What Exactly Is the Model?
AI model escrow and defining deposit scope for AI systems
AI model escrow is the deposit of the materials needed to continue operating, retraining or replacing an AI system if its supplier fails. The drafting difficulty is definitional. A licence that refers to “the model” may mean the weights alone, or the weights with training code, evaluation data, prompts and the surrounding application. Each reading produces a very different deposit.
What does an AI system consist of?
An AI system is a stack of components, each with its own deposit consideration:
Component | Needed to | Deposit consideration |
Model weights | Run the model | Base-model licences may restrict redistribution |
Inference code and serving configuration | Operate in production | A conventional source code deposit |
Fine-tuning and training pipelines | Retrain or update | Code, hyperparameters and environment |
Training and evaluation data | Retrain and validate | Personal data and third-party rights; often described rather than copied |
Prompts, system instructions and guardrails | Reproduce behaviour | Frequently overlooked; treat as source code |
Third-party model services | Operate | Often impossible to escrow; document substitutes |
How should counsel define the deposit scope?
Counsel should define the deposit by purpose. Continuity of operation requires the weights, inference code and prompts. Continuity of development adds the training pipelines and either the datasets or a lawful route to them. The escrow agreement should state which purpose applies and list every component in a schedule, rather than relying on the word “model”.
What limits AI model escrow?
Three constraints recur:
- Base-model licences may prohibit sublicensing or redistribution of weights, which should be checked before any deposit is promised.
- Hosted third-party models called through an API cannot be preserved by escrow. The deposit should instead document how to substitute another model and include the evaluation set used to test the substitution.
- Personal data in training sets engages the Digital Personal Data Protection Act, 2023, which favours depositing descriptions and access routes over copies.
Quick answers
Are weights alone sufficient? For operation, sometimes. For adapting or retraining the system, no.
How is an AI deposit verified? By running inference on a benchmark set and comparing outputs with a recorded baseline.



