Why AI shouldn’t get special copyright privileges humans don’t have

The paper probes deeply into the mechanics of how GenAI models are trained - by copying trillions of tokens scraped from the internet, much of it copyright-protected - and asserts that these acts of reproduction are clear-cut cases of infringement unless protected by fair use.

A new academic study challenges the notion that generative artificial intelligence (GenAI) should get special exemptions from copyright law, arguing that such privileges would be inconsistent with existing legal protections for human creators. In the working paper titled “Unfair Learning: GenAI Exceptionalism and Copyright Law”, submitted on the preprint server arXiv, David Atkinson, Assistant Professor at the University of Texas at Austin, critically examines legal and policy justifications for allowing GenAI to use copyrighted material without authorization.

At its core, the study challenges a critical assumption: Why grant GenAI legal exceptions denied to humans? The author argues that every legal and substantive rationale advanced in favor of GenAI also applies to human creators - often even more so. To permit blanket fair use for GenAI but deny it to individuals, he suggests, would be both illogical and legally unsound.

The paper probes deeply into the mechanics of how GenAI models are trained - by copying trillions of tokens scraped from the internet, much of it copyright-protected - and asserts that these acts of reproduction are clear-cut cases of infringement unless protected by fair use. Fair use, a legal doctrine grounded in four statutory factors, has become the cornerstone defense for GenAI companies facing a wave of litigation, including a high-profile suit by The New York Times against OpenAI and Microsoft.

Yet the author challenges whether this reliance on fair use withstands scrutiny, especially when applied at scale. The study evaluates the traditional four-factor fair use test as well as a suite of eight common policy arguments used to justify GenAI’s access to copyrighted material without permission.

Another key question addressed is whether GenAI’s ways of learning and transformation are fundamentally different from those of humans. Atkinson’s answer: not meaningfully. He illustrates that humans also learn by copying, through reading, listening, and observing, and transform inputs into novel outputs. However, unlike GenAI systems, humans are limited by cognitive and temporal constraints and cannot regurgitate massive volumes of text verbatim.

By contrast, GenAI systems can replicate long stretches of copyrighted text, often word-for-word, and at scales that defy human capacity. Despite this disparity, Atkinson argues, if courts accept that such use by GenAI qualifies as fair use, it would logically follow that humans should be allowed to do the same. That, in turn, would collapse the copyright marketplace entirely, as individuals could download and consume copyrighted books, films, and music without paying for them - so long as they didn’t produce identical outputs.

Another foundational issue addressed is whether the supposed "transformativeness" of GenAI outputs, central to the first fair use factor, should shield companies from liability. The author argues that courts have historically limited fair use to narrow, purpose-driven uses such as parody, criticism, or functional repurposing (as in search engines or snippet libraries). GenAI, he points out, is not serving such a targeted, functional purpose. It is general-purpose and expressive by design - intended to generate outputs “in the style of” copyrighted creators, a use that courts have repeatedly rejected as fair.

The study also addresses the question of intermediary copying, whether reproducing copyrighted content during the training phase should be considered infringing even if outputs are non-infringing. GenAI defenders claim that intermediate copying is a technological necessity. The author acknowledges the long-standing legal recognition of such copies in cases involving software interoperability, but contends that these precedents do not apply to GenAI, where the intent is often expressive replication rather than functional transformation.

Another critical knowledge question raised is whether “public availability” of content equates to permission for its use in GenAI training. Atkinson emphatically rejects this logic, noting that publicly available is not synonymous with public domain. He warns that adopting such a view would obliterate the economic incentives underpinning content creation online, particularly in journalism, education, and independent media.

The paper dismantles the frequent claim that GenAI deserves a relaxed regulatory environment to foster innovation. Atkinson counters that real innovation has always come from humans, and that there is no historical precedent for granting machines greater legal privileges than the people who create and use them. He critiques the assumption that GenAI is inherently progress-enhancing, pointing to early research showing its tendency to reproduce clichés and surface-level insights rather than original ideas.

Simply put, the study asserts that the law should not afford GenAI companies legal leeway that is denied to humans. If society is serious about preserving copyright's role in promoting the arts and sciences, then courts and policymakers must resist arguments that treat GenAI as an exceptional class of user. Instead, the author suggests, the only defensible policy choice is to hold GenAI and humans to the same legal standard.