IP ownership in AI projects

Five assets, not one

A typical AI project produces source code, a trained model (weights and architecture), training and evaluation datasets, prompts, fine-tuning recipes and documentation, and outputs generated for users. Each has a different legal character. Code and documentation are copyright works. Model weights are protected, if at all, as trade secrets and possibly as copyright or database rights depending on the jurisdiction. Datasets may be copyright compilations, database rights (in the United Kingdom and Malta) or simply confidential information, and may carry third-party licence terms. Outputs may have no copyright at all where no human authored them. Ownership documents that speak only of 'software' leave most of this unaddressed.

Where ownership leaks

The default rules allocate IP by who created it and under what relationship. Employees' work in the course of employment generally belongs to the employer in all ten LegalMatrix jurisdictions, with nuances (Japan's rules on employee inventions, California's section 2870 carve-out, moral rights in civil law countries). Everything else leaks:

• Founders who wrote code or collected data before the company existed own it personally until they assign it.
• Contractors and outsourced developers own what they create unless the contract assigns it; a licence is not ownership.
• University or research collaborators are bound by their institution's IP policy, which often claims a share.
• Customers whose data trained the model may have terms that prohibit training or claim derived IP.
• Open-source components bring licence obligations that can reach the whole model or codebase.
• Cloud and model providers' terms may restrict use of outputs or claim rights in fine-tuned models.

Investors and acquirers find these gaps in due diligence; the time to close them is before the round.

Assignment language also has to match the asset. A clause that assigns 'all copyright in the software' does not clearly reach model weights, datasets or know-how, and a clause that assigns 'all intellectual property' may be read in some jurisdictions as excluding future rights or confidential information. The LegalMatrix forms assign present and future rights in listed categories (code, models, data, documentation, prompts, inventions, confidential information), include a further assurance obligation to sign office recordals, and deal with moral rights in the form the jurisdiction permits. Patents are the one category handled separately: an invention in an AI method may be patentable in some of the ten jurisdictions and not others, and a registered patent agent should scope any filing before the assignment language is finalised.

Data: provenance and permission

The question in diligence is no longer only 'do you own the model' but 'were you allowed to train it'. For each dataset the company should be able to show its source, the licence or contract under which it was obtained, whether it contains personal data and the legal basis for processing it, and any restrictions on commercial use or model training. Scraped data is the highest-risk category; the position on text and data mining differs between jurisdictions (Singapore and Japan have computational analysis exceptions; the United Kingdom's is limited to non-commercial research; United States fair use is contested in litigation). The data protection laws of every LegalMatrix jurisdiction apply to personal data in training sets, and the EU AI Act, which applies in Malta, adds transparency obligations for general-purpose models. The LegalMatrix technology licence agreement includes a data provenance schedule for this reason.

Models and outputs

Keep model weights as trade secrets: access controls, confidentiality agreements and a register of who has had access are what a court will ask for. Where the model is licensed to customers, the licence should prohibit extraction, reverse engineering and use of the model to train a competing model, and should say who owns fine-tuned versions and improvements. For outputs, state the position contractually even where copyright is doubtful: the customer owns outputs generated for it, subject to the provider's right to use them for service improvement if that is the commercial deal, and neither party warrants that outputs are free of third-party rights beyond defined limits.

Putting the IP in the right entity

Groups usually want their AI assets in one IP-holding company that licenses to operating subsidiaries. That requires a chain of assignments: founders and contractors to the company that engaged them, then that company to the IP holder, each in writing and, for registered rights, recorded at the office. Moral rights are waived or consented to where the law allows (common law jurisdictions) or made the subject of an agreement not to exercise them (Japan, Taiwan, Malta). The licence back to operating companies should match the transfer pricing documentation; see the LegalMatrix guide on intercompany agreements. The LegalMatrix IP assignment and licence product delivers the chain for one jurisdiction, signed off by a lawyer admitted there.

Contract checklist for AI vendors and customers

• Define the licensed technology to include weights, prompts and fine-tuning materials, not only 'software'.
• Allocate ownership of improvements, fine-tuned models and feedback explicitly.
• Prohibit training competing models on outputs or the model.
• Warrant data provenance to a defined standard and allocate liability for third-party claims.
• Address personal data under the applicable privacy law with a processing schedule.
• Carry open-source obligations through to the customer where the licence requires.

Sequence for a company preparing for investment

Inventory the assets and who created them; obtain founder and contractor assignments; audit open-source and third-party data terms; document trade secret controls for the model; file trade marks for the product name; and assign and license into the IP holder. LegalMatrix prepares the assignment, licence and NDA documents, a registered trade mark agent or attorney files the marks, and GoGoTrust records the IP in the company's register so that the diligence answer is a list, not a search.