Artificial Intelligence Models May Not Have Owners

44 Pages Posted: 4 Feb 2025

See all articles by Devin Owens

Devin Owens

University of Akron, School of Law

Date Written: May 01, 2024

Abstract

While new artificial intelligence models see unprecedented investment, serious questions exist about the ownership of the models themselves under existing intellectual property structures. AI models, as compilations of information created largely autonomously by algorithms from sets of training data, may not be suited for the subject matter and inventorship/authorship requirements of traditional patent and copyright protection. The literature assumes that trade secrecy will protect AI models, which are largely kept secret on remote servers away from direct inspection by users, but model extraction attack methods known since 2016 are effective in copying any AI model that can be queried. This potential for copying compromises the viability of trade secrecy as an intellectual property framework for AI models. Model extraction attacks can also extract substantial amounts of training data that the model memorizes during its training period, potentially also compromising the trade secrecy of any proprietary training data. With protection for AI models by other avenues-patent and copyright-also in doubt, it is possible that AI models are not protectable or owned under current law.

Keywords: Artificial Intelligence, Trade Secrets, Trade Secrecy, Patent, Artificial Intelligence Models, Deep Learning, Patents, LLMs, Model Extraction Attacks, Copyright, Anti-reverse Engineering Clause

Suggested Citation

Owens, Devin, Artificial Intelligence Models May Not Have Owners (May 01, 2024). Available at SSRN: https://ssrn.com/abstract=5055010 or http://dx.doi.org/10.2139/ssrn.5055010

Devin Owens (Contact Author)

University of Akron, School of Law ( email )

Akron, OH
United States

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
60
Abstract Views
303
Rank
744,156
PlumX Metrics