The complaints against OpenAI and Microsoft in New York Times Company v. Microsoft Corporation and Daily News, LP v. Microsoft Corporation include multiple theories––for instance, vicarious copyright infringement, contributory copyright infringement, and improper removal of copyright information. Those theories, however, are ancillary to both complaints’ primary cause of action: direct copyright infringement. While the defendants’ motions to dismiss focus primarily on jettisoning the ancillary claims and acknowledge that “development of record evidence” is necessary for resolving the direct infringement claims, they nonetheless offer insight on how the direct infringement fight might unfurl.

Direct Infringement Via Inputs and Outputs: The Daily News plaintiffs claim that by “building training datasets containing” their copyrighted works without permission, the defendants directly infringe the plaintiffs’ copyrights. Inputting copyrighted material to train Gen AI tools, they aver, constitutes direct infringement. Regarding outputs, the Daily News plaintiffs assert that “by disseminating generative output containing copies and derivatives of the” plaintiffs’ content, the defendants’ tools also infringe the plaintiffs’ copyrights. The Daily News’s input (illicit training) and output (disseminating copies) allegations track earlier contentions of The New York Times Company.

Fair Use Inputs and “Fringe” Outputs: OpenAI’s June arguments in Daily News frame “the core issue”––one OpenAI says “is for a later stage of the litigation” because discovery must first generate a factual record––facing New York City-based federal judge Sidney Stein as “whether using copyrighted content to train a generative AI model is fair use under copyright law.” Fair use, a defense to copyright infringement, involves analyzing four statutory factors: 1) the purpose and character of the allegedly infringing use; 2) the nature of copyrighted work allegedly infringed upon; 3) the amount of the copyrighted work infringed upon and whether the amount, even if small, nonetheless goes to the heart of the work; and 4) whether the infringing use will harm the market value of (or serve as a market substitute for) the original copyrighted work.

So, how might ingesting copyrighted journalistic content––the training or input aspect of the alleged infringement––be a protected fair use? Microsoft argues in Daily News that its “and OpenAI’s tools [don’t] exploit the protected expression in the Plaintiffs’ digital content.” (emphasis added). That’s a key point because copyright law does not protect things like facts, “titles, names, short phrases, and slogans.” OpenAI asserts, in response to The New York Times Company’s lawsuit, that “no one . . . gets to monopolize facts or the rules of language.” Learning semantic rules and patterns of “language, grammar, and syntax”––predicting which words are statistically most likely to follow others––is, at bottom, the purpose of the fair use to which OpenAI and Microsoft say they’re putting newspaper articles. They’re ostensibly just leveraging copyrighted articles “internally” (emphasis in original) to identify and learn language patterns, not to reproduce the articles in which those words appear.

More fundamentally, OpenAI and Microsoft aren’t attempting to disseminate copies of what copyright law is intended to incentivize and protect––“original works of authorship” and “writings.” They aren’t, the defendants claim, trying to unfairly produce market substitutes for actual newspaper articles.

How, then, do they counter the newspapers’ output infringement allegations that the defendants’ tools sometimes produce verbatim versions of the newspapers’ copyrighted articles? OpenAI contends such regurgitative outcomes “depend on an elaborate effort [by the defendants] to coax such outputs from OpenAI’s products, in a way that violates the operative OpenAI terms of service and that no normal user would ever even attempt.” Regurgitations otherwise are “rare” and “unintended,” the company adds. Barring settlements, courts will examine the input and output infringement battles in the coming months and years.

Source: Inputs, Outputs, and Fair Uses: Unpacking Responses to Journalists’ Copyright Lawsuits | American Enterprise Institute – AEI

Even though parents insist that it is good and right to share things, the copyright world has succeeded in establishing the contrary as the norm. Now, sharing is deemed a bad, possibly illegal thing. But it was not always thus, as a fascinating speech by Ryan Cordell, Associate Professor in the School of Information Sciences and Department of English at the University of Illinois Urbana-Champaign, underlines. In the US in the nineteenth century, newspaper material was explicitly not protected by copyright, and was routinely exchanged between titles:

Nineteenth-century editors’ attitude toward text reuse is exemplified in a selection that circulated in the last decade of the century, though often abbreviated from the version I cite here, which insists that “an editor’s selections from his contemporaries” are “quite often the best test of his editorial ability, and that the function of his scissors are not merely to fill up vacant spaces, but to reproduce the brightest and best thoughts…from all sources at the editor’s command.” While noting that sloppy or lazy selection will produce “a stupid issue,” this piece claims that just as often “the editor opens his exchanges, and finds a feast for eyes, heart and soul…that his space is inadequate to contain.” This piece ends by insisting “a newspaper’s real value is not the amount of original matter it contains, but the average quality of all the matter appearing in its columns whether original or selected.”

Material was not only copied verbatim, but modified and built upon in the process. As a result of this constant exchange, alteration and enhancement, newspaper readers in the US enjoyed a rich ecosystem of information, and a large number of titles flourished, since the cost of producing suitable material for each of them was shared and thus reduced.

That historical fact in itself is interesting. It’s also important at a time when newspaper publishers are some of the most aggressive in demanding ever stronger – and ever more disproportionate – copyright protection for their products, for example through “link taxes”. But Cordell’s speech is not simply backward looking. It goes on to make another fascinating observation, this time about large language models (LLMs):

We can see in the nineteenth-century newspaper exchanges a massive system for recycling and remediating culture. I do not wish to slip into hyperbole or anachronism, and will not claim historical newspapers as a precise analogue for twenty-first century AI or large language models. But it is striking how often metaphors drawn from earlier media appear in our attempts to understand and explain these new technologies.

The whole speech is well worth reading as a useful reminder that the current copyright panic over LLMs is in part because we have forgotten that sharing material and helping others to build on it was once the norm. And despite blinkered and selfish views to the contrary, it is still the right thing to do, just as parents continue to tell their children.

Source: Sharing material used to be the norm for newspapers, and should be for LLMs – Walled Culture