IP, Licensing & Feedback Loops

The market has converged on output ownership: the customer owns what the AI generates. Most vendors assign their interest in outputs. That sounds clean. But LLMs are stateless and probabilistic — two customers can receive identical outputs from similar prompts. Ownership does not mean exclusivity. The standard vendor clause assigns output ownership but says nothing about whether the vendor retains any license back to use those outputs, and does not address the reality that identical outputs can reach multiple customers.

The enterprise counter adds two critical elements: an explicit acknowledgment that outputs are not exclusive (reflecting how the technology works), and a limitation of the vendor's license back to service provision only. In one financial services negotiation, the deal nearly collapsed over a perpetual license-back clause buried in the vendor's output IP section — the vendor had offered clean output assignment but had also drafted an unbounded right to use all outputs for 'model improvement.' The fix: limit the license to the term of the agreement, exclude outputs used in production, and add a deletion obligation at termination.

Training data provenance is the gap the market has not closed. The vendor indemnifies for output infringement, but no vendor warrants that their foundation model was trained on lawfully obtained data. This is the invisible layer of liability. A training data class action can reach the vendor's model behavior in ways no output indemnity covers. Enterprise buyers should ask for a representation that the vendor's model provider has obtained training data lawfully to the best of its knowledge. Most vendors will resist. The fallback: a commitment to notify the customer if the vendor becomes aware of a training data challenge, plus a right to terminate if the challenge materially impacts the service.

The derivative data gap must be closed at the definition level (Chapter 5) and acknowledged in the IP section. If derivatives are not in the Customer Data definition, the vendor owns them by default. The IP section must confirm: 'For clarity, Customer retains all right, title, and interest in and to Customer Data as defined in Section [X], including all derivatives thereof.'

The training prohibition has reached near-consensus — vendors will not train on customer data. The debate has shifted from whether to how verified. The vendor's no-training promise must flow down to every third-party AI provider in the supply chain. The standard vendor language — 'Provider will not use Customer Data to train, improve, or modify any models' — does not specify which models (the vendor's or the model provider's), does not flow down to the model provider, and the vendor may not control the provider's data practices at all. The enterprise counter requires the vendor to contractually bind every Third-Party AI Provider.

The feedback trap is the most dangerous clause in a standard AI MSA: 'Customer grants Provider a perpetual, irrevocable, worldwide, royalty-free license to use any feedback, suggestions, or ideas provided by Customer regarding the Services.' This is a data capture mechanism. Your feature requests, usage patterns, and improvement ideas are owned by the vendor forever. In an AI context, this is a training data pipeline disguised as a collaboration clause. The fix: void the assignment, license feedback for the vendor's internal use only, and explicitly exclude Customer Data and Customer Confidential Information from the feedback license.

Red Flags

• A feedback clause that assigns all customer suggestions, feature requests, or improvement ideas to the vendor in perpetuity
• A license grant surviving termination for 'improving and enhancing' the service — a training license by another name
• An IP section silent on derivatives — no acknowledgment that derivative data belongs to the customer
• A no-training clause with no flowdown language — applies only to the vendor's models, not the model provider's
• Output assignment without a matching prohibition on the vendor's license-back for model improvement
• No training data provenance representation and no notification obligation if a training data challenge arises

Sample Clause (Illustrative)

"As between the parties, Customer retains all rights in Customer Data. Provider retains all rights in the AI models, algorithms, and related technology. Customer is granted a non-exclusive, perpetual, worldwide, irrevocable, royalty-free license to use AI-generated outputs for Customer's business purposes. Provider shall not, and shall contractually require that no Third-Party AI Provider shall: (a) use Customer Data to train, improve, or modify any machine learning models (including foundation models, fine-tuned models, and classification models), (b) retain Customer Data after it is no longer needed to provide the Services, or (c) extract or derive information from Customer Data for any purpose other than providing the Services to Customer. Customer may provide feedback regarding the Services. Provider may use such feedback for internal purposes only. For clarity, the feedback license does not include the right to use Customer Data or Customer Confidential Information."

Grants a broad output license instead of assigning IP — functionally identical for the buyer's business purposes, but the vendor avoids the complexity of assigning ownership in model-generated content. The license is perpetual and irrevocable, so the buyer keeps outputs regardless of the relationship. The training prohibition flows down to every third-party model provider — this is the critical piece the standard vendor clause omits. The feedback clause is narrowed to internal use only, with an explicit exclusion of Customer Data and Confidential Information, closing the training pipeline that standard feedback clauses leave open.

Illustrative only. Clause language requires adaptation to your jurisdiction, deal context, and risk profile. Not legal advice.