II. The Core Defense: Fair Use, or the Fiction of Learning

Virtually all legal defenses by LLM providers to their unauthorized use of copyrighted works for training originate from their invocation of the fair use doctrine.

We will reserve our deep-dive into the fair use defense for next time, but one thread of the argument goes like this: just as humans learn by reading and synthesizing knowledge, so too do AI models “learn” from the data they consume. If human learning isn’t infringement, the reasoning goes, neither should machine learning be. LLM developers contend that while a particular output may be informed by an original copyrighted work, it is also shaped by countless other sources, and any resemblance to one specific work is incidental—not the result of copying in the legal sense.

Whether courts will accept this analogy, however, remains very much in dispute—and depends in part on how they interpret the scale, precision, and retention power that distinguish LLMs from human minds.

III. From Surveillance to Substitution: Technology Changes Doctrine

Legal doctrine often lags behind technological disruption. The law’s assumptions—about what’s visible, what’s public, and what’s fair—can collapse when scale and precision transform the nature of harm. That’s exactly what happened with the Fourth Amendment in the era of government surveillance, and it’s what’s now unfolding in copyright law as AI systems convert public access into the great replacement—of human authorship, not humanity.

A. The evolution of Fourth Amendment doctrine with modern technological surveillance

For much of the 20th century, Fourth Amendment jurisprudence followed a rigid logic: what was exposed to public view was unprotected. If someone left their curtains open or spoke in a public place, they had no “reasonable expectation of privacy.” But as surveillance capabilities advanced, that premise began to break down.

The turning point came in Katz, where the Supreme Court held that a person using a public phone booth could still retain a reasonable expectation of privacy—not because the setting was private in the physical sense, but because the user’s intent to keep the conversation private deserved constitutional recognition.⁴ What mattered was not merely visibility, but how technology amplified the power to observe and exploit.

Subsequent cases—addressing GPS trackers, cell-site location data, and dragnet surveillance—have reaffirmed that the Fourth Amendment must evolve when new technologies allow public behavior to be captured, aggregated, and weaponized in ways never before possible.⁵

[cont’d ↗]

B. Copyright and AI: Why AI training is not analogous to human learning

That same logic now confronts copyright law. Courts, creators, and companies are grappling with whether the public availability of creative works online gives AI systems license to ingest, retain, and reproduce their expressive features—without consent or compensation.

It’s true that copyright regulates private conduct, not state actors. But the doctrinal echo is clear: both surveillance and AI training exploit a core assumption that public availability implies lawful use. In reality, scale and automation have changed the stakes. A human browsing a few blog posts is categorically different from an LLM trained on millions of blog posts to mimic their structure and style. And just as courts came to see that a thermal imaging scan or GPS tracker was not “just observation,” courts must now recognize that LLM training is not “just reading.”

Just because works are publicly posted, it does not mean they are in the public domain. Even without a copyright notice, original works fixed in a tangible medium are presumptively protected under U.S. law.⁶ The Copyright Act of 1976 eliminated formalities like mandatory notice or registration as a condition of protection.⁷ Yet the myth persists that if a work is online—and unmarked—it’s free for any use. That misunderstanding is now weaponized at industrial scale.

1. Machines don’t learn like humans

The fair use defense often hinges on the analogy between machine and human learning. But legally and functionally, the analogy breaks down.

Humans interpret. Machines retain. A person reading ten articles may be influenced by them, but they forget, misremember, and synthesize imperfectly. An LLM ingests entire corpora, transforms expressive language into vectorized representations, and reuses those patterns with near-perfect recall.

The concern isn’t that any given AI-generated output is a direct or substantially similar copy of one original work. In most cases, it isn’t. The deeper issue lies in the front-end: LLMs are trained through the wholesale ingestion of protected works—many in their entirety—without authorization.

The real concern is that AI models now generate content that fulfills the same purpose, echoes the same tone, and competes in the same market—replacing the original work and potentially displacing the author who created it.

2. The quantitative and qualitative gap we cannot ignore

What makes LLM training an entirely different animal isn’t just how it functions—it’s how much it consumes and how broadly it can deploy.

• Volume: LLMs are trained on hundreds of thousands, even millions, of copyrighted works—often in their entirety.
• Speed: A human might read a few books in a week. An AI model processes terabytes of content in hours.
• Scale of output: A person might publish a few articles a year. A deployed LLM can generate thousands of derivative works daily, many competing directly with human-authored content.

These differences are not incidental—they are foundational. Just as courts eventually recognized that digital surveillance changed the nature of privacy violations, they must now grapple with how generative AI changes the nature of unauthorized use.