5

"Individual and Representative Plaintiffs v. GITHUB, INC., a Delaware corporation; MICROSOFT CORPORATION, a Washington corporation; OPENAI, INC., a Delaware nonprofit corporation; [...]" is a class action lawsuit filed against OpenAI et al which includes the allegation of "violation of the Digital Millennium Copyright Act" (p.4) wrt the way Codex, a GPT-3 series large language model, was trained using open source code from Github user repositories. They note "Numerous questions of law or fact common to the entire Class arise from Defendants’ conduct—including [...] Whether Defendants’ conduct violated the Class’s rights under the DMCA when GitHub and OpenAI caused Codex and Copilot to ingest and distribute Licensed Materials without including any associated Attribution, Copyright Notice, or License Terms. (VI)(D)(37)-(VI)(D)(37)(1) The defendants have made a motion to dismiss the case on several grounds stating "the crux of this claim asserts that OpenAI improperly benefited from using Licensed Materials to create Derived Works" (VI)(A)(3)(B) but refuting that "Plaintiffs do not asserts a copyright infringement claim. Instead they allege that Defendants violated the DMCA by [...] (3) knowingly providing CMI that is false by "asserting and/or implying that CoPilot is the author of Licensed Materials"". (VI)(C)(1).

If the DMCA abuse claim of this case gets dismissed, I'm wondering if it would have the effect of nullifying the legal basis to prosecute commercial use of large language models (LLMs) derived from data selected by an even larger model (henceforth, "data derived LLM") -- or at least setting the precedent to do so. Examples of data derived LLMs include Alpaca, Koala, GPT4ALL, and Vicuna, all of which were trained on data produced by OpenAI's models and are currently only available for non-commercial purposes. EG, "the instruction data is based on OpenAI’s text-davinci-003, whose terms of use prohibit developing models that compete with OpenAI." Also worth consideration is this provision in OpenAI's ToS:

(c) Restrictions. You may not [...] (iv) except as permitted through the API, use any automated or programmatic method to extract data or output from the Services, including scraping, web harvesting, or web data extraction; [...]

The precedent set by the action to dismiss or not dismiss the class action lawsuit against OpenAI et al could have an enormous impact on the economics of training LLMs, since (as of Apr 29) it is orders of magnitude cheaper to train a data derived LLM like Alpaca than a new one. Surely OpenAI would attempt to restrict others from using their LLMs to train new ones for fractions of the price (the maximum penalty for violation is suspension and termination ToS 6(a)), but under their current ToS, could they prosecute such incidents? The big question I'm asking is: Is it considered intellectual property theft to capitalize on synthetic datasets produced by OpenAI's LLMs?

1 Answer 1

5

A motion to dismiss sets no precedent

Whether it succeeds or not, it does not result in a judgement on the merits, it is simply an analysis on whether the case as pleaded shows the defendant has a case to answer.

The case would have to go to trial, have a judgement issued (i.e. not settle), and await the result of final appeals (if any) before it would be considered precedent.

As to your final question

Is it considered intellectual property theft to capitalize on synthetic datasets produced by OpenAI's LLMs?

No one knows. Hence the lawsuit.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .