An investment firm co-develops a custom large language model with an AI vendor under a joint development agreement (JDA). The firm provides proprietary trading data for fine-tuning; the vendor provides the base model architecture and engineering labor. The agreement is silent on IP ownership of the fine-tuned model weights. Under general U.S. intellectual property principles, what is the most likely default ownership outcome if the JDA is not clarified?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Todd Lammle: A silent JDA on IP is like a prenuptial agreement that forgets to mention the house — whoever contributed something gets a claim, and that means both parties own the keys. Under 35 U.S.C. § 262, joint inventors (or co-developers under work-for-hire ambiguity) can each use and license jointly owned IP without permission from the other. That's a massive commercial risk for the firm.
Full explanation below image
Full Explanation
When a joint development agreement is silent on intellectual property ownership, U.S. law defaults to principles that frequently produce uncomfortable outcomes for investment firms. Under 35 U.S.C. § 262, each co-owner of a jointly held patent (or analogously, jointly developed IP) may make, use, sell, or license the work without accounting to the other owner. Applied to AI model weights, this means the vendor could theoretically license the fine-tuned model — which encodes the firm's proprietary trading patterns — to a competitor.
Option A overstates the significance of data contribution. While the firm's proprietary data drove the fine-tuning, data itself is often not copyrightable (collections may be, under thin copyright doctrine, but the act of providing data for training does not automatically transfer ownership of the trained artifact). The legal treatment of training data contributions in AI IP is an evolving area, but silence in the contract does not resolve it in the data provider's favor.
Option C similarly overstates the vendor's position. Fine-tuned model weights are a transformation of the base model, but they also embody the engineering contribution of the training process and the informational contribution of the training data. Copyright derivative work doctrine applies to expressive works, and its application to neural network weights is legally contested.
Option D correctly identifies that AI-generated outputs face copyright barriers (as confirmed by the U.S. Copyright Office's 2023 guidance on AI-generated works), but the weights themselves — as a product of human-directed development — may qualify for trade secret protection or patent claims on novel architecture modifications. Omitting these protections from the JDA is the error.
CFIA candidates should know: always negotiate explicit IP assignment or license-back provisions in JDAs before development begins. The WIPO AI and IP Policy guidance and the USPTO's 2024 AI Inventorship guidance are the relevant reference frameworks.