When the contract is silent on AI output: what it costs to be wrong
Silence does not default in your favour. When a technology contract says nothing about AI output, ownership, licence scope and liability for that output stay open questions. The party best able to litigate usually wins the gap — not the party who commissioned the work. Fixing this after delivery costs far more than a clause would have cost before signature.
The legal basis
EU copyright law starts from Directive 2009/24/EC, which protects a computer program as the expression of its author's own intellectual creation. It does not itself decide who owns output that a generative system produces without a human author's creative input. That question is left to the contract, or to whichever court is asked to apply it in the absence of one.
A second layer arrives with Regulation (EU) 2024/1689 (AI Act). Its core obligations on providers and deployers of AI systems become directly applicable across the Union from 2 August 2026 (status confirmed as of this review). A contract signed today without an AI clause will still be in force on that date, and the roles it never allocated will still need to be filled by someone.
What silence actually costs
Without a clause, you cannot always show that you own the output your supplier generated. If that output is meant to become a tokenised asset or a registrable right, the gap becomes existential — see The token refers to an unregistered right tokenised rights, which sets out the same failure in the token context. Left unaddressed, the output may be free for anyone to copy, including a competitor, because no clause ever restricted its use.
A second cost sits in liability. If the AI output infringes a third party's rights, an unallocated contract leaves both sides exposed at once. In practice the exposed party is normally the one who deploys the output commercially, whether or not that party wrote a single line of the prompt.
| Element the contract leaves silent | Question it leaves open | Who typically absorbs the cost |
|---|---|---|
| Ownership of the AI output | Can it be registered, licensed or resold? | Whoever holds the weaker documentation |
| Provider or deployer role under the AI Act | Who carries the transparency and marking duty? | Whoever is easier to identify as the acting party |
| Training-data provenance | Does the output infringe a third party's rights? | Whoever deploys the output commercially |
| Notice period for disputes | How long do you have to raise a defect? | Whoever discovers the problem last |
What to do
- Legal counsel audits every AI tool feeding the deliverable, including any sub-processor the vendor uses without naming it.
- The contract owner inserts an express output-ownership clause, kept separate from the general IP assignment wording.
- Procurement allocates the AI Act's provider and deployer roles between the parties, in writing, ahead of 2 August 2026.
- Legal counsel adds a warranty on training-data provenance and a liability cap for third-party infringement claims tied to the output.
- The contract owner sets a notice-of-claim period for disputes about AI output quality or ownership.
- Escalate the gap into the next renewal negotiation rather than wait for a dispute to expose it.
Who decides, and on what basis
An ownership dispute over AI output is decided by the court or arbitral tribunal named in the contract's governing-law clause. It applies that law to the facts of authorship and creative input. Evidence is version history, prompt records and the vendor's own documentation of how the output was produced.
A compliance question that arises once the AI Act applies is a matter for the market surveillance authority designated under that Regulation. It decides on the basis of the technical documentation the provider or deployer is required to hold.
Three situations, three answers
- If your general IP clause assigns "all deliverables" without naming AI output, the assignment is still ambiguous. A court still asks whether the clause contemplated generative output at the time of signature.
- If the vendor runs the AI tool under its own terms of service, those terms may govern the output regardless of what your contract says, unless your contract expressly displaces them.
- If a dispute surfaces after 2 August 2026 and neither party is named as "deployer," the AI Act's own definitions decide who carries the compliance duty — usually whoever put the system into use.
Sources and limit
This page does not cover data-protection settings for AI tools; the default state of tracking consent is addressed separately — see Cookies and what is switched off by default step.
Where the output is meant to end up as a tokenised or assignable right, the same ownership question resurfaces at onboarding. A Counterparty report before onboarding a distributor in practice checks the counterparty side of that gap before you sign anything.
Where the assignment underlying the output is missing a signature rather than a clause, that gap is handled directly by the Document report when an assignment is unsigned and what check.
Verified against: Regulation (EU) 2024/1689 (AI Act); Directive 2009/24/EC on the legal protection of computer programs. Status of the AI Act's application date confirmed as of the date of this review.
Related situations
A similar gap appears when a technology contract ends: see When a migration is blocked by an exit clause.
FAQ
What is the first step when the contract is silent on AI output?
Audit every AI tool in the deliverable chain, including any sub-processor the vendor uses, before touching the clause itself.
Which deadline is already running?
The clock toward 2 August 2026, when the core obligations of Regulation (EU) 2024/1689 (AI Act) become directly applicable, is already running for any contract in force today.
Who decides, and on what evidence?
An ownership dispute goes to the court or tribunal named in the governing-law clause. It is decided on prompt records, version history and the vendor's documentation of how the output was produced.