EXTRA From the days of player pianos if not earlier, new technologies have often discomfited creators and copyright owners, at least when the technology is first introduced. Radio was initially seen as a mortal threat to songwriters and music publishers fearful of its effect on sheet music sales. The VCR, in the immortal words of then-Motion Picture Assn. of America CEO Jack Valenti, was to the Hollywood studios “like the Boston Strangler to a woman home alone.” The MP3 compression format was the bane of the recorded music business.
Yet, for all the angst (and litigation) that greeted them, each of those technologies represented simply new means of reproducing, distributing or performing existing creative works. As such, they fell squarely within the design and scope of copyright law, even as they threatened or disrupted existing business models, and rights owners eventually figured out how to turn them to their advantage: Collective licensing redeemed radio for songwriters and publishers; the VCR became the foundation of the highly profitable video rental business; the MP3 format became the basis of the paid streaming business and ultimately boosted the asset value of song catalogs.
The rise of advanced generative artificial intelligence tools like ChatGPT, Midjourney and Stable Diffusion, however, presents creators, copyright owners, and now the courts, with an entirely different question: Can this new technology be fitted into the current copyright framework at all for the purposes of licensing or regulation?
The U.S. Copyright Office has taken the position that works autonomously produced by machines (i.e. without identifiable human involvement) are not eligible for copyright protection and has refused to register such works as have been submitted. Given the scale at which AI systems are able to churn out new text, images, music and other types of media, defining all of that output as falling within the public domain by default could well create significant commercial and competitive problems for artists and publishers. But the legal question presented, which is now before a U.S. district court, is a fairly straightforward matter of statutory construction: Does the plain language of the Copyright Act unambiguously restrict copyright protection to works of genuine human authorship? And the Copyright Office has precedent on its side.
There is nothing straightforward about the other side of that coin, however: the use of existing copyrighted works to train AI systems to produce similar works.
The rub lies in which elements of the existing works an AI is actually using to produce (derive?) the new work, and whether anything like an actual copy of an existing work is ever present within the system beyond RAM. And would either operation offer a legal and technical foundation on which to erect a licensing system?
As a non-legal matter, the idea that creators whose works are being fed into the maw of an AI to train it to produce basically an unlimited number of potentially competing products should be compensated, has a certain intuitive appeal. Without it, it’s almost like asking soon-to-be laid off employees to train their lower-paid replacements. It just feels like a dick move.
Even as a legal matter, the case for compensation can seem straightforward enough. The AI is copying millions of copyrighted works, at least temporarily into RAM, in order to analyze them, and then producing new works based on what it has been fed. Both of those functions could seem to fall within the scope of copyright, and therefore require a license.
Or do they?
I asked musician, technologist and attorney Damien Riehl for a quick and dirty overview of the actual process, and its implications for licensing. Here’s what he non-artificially generated (lightly edited for space and clarity):
INPUTS → OUTPUT PROVENANCE — and Licensing:
Here’s how the technology works:
1. A large language model (LLM) ingests all the things (millions? billions? trillions?)
- Writing (text)
- Images
- Music
2. It puts all the components of all the things (e.g., words, sentences, pixels, pitches, rhythms) into vector space — 12,288 dimensions — creating a mathematical model of “how things work.”
- How words work: This word is usually followed by that word
- How images work: This pixel is usually next to that pixel
- How music works: This note is usually followed by that note
- Etc.
3. These are ideas, mathematical probabilities. Ideas are not copyrightable, only particular original expressions of those ideas. That is the fundamental idea/expression dichotomy in copyright law. Thus:
- Mathematics = uncopyrightable
- Statistics = uncopyrightable
- Facts (these words are statistically likely to appear together) = uncopyrightable
4. LLM systems do not store the things themselves (writing, images, music). The things themselves are jettisoned.
- Instead, the LLM stores a mathematical model of the things
- That mathematical model is the idea of things, not the expression of things
- Ideas cannot be copyrighted
5. LLMs, therefore, never hold any provenance of which particular inputs (the things) have contributed to the particular output (text, images, music)
6. So, one can never find this chain: Author → Source → LLM model (which jettisons references to billions of sources) → LLM output
7. Without such provenance, how can one do licensing? All that’s stored is the mathematical model, the ideas. Not the millions/billions/trillions of particular sources.
In addition to his legal work, Riehl is a co-founder of All The Music, a project using an algorithm to generate, by brute computational force, every mathematically possible melody within the diatonic and chromatic scales by varying pitch, rhythm, length and other parameters. So he may be predisposed to view the problem through the prism of mathematics. ATM’s goal is to demonstrate that what we call melody is simply a function of math — that is, facts, and a finite number of them at that — and therefore, by itself (i.e. absent lyrics or other musical elements) should perhaps not be considered original within the meaning of the Copyright Act or the basis of an infringement claim.
But if Riehl’s analysis is correct — and I have neither the legal nor technical chops to dispute it though it jibes with what other who have those credentials have said — and the courts and the Copyright Office stick to their position that works created by a computer cannot be copyrighted, then we’re in very murky waters indeed.
Taken together, they would seem to place the entire operation of generative AI systems — from the ingestion of existing works to constructing its mathematical model to the output — outside the realm of copyright.
To be sure, there could be other grounds on which to challenge the use of copyrighted works in AI training sets. The lawsuit filed Friday by Getty Images against Stability AI over its use of Getty-owned images in training Stable Diffusion, for instance, charges the AI developer with trademark infringement, unfair competition, trademark dilution and deceptive trade practices, in addition to copyright infringement.
AI systems that advertise their ability to create works “in the style of” named artists, authors or musicians, could also be vulnerable to charges of misuse of their name, image or likeness (NIL).
But those laws are more likely to provide rights owners with individual causes of action than with the basis for a comprehensive and systematic licensing regime.
For that, it may require Congress to generate new rules.