Buying a Model: AI Vendor Contracts, Training Rights, Output Ownership, and the Indemnity That Is Not There
By Casey Scott McKay ·
Buying access to a model is not buying software and the standard technology agreement does not fit it. This article works the five terms that actually matter - whether the vendor may train on customer data, who owns the output, what the infringement indemnity covers and what it carves out, what the vendor represents about its training data, and what happens when the model changes underneath you. It explains why output ownership is largely a contractual fiction given the human authorship requirement, and what a customer should therefore ask for instead. It covers the confidentiality problem created by prompts, the subprocessor chain nobody maps, and the evaluation and monitoring obligations that turn a procurement into a governance programme. It closes with the deployment restrictions that matter more than any of the drafting.
IP and Technology > Information Technology | Article | Published 29 December 2024 - Updated 21 January 2025 | Casey Scott McKay - marksy.us
Summary. Buying access to a model is not buying software and the standard technology agreement does not fit it. This article works the five terms that actually matter — whether the vendor may train on customer data, who owns the output, what the infringement indemnity covers and what it carves out, what the vendor represents about its training data, and what happens when the model changes underneath you. It explains why output ownership is largely a contractual fiction given the human authorship requirement, and what a customer should therefore ask for instead. It covers the confidentiality problem created by prompts, the subprocessor chain nobody maps, and the evaluation and monitoring obligations that turn a procurement into a governance programme. It closes with the deployment restrictions that matter more than any of the drafting.
Keywords: AI vendor agreement · training on customer data · input and output definitions · output ownership · no copyright in machine output · human authorship requirement · infringement indemnity · indemnity carve-outs · model provenance · training data warranties · confidentiality and prompts · subprocessor chains · model version changes · evaluation and acceptance · hallucination and accuracy · deployment restrictions · regulated use cases · audit rights · exit and portability · governance triggers
This is premium Marksy content — the full document is available to subscribers.