How to Review Ownership of AI Outputs in a Vendor Contract

Tamsin, founder of a 14-person UK SaaS, is about to treat “Customer owns all Outputs” as a title deed because Sales said you own what you generate. Knowing how to review ownership of ai outputs in a vendor contract is a 25-minute hunt: three grants, an if-any assignment, a perpetual licence if copyright never attaches, then a block on reuse to compete.

September 2026. Finance has said yes on a drafting copilot that will sit on customer-facing copy. English law; exclusive courts of England and Wales. The PDF is an MSA — master services agreement, the frame the order form hangs on. Clause 8 is Intellectual Property. Typical mistake: treating “you own the outputs” as a title deed. AE: “same as ChatGPT.”

Clause 8.2: “Customer owns all Outputs.” No “hereby assigns … if any.” No perpetual licence if copyright never attaches. Clause 8.4: “Vendor may reuse similar Outputs for other customers.” Schedule 3 is an exhibit — an attached schedule or appendix — with the train sentence. Friday is booked. The rush is the problem.

The hidden risk is a heading that assigns nothing. Promise Legal (7 July 2026): “if any is doing heavy lifting.” OpenAI (effective 1 January 2026) §4.1 is the sentence she does not have: hereby assigns … if any, in and to Output.

Disclaimer: Checkory provides AI support, not legal advice. Consult a qualified lawyer for binding decisions.

Which grants actually move — inputs, prompts, or outputs?

Inputs, prompts, and outputs are three grants, not one data line. Circle whether files stay yours with only a service-delivery licence to the vendor, whether prompts are named, and whether outputs are owned. Tamsin’s 8.2 names Outputs only — write inputs ? / prompts ? / outputs headline only.

For example, OpenAI §4.1 and Common Paper AI Addendum 1.0 (7 August 2025) split retain Input / own Output. Venable (12 January 2026): name those buckets separately. Do: write three rows. Don’t: treat one “data” heading as three grants.

What does “Customer owns all Outputs” assign if copyright never attaches?

A heading that customer owns AI generated outputs is not an assignment. Demand: Vendor hereby assigns all right, title and interest, if any, in Outputs from Customer’s use. If copyright never attaches, that assignment moves nothing against the world. Tamsin’s 8.2 = headline, no if-any — fail.

England and Wales is not the US fork. CDPA 1988 s.9(3) still treats the author of a computer-generated literary work as the person who made the arrangements necessary. GOV.UK (March 2026) proposes removing that protection; it is not repealed. Subsistence is contested.

US Thaler v. Perlmutter (cert. denied 2 March 2026) is why US vendor paper uses “if any”; it is not Tamsin’s statute. Shumaker / JD Supra: the allocation binds the parties even where output may not qualify for copyright. Do: circle “hereby assigns” and “if any.” Don’t: treat the heading as a title deed.

Why does a perpetual licence still matter when copyright never attaches?

If copyright never attaches, an assignment of copyright is empty. Demand a perpetual, irrevocable, worldwide, royalty-free licence to use, modify, commercialise, and create derivatives from Outputs regardless of whether IP rights subsist. Tamsin’s paper has no fallback — fail.

Promise Legal (7 July 2026): do not accept “if any” without a broad licence that operates regardless of IP status. BindLegal (2026): an assignment of “all copyright” in AI-generated portions can be hollow — back it with a perpetual post-termination licence. In practice the licence is what she can still use after the subscription dies. Do: require the fallback. Don’t: assume “owns” conjures a copyright.

When to treat similar outputs as reuse to compete?

Split two doors. Stochastic similarity is another customer typing a similar prompt and getting similar text. Reuse to compete is the vendor serving her customer-specific output to a rival. Accept the first for generic text if confidentiality still covers confidential prompts. Do not accept the second. Tamsin’s 8.4 opens door two — fail.

OpenAI §4.4 (1 January 2026): Output may not be unique; responses generated for other users are not Customer’s Output. Common Paper §3.2 copies a Similarity of Output disclaimer. tish.law: assign, waive, and agree not to use customer-specific outputs for other customers. Redline: no use of Customer-specific Outputs for other customers or to compete, except to transmit them. Do: split chance-similar from reuse-to-compete. Don’t: let 8.4 stand as “how models work” under Friday pressure.

How do you keep training rights and feedback off this hunt?

If the fight is whether they train or improve the service on her data, that is a different hunt — see vendor AI training rights. Write one sentence, send it there, and come back to 8.2 and 8.4.

If the fight is Slack ideas assigned as Feedback, that is a different hunt — see feedback license of customer ideas. Freeze the packet — the MSA, order form, live terms URL, and every schedule you will execute — then return to outputs. Do: send one sentence each and open 8.2 / 8.4. Don’t: clone those hunts here.

How do you decide assign, licence, or walk?

Assign — keep — only if the log shows three grants named, a present if-any assignment, a perpetual licence fallback, and no reuse of customer-specific outputs to compete. Licence means add the fallback and the non-reuse covenant even if they will not move on copyright theory. Walk — do not sign Friday — if headline-owns, no assignment, no licence, and reuse-for-other-customers remain as a package.

Success bar: a one-page log plus one sentence that would pause Friday. Tamsin already has three: no if-any assignment; no perpetual licence; 8.4 reuse similar. Workflow: three grants → if-any → perpetual licence → no reuse to compete → assign / licence / walk.

Escalate to counsel — your England-and-Wales solicitor — on that package. Verify 8.2 and 8.4 before you sign. Do: spend counsel time on the missing assignment plus no licence plus reuse. Don’t: mark this paper a green light under AE pressure. Optional: upload the same PDF to document analysis for a first-pass — a machine extract of clauses before a human reads every page. A human still opens 8.2 and 8.4 and verifies every High flag — a high-severity extract a human must still open.