Watching the recent decisions around training large language models and the intersection with copyright, a balancing of interests is clearly at play. Large language models need copious amounts of high-quality material to produce content that is accurate and therefore useful. Creators also need protection for their work, which is the role IP has long served. Failure to protect IP in the face of AI would likely lead to fewer reliable high-quality materials to train LLMs. And, if copyright prohibits training LLMs using copyrighted material, then we will not be able to rely on LLM output any more than we could rely on a biography written without review of prior source references.  

For as long as thought pieces and other works of authorship have been informed by reading prior references, there has been no foul so long as the informed work of authorship is new (not copied). It seems the game is not so different with AI at bat. The key (or at least one of them) is still whether the new work is transformative.  

The Third Circuit's ruling in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., found infringement of Westlaw headnotes where Ross's use was “minimally transformative, at best." Ross's system could only point users to existing opinions, unlike LLMs, which can generate original expression and therefore create highly transformative output. 

The U.S. Government's Statement of Interest in The New York Times Company v. Microsoft Corporation et al., calls LLM training “extraordinarily transformative,” and asks courts to analyze training separately from outputs. This aligns with Judge Alsup's reasoning in the Bartz et al. v. Anthropic PBC, finding that the training there was highly transformative and comparing the model to a person who reads widely to learn how to write. 

Of course, there are many more complications around training LLMs with copyrighted materials, including (as in Anthropic), whether the source material was lawfully acquired. But at a high level, it seems that whether the learning is being done by a human or AI, the protection in place to maintain the delicate balance between informing new works and protecting the old ones is the same: where the source references were not pirated, and the output was not copied, the use is probably fair.