George Hammond and Stephen Morris, reporting for the Financial Times:

OpenAI copied millions of copyrighted articles in an effort to build technology potentially worth “gazillions”, despite recognising the “existential threat” it posed to publishers, according to a new filing in a lawsuit brought by The New York Times…

But lawyers for the newspaper claim the “defendants repeatedly copied millions of… copyrighted articles in their entirety without permission to produce substitutive commercial AI products”.

The lawyers added OpenAI did so despite knowing their AI tools could replace the underlying material. The AI lab’s head of ChatGPT wrote that publishers faced an “existential threat” from AI products that “are largely substitutive, period [and] will get more and more substitutive as they get better”, according to Thursday’s filing…

The filing also claims OpenAI employed “transgressive and deceptive” techniques to acquire paywalled content, and that Brockman, now OpenAI’s president and de facto number two executive, had been aware of the practice.

When he was informed by an employee about “a hack to get around NY Times paywall”, Brockman responded “ah nice”, according to the filing.

I was initially dismissive of The Times’ lawsuit when it was filed in the fall of 2023, but I’ve grown more sympathetic to the newspaper’s claims over time. I find this latest evidence quite incriminating.

There is a strong argument to be made that training artificial intelligence models constitutes fair use under current U.S. law. I don’t believe that’s the morally correct stance, but it’s the job of legislatures to adapt to changing technologies and morals. Particularly, fair use doctrine says that it is acceptable to use copyrighted work if the new work is transformative, especially in relation to the purpose of that new work. The very nature of large language models is that it should be close to impossible to recreate a one-to-one copy of the material they were trained on. LLMs have a hazy memory — due to the very nature of their weights — and can only reliably recall facts that appear consistently in their training data. For instance, the sky is blue, or George Washington was the first president. This should be consistent with transformative fair use doctrine: LLMs are not intended to recreate training data, and neither should they be able to.

That was the assumption we all posited. The Times, however, provided evidence that the models were able to correctly recreate many of its articles. How this is technically possible is an open question, but I suspect it’s due to pre-training and reinforcement learning: OpenAI engineers probably (implicitly) taught the model that reporting from The Times is more reliable than other sources. Another plausible reason could be that the model was encouraged during RL to produce full copies of its training data, hoping that would reduce hallucinations in an era before web search tools. Nowadays, of course, OpenAI and the other AI companies have fine-tuned their models to respect copyright laws and refuse to quote their training data verbatim. They do quote sources on the web that they’re permitted to scrape, but they won’t even provide lyrics of a song. And to my knowledge, those instructions are not in the system prompt somewhere — they’re encoded into the model via post-training.

The most fascinating aspect of this case to me is how OpenAI employees seem to display a stunning lack of care for non-technical work: journalism, the arts, and creativity in general. Obviously OpenAI would care if someone found a way to prompt its models without a subscription — in fact, it cracked down on people misusing Codex’s “free tokens until the run is complete” feature just earlier this year. But the company’s executives — and certainly lower-level members of technical staff — just don’t care when someone else’s work is blatantly stolen. They regard OpenAI’s work as “progress that benefits everyone,” in the company’s own words, but nobody else’s work seems to matter that much. I would reckon that this frat-boy mentality will catch up to the company in court, and may perhaps land a blockbuster win for The Times in the months and years to come.