quartz/content/Essays/ai-infringement.md

title	tags	date	draft	lastmod
Generative AI: Copyright Infringement's New Trench Coat	essay ai legal copyright	2023-11-04	true	2024-07-25

One ticket to the original, authorized, or in the alternative, properly licensed audiovisual work, please!

A film roll clatters to the ground from underneath a suspiciously camera-shaped bulge in the figure's oversized trench coat.

[!info] I’m looking for input! Critique my points and make your own arguments. That’s what the comments section is for.

Warning

CW: US law and politics; memes

This site contains my own opinion in a personal capacity, and is not legal advice, nor is it representative of anyone else's opinion. Not every citation is an endorsement, and none of the authors I cite have endorsed this work.

• Also a reminder that I won’t permit inputting my work in whole or part into an LLM.

I've seen many news articles and opinion pieces recently that support training generative AI and LLMs (such as ChatGPT/GPT-4, LLaMa, and Midjourney) on the broader internet as well as more traditional copyrighted works, without respect to the copyright holders for all of the above. For now, this will be less of a response to any one article and more of a collection of points of consideration that tie together common threads in public perception. I intend for this to become comprehensive over time.

My opinion here boils down to three main points. Under existing US law:

• Training a generative AI model on copyrightable subject matter without authorization is copyright infringement (and the proprietors of the model should be responsible);
• Generating something based on copyrightable subject matter is copyright infringement (and the proprietors and users of the model should be jointly responsible); and
• Fair use is not a defense to either of the above.

I also discuss policy later in the essay. Certain policy points are instead made in my Essays/plagiarism, and links to that entry will be labeled with 🅿️.

Prologue: why these arguments are popping up

In short, there's a growing sentiment against copyright in general. Copyright can enable centralization of rights when paired with a capitalist economy, which is what we've been historically experiencing with the advent of copyright repositories like record labels and publishing companies. It's even statutorily enshrined as the "work-for-hire" doctrine. AI has the potential to be an end-run around these massive corporations' rights, which many see as a benefit.

However, this argument forgets that intangible rights are not yet so centralized that independent rights-holders have ceased to exist. While AI will indeed affect central rights-holders, it will also harm individual creators and the bargaining power of those that choose to work with central institutions. Instead, I see AI as a neutral factor to the disestablishment of copyright. Due to my roots in the indie music and open-source communities, I'd much rather keep their/our/your present rights intact.

Reconciling the two views, I'm sympathetic to arguments against specific parts of the US's copyright regime as enforced by the courts, such as the DMCA or the statutory language of fair use. We as a voting population have the power to compel our representatives to enact reforms that take the threat of ultimate centralization into account, and can even work to break down what's already here. But I don't think that AI should be the impetus for arguments against the system as a whole.

The Legal Argument

Fair warning, this section is going to be the most law-heavy, and probably pretty tech-heavy too. Feel free to skip #The First Amendment and the "Right to Read"

The field is notoriously paywalled, but I'll try to link to publicly available versions of my sources whenever possible. Please don't criticize my sources in this section unless a case has been overruled or a statute has been repealed/amended (i.e., I can't rely on it). This is my interpretation of what's here (again, not legal advice or a professional opinion. Seek legal counsel before acting/refraining from action re: AI). Whether a case is binding on you personally doesn't weigh in on whether its holding is the nationally accepted view.

The core tenet of copyright is that it protects original expression, which the Constitution authorizes regulation of as "works of authorship." This means you can't copyright facts. It also results in two logical ends of the spectrum of arguments made by authors (seeking protection) and defendants (arguing that enforcement is unnecessary in their case). For example, you can't be sued for using the formula you read in a math textbook, but if you scan that math textbook into a PDF, you might be found liable for infringement because your reproduction contains the way the author wrote and arranged the words and formulas on the page.

One common legal argument against training as infringement is that the AI extracts facts, not the author's expression, from a work. But that position assumes that the AI is capable of first differentiating the two, and then separating them in a way analogous to the human mind's.

Training

Everything AI starts with a dataset. And most AI models will start with the easiest, most freely available resource: the internet. Hundreds of different scrapers exist with the goal of collecting as much of the internet as possible to train modern AI (or previously, machine learners, neural networks, or even just classifiers/cluster models). I think that just acquiring data without authorization to train an AI on it is copyright infringement standing by itself.

Acquiring data for training is an unethical mess. In human terms, scrapers like Common Crawl will take what they want, without asking (unless you know the magic word to make it go away, or just Projects/Obsidian/digital-garden#Block the bot traffic!), and without providing immediately useful services in return like a search engine. For more information on the ethics of AI datasets, read my take on Essays/plagiarism#AI shouldn't disregard the need for attribution, and have a look at the work of Dr. Damien Williams (Mastodon).

The first reason that it's copyright infringement? MAI Systems v. Peak Computer. It holds that RAM copying (ie, moving a file from somewhere to a computer's memory) is an unlicensed copy. As of today, it's still good law, for some reason. Every single file you open in Word or a PDF reader; or any webpage in your browser, is moved to your memory before it gets displayed on the screen. Bring it up at trivia night: just using your computer is copyright infringement! It's silly and needs to be overruled going forward, but it's what we have right now. And it means that a bot drinking from the firehose is committing infringement on a massive scale.

But then a company actually has to train an AI on that data. What copyright issues does that entail? First, let's talk about The Chinese Room.

The Chinese Room is a philosophical exercise authored by John Searle where the (in context, American) subject is locked in a room and receives symbols in Chinese slipped under the door. A computer program tells the subject what Chinese outputs to send back out under the door based on patterns and combinations of the input. The subject does not understand Chinese. Yet to an observer of Searle's room, it appears as if whoever is inside it has a firm understanding of the language.

Searle's exercise was at the time an extension of the Turing test. He designed it to refute the theory of "Strong AI." At the time that theory was well-named, but today the AI it was talking about is not even considered AI by most. The hypothetical Strong AI was a computer program capable of understanding its inputs and outputs, and importantly why it took each action to solve a problem, with the ability to apply that understanding to new problems (much like our modern conception of Artificial General Intelligence). A Weak AI, on the other hand, was just the Chinese Room: taking inputs and producing outputs among defined rules. Searle reasoned that the "understanding" of a Strong AI was inherently biological, thus one could not presently exist.

• Note that some computer science sources like IBM have taken to using Strong AI to denote only AGI, which was a sufficient, not necessary quality of a philosophical "intelligent" intelligence like the kind Searle contemplated.

Generative AI models from different sources are architected in a variety of different ways, but they all boil down to one abstract process: tuning an absurdly massive number of parameters to the exact values that produce the most desirable output. (note: CGP Grey's video on AI and its follow-up are mainly directed towards neural networks, but do apply to LLMs, and do a great job illustrating this). This process requires a gargantuan stream of data to use to calibrate those parameters and then test the model. How it parses that incoming data suggests that, even if the method of acquisition is disregarded, the AI model still infringes the input.

The Actual Tech

At the risk of bleeding the #Generation section into this one, generative AI is effectively a very sophisticated next-word predictor based on the words it has read and written previously.

First, this training is deterministic. It's a pure, one-way, data-to-model transformation (one part of the process for which "transformer models" are named). The words are ingested and converted into one of various types of formal representations to comprise the model. It's important to remember that given a specific work and a step of the training process, it's always possible to calculate by hand the resulting state of the model after training on that work. The "black box" that's often discussed in connection with AI refers to the final state of the model, when it's no longer possible to tell what effect of certain portions of the training data have had on the model.

If some words are more frequently associated together, then that association is more "correct" to generate in a given scenario than other options. And the only data to determine whether an association is correct would be that training input. This means that an AI trains only on the words as they are on the page. Training doesn't have some external indicator of semantics that a secondary natural-language processor on the generation side can incorporate. Training thus can't be analogized to human learning processes, because when an AI trains by "reading" something, it isn't reading for the forest—it's reading for the trees. Idea and expression are meaningless distinctions to AI.

As such, modern generative AI, like the statistical data models and machine learners before it, is a Weak AI. And weak AIs use weak AI data. Here's how that translates to copyright.

• Sidebar: this point doesn't consider an AI's ability to summarize a work since the section focuses on how the training inputs are used rather than how the output is generated from real input. This is why I didn't want to get into generation in this section. It's confusing, but training and generation are merely linked concepts rather than direct results of each other when talking about machine learning. Especially when you introduce concepts like "temperature", which is a degree of randomness added to a model's (already variant) choices in response to an user in order to simulate creativity.
• ...I'll talk about that in the next section.

"The Law Part"

All of the content of this section has been to establish how an AI receives data so that I can reason about how it stores that data. In copyright, reproduction, derivatives or compilations of works without authorization can constitute infringement. I believe that inputting a work into a generative AI creates a derivative representation of the work. Eventually, the model is effectively a compilation of all works passed in. And finally (on a related topic), there is nothing copyrightable in how it's arranged the works in that compilation even if every work trained on is authorized.

• Sidebar: fair use analysis for both training and generation is located in the #Fair Use section.

Recall that training on a work incorporates its facts and the way the author expressed those facts into the model. When the training process takes a model and extracts weights on the words within, it's first reproducing copyrightable expression, and then creating something directly from the expression. You can analogize the model at this point to a translation (a specifically recognized type of derivative) into a language the AI can understand. But where a normal translation would be copyrightable (if authorized) because the human translating a work has to make expressive choices and no two translations are exactly equal, an AI's model would not be. A given AI will always produce the same translation for a work it's been given, it's not a creative process. Even if every work trained on expressly authorized training, I don't think the resulting AI model would be copyrightable. And absent authorization, it's infringement.

As the AI training scales and amasses even more works, it starts to look like a compilation, another type of derivative work. Normally, the expressive component of an authorized compilation is in the arrangement of the works. Here, the specific process of arrangement is predetermined and encompasses only uncopyrightable material. I wasn't able to find precedent on whether a deterministically-assembled compilation of uncopyrightable derivatives passes the bar for protection, but that just doesn't sound good. Maybe there's some creativity in the process of creating the algorithms for layering the model (related: is code art?). More in the #Policy section.

More cynically, I don't think any of this could be workable in a brief. Looking at how much technical setup I needed to make this argument, there's no way I could compress this all into something a judge could read (even ignoring court rule word limits) or that I could orate concisely to a jury. I'm open to suggestions on a more digestible way to go about arguing the principles I'm concerned about based on this technological understanding of AI.

Detour: point for the observant

The idea and expression being indistinguishable by AI may make one immediately think of merger doctrine. That argument looks like: the idea inherent in the work trained on merges with its expression, so it is not copyrightable. That would not be a correct reading of the doctrine. Ets-Hokin v. Skyy Spirits, Inc. makes it clear that the doctrine is more about disregarding the types of works that are low-expressivity by default, and that this "merger" is just a nice name to remember the actual test by. Confusing name, easy doctrine.

Generation

The model itself is only one side of the legal AI coin. What of the output? It's certainly not copyrightable. The US is extremely strict when it comes to the human authorship requirement for protection. If an AI is seen as the creator, the requirement is obviously not satisfied. And the human "pushing the button" probably isn't enough either. But does it infringe the training data? It depends.

Human Authorship

As an initial matter, AI-generated works do not satisfy the human authorship requirement. This makes them uncopyrightable, but more importantly, it also gives legal weight to the distinction between the human and AI learning process. Like I mentioned in the training section, it's very difficult to keep discussions of training and generation separate because they're related concepts, and this argument is a perfect example of that challenge.

Summaries

This section is the most direct refutation of the "AI understands what it trains on" conclusion. I also think it's the most important aspect of generative models for me to discuss. The question: If an AI can't understand what it reads, how does it choose what parts of a work should be included in a summary of that work? A book, an article, an email?

Once again, the answer is mere probability. In training, the model is told what word to come after a word is more "correct" by how many times that sequence of words occurs in its training data. And in generation, if more of the work mentions a particular subject than the actual conclusion of the work, the subject given most attention will be what the model includes in a summary.

Empirical evidence of this fact can be found in the excellent post, When ChatGPT Summarizes, it Actually does Nothing of the Kind. It's funny how this single approach is responsible for nearly all of the problems with generative AI, from the decidedly unartistic way it "creates" to its Essays/plagiarism##1 Revealing what's behind the curtain. I don't want this sort of technology to take any place in daily life.

Dr. Edgecase, or how I learned to stop worrying (about AI) and love the gig worker

So how do corporations try to solve the problem? Human-performed microtasks.

AI can get things wrong, that's not new. Take a look at this:

!limmygpt.png Slight variance in semantics, same answer because it's the most popular string of words to respond to that pattern of a prompt. Again, nothing new. Yet GPT-4 will get it right. This probably isn't due to an advancement in the model. My theory is that OpenAI looks at the failures published on the internet (sites like ShareGPT, Twitter, etc) and has remote validation gig workers (already a staple in AI) "correct" the model's responses to that sort of query. In effect, corporations are exploiting (yes, exploiting) developing countries to create a massive network of edge cases to fix the actual model's plausible-sounding-yet-wrong responses. So that begs the question: who's responsible for the expressive, copyrightable content of these edge cases?

Expression and Infringement; "The law part" again

Like training, generation also involves reproduction of But where a deterministic process creates training's legal issues, generation is problematic for its non-deterministic output.

It can be said that anything a human produces is just a recombination of everything that person's ever read. Similarly, that process is a simplified understanding of how an AI trains.

==MORE==

Detour: actual harm caused by specific uses of AI models

My bet for a strong factor when courts start applying fair use tests to AI output: harm. { and I actually wrote this before the Essays/no-ai-fraud-act 's negligible-harm provision was published, -ed. } Here's a quick list of uses that probably do cause harm, some of them maybe even harmful per se (definitely harmful without even looking at specific facts).

• Election fraud and misleading voters, including even more corporate influence on US elections (not hypothetical in the slightest, and knowingly unethical)
  • Claiming misleading voters?
• Other fraud, like telemarketing/robocalls, phishing, etc
• Competition with actual artists and authors (I am VERY excited to see where trademark law evolves around trademarking one's art or literary style. Currently, the arguments are weak and listed in the mini-argument section).
• Obsoletes human online workforces in tech support, translation, etc
• Essays/plagiarism##1 Revealing what's behind the curtain
• Violates the GDPR on a technological level
  • I also think being unable to delete personal data that it has acquired and not just hallucinated is a big problem

Detour 2: An Alternative Argument

There's a much more concise argument that generative AI output infringes on its training dataset. I don't plan to engage with it much because I can only see it being used to sue a user of a generative AI model, not the corporation that created it.

Recall that AI output taken right from the model (straight from the horse's mouth) is not copyrightable according to USCO. If the model's input is copyrighted, and the output can't be copyrighted, then there's nothing in the AI "black box" that adds to the final product, so it's literally just the training data reproduced and recombined. Et voila, infringement.

This isn't to say that anything uncopyrightable will infringe something else, but it does mean that the defendant's likelihood of prevailing on a fair use defense could be minimal. Additionally, the simpler argument makes damages infinitely harder to prove in terms of apportionment.

Note that there are many conclusions in the USCO guidance, so you should definitely read the whole thing if you're looking for a complete understanding of the (very scarce) actual legal coverage of AI issues so far.

Where do we go from here?

Well, getting to evaluation of the above by courts would be a start. Right now, courts are ducking AI issues left and right on standing and pleading grounds. Once there's more solid (or honestly any) coverage of the legal arguments on the merits, whether the law should be enforced will become prudent.

Policy

These arguments will be more or less persuasive to different people. I think there's a lot more room for discussion here because they become relevant to the future direction of the law as well as current enforcement. The most important debate is up first, but the others are not particularly ordered.

[!info] Section Under Construction More topics under this section forthcoming! I work and edit in an alternate document and copy over sections as I finish them.

Fair Use

WIP

Who's holding the bag?

WIP https://www.wsj.com/tech/ai/the-ai-industry-is-steaming-toward-a-legal-iceberg-5d9a6ac1?st=5rjze6ic54rocro&reflink=desktopwebshare_permalink

Detour: Section 230 (again)

Well, here it is once more. There's strangely an inverse relationship between fair use and § 230 immunity. If the content by an AI is not just the user's content and is in fact transformative, then it's the website's content, not user content. That would strip Section 230 immunity from the effects of whatever the AI says. Someone makes an investment decision based on the recommendation of ChatGPT? Maybe it's financial advice. I won't bother with engaging the effects further here. I have written about § 230 and AI no-ai-fraud-act#00230: Incentive to Kill, albeit in reference to AI-generated user content hosted by the platform.

The First Amendment and the "Right to Read"

This argument favors allowing GAI to train on the entire corpus of the internet, copyright- and attribution-free, and bootstraps GAI output into being lawful as well. The position most commonly taken is that the First Amendment protects a citizen's right to information, and that there should be an analogous right for generative AI.

The right to read, at least in spirit, is still being enforced today. Even the 5th Circuit (!!!) believes that this particular flavor of First Amendment claim will be likely to succeed on appeal after prevailing at the trial level. Book People v. Wong, No. 23-50668 (5th Cir. 2024) (not an AI case). It also incorporates principles from intellectual property law. Notably, that you can read the content of a work without diminishing the value of the author's expression (i.e. ideas aren't copyrightable). As such, the output of an AI is not taking anything from an author that a human wouldn't take when writing something based on their knowledge.

I take issue with the argument on two points that stem from the same technological foundation.

First, as a policy point, the argument incorrectly humanizes current generative AI. There are no characteristics of current GAI that would warrant the analogy between a human reading a webpage and an AI training on that webpage.

Second and more technically, #Training above is my case for why an AI does not learn in the same way that a human does in the eyes of copyright law. ==more==

But for both of these points, I can see where the confusion comes from. The previous leap in machine learning was called "neural networks", which definitely evokes a feeling that it has something to do with the human brain. Even more so when the techniques from neural network learners are used extensively in transformer models (that's those absurd numbers of parameters mentioned earlier).

Mini-arguments

A list of smaller points that would cast doubt on the general zeitgeist around the AI boom that I found compelling. These may be someone else's undeveloped opinion, or it might be a point that I don't think I could contribute to in a valuable way. Many are spread across the fediverse; others are blog posts or articles. Others still would be better placed a Further Reading section, but I don't like to tack on more than one post-script-style heading. { ed.: #Further Reading }

• Cartoonist Dorothy’s emotional story re: midjourney and exploitation against author intent
• Misinformation worries
• Stronger over time
  • One of the lauded features of bleeding-edge AI is its increasingly perfect recall from a dataset. So you're saying that as AI gets more advanced, it'll be easier for it to exactly reproduce what it was trained on? Sounds like an even better case for copyright infringement.
• Inevitable harm
  • Temperature and the very fact that word generation is used mean that there's no way to completely eliminate hallucination, so truth in AI is unobtainable. Xu, et al.
• Unfair competition
  • This doctrine is a catch-all for claims that don't fit neatly into any of the IP categories, but where someone is still being wronged by a competitor. I see two potential arguments here.
    • First, you could make a case for the way data is scraped from the internet being so comprehensive that there's no way to compete with it by using more fair/ethical methods. This could allow a remedy that mandates AI be trained using some judicially devised (or hey, how about we get Congress involved if they don't like the judicial mechanism), ethical procedure. The arguments are weaker, but they could be persuasive to the right judge.
    • Second, AI work product is on balance massively cheaper than hiring humans, but has little other benefit, and causes many adverse effects. A pure cost advantage providing windfall for one company but not others could also be unfair. Again, it's very weak right now in my opinion.

Further Reading

• Copyleft advocate Cory Doctorow has written a piece on why copyright is the wrong vehicle to respond to AI. Reply-guying his technical facts and legal conclusions is left as an exercise for the reader; I articulated #Training#The Actual Tech #Generation in this write-up as comprehensively as I could so that readers can reference it to evaluate the conclusions of other works. What's more interesting is his take on the non-fair use parts of the #Policy debate. This entry holds my conclusions on why copyright can be enforced against AI; reasonable minds can and should differ on whether it ought to be.
• TechDirt has a great article that highlights the history of and special concerns around fair use. I do think that it's possible to regulate AI via copyright without implicating these issues, however. And note that I don't believe that AI training is fair use, for the many reasons above.