise develop machine learning models, and shall impose the same restriction on downstream licensees. Whether a client accepts it is a negotiation — but it is a normal ask in 2026, not an exotic one. Note this cuts both ways: if you reuse your own templates, components, or prior work across clients, check that this restriction does not accidentally limit your own background IP.

Quick reference

If the clause says…It is really about…Reasonable thing to ask for
"disclose any AI tools used"record-keepinglimit to generative AI used for substantive work; disclose once at delivery
"shall not use generative AI"breach risk on tool choicelimit to final deliverables; allow research, drafts, embedded features
"do not input Confidential Information into AI services"confidentiality, via a specific channelan approved-tools list, or a no-training-on-our-data standard
"warrants the work is human-authored… and assigns all rights"warranty + indemnity + IPsplit the three promises; align the AI promise with how you actually work
"Client may use Deliverables for any purpose"possible training-rights grantexplicit no-training restriction, flowing down to licensees

What to do before you sign

  1. Search the document for the strings "AI", "artificial intelligence", "machine learning", "generative", and "automated". These clauses are rarely in a section called "AI." They hide inside Confidentiality, Warranties, Deliverables, and IP.
  2. Check whether the clause has a scope word — "any", "material", "substantive". That single word is the difference between a rule you can follow and a rule you will break by accident.
  3. Check what the clause is wired to. A duty sitting next to an indemnity is a financial exposure; the same duty on its own is an administrative task.
  4. Check the dates it cites. A contract that describes California's rules as effective 1 January 2026 was drafted from stale material.
  5. Say what you actually do. The safest position is not the strictest promise — it is the accurate one.

Sources

Rules described here are US and EU. Contract law and copyright rules differ by country; nothing above should be read as a universal rule.

Reading a clause is one thing; knowing what to ask for is the other half. FixMyContract reads an agreement from your side of it — tenant, freelancer, employee, or business — and tells you both what the clauses that matter do and what to ask for before you sign. Analyze my contract — free

Keep reading