Fair use for me, EULAs for thee: the AI industry's anti-distillation grift

Watching OpenAI whine about open-weight competitors scraping its data is a masterpiece of cognitive dissonance. I was reading a recent piece from Stratechery analyzing the frontier labs' absolute panic over foreign and open-source competitors distilling their models, and the hypocrisy is staggering. Over on Hacker News, the technical community immediately identified the grift, with users pointing out that while the labs insist scraping our copyrighted code is protected under fair use, they somehow view the distillation of their own uncopyrightable outputs as a criminal offense.

Structurally, these companies behave exactly like the tropical strangler fig. The fig germinates high in the canopy of a host tree, dropping parasitic roots that slowly envelop the trunk, extracting every drop of sunlight and water until the original host dies and rots away. Then, having become the new forest canopy, the strangler fig secretes toxic allelopathic chemicals into the soil to ensure that no other seed can ever sprout in its shadow. The AI industry strip-mined human culture to build their computational empires, and now they are desperately deploying legal toxins to pull up the ladder.

The corporate euphemism for this toxin is a ban on model distillation (the industry term for forbidding the open-source community from feeding a monopolist's expensive outputs into a smaller, cheaper model to teach it how to behave). Strip away the sci-fi mystique, and distillation is just the 2020s equivalent of 1990s clean-room reverse engineering. It is a foundational act of adversarial interoperability—a blue-collar hacker tradition of figuring out how a monopolist's black box works and building a cheaper, better version to break their stranglehold. When the frontier labs frame distillation as a violation of their property rights, they are prosecuting a completely made-up crime. It is the textbook definition of Felony Contempt of Business Model. Structurally, a distillation ban is merely a class war against the tinkerer, designed to artificially inflate switching costs and punish anyone who refuses to buy their overpriced API credits forever.

To understand the depth of the grift, you have to compare the rules they demand for their inputs against the rules they enforce for their outputs. When building their multi-billion-dollar empires, Western AI labs ingested trillions of copyrighted works. They correctly argued that this ingestion was protected by fair use, insisting that parsing data to extract mathematical patterns does not infringe on the underlying copyright. But look at the glaring structural hypocrisy when those same models generate text or code. As the US Copyright Office definitively ruled in the Federal Register, AI outputs are entirely devoid of human authorship. They cannot be copyrighted. They belong to the public domain.

So, if the outputs are legally recognized as uncopyrightable, how exactly do OpenAI and Anthropic ban competitors from distilling them?

They weaponize contract law to override federal copyright jurisprudence, inventing a pseudo-copyright out of thin air using restrictive end-user agreements to lock up the public domain. Knowing they would unequivocally lose an intellectual property battle, these firms bury anti-compete mandates in their boilerplate to ensure no one can train a rival architecture using their machine-barf:

https://openai.com/policies/terms-of-use

This is an industry-wide cartel behavior, mirrored directly in Anthropic's Commercial Terms. This legal fiction revives a toxic precedent from the 1990s, specifically Bowers v. Baystate Technologies (a case where a federal court decided that a flimsy shrink-wrap contract was magically more powerful than the US copyright code). In that era, software monopolies pioneered the tactic of using impenetrable volumes of legalese to ban reverse engineering, successfully proving that end-user licenses could obliterate federal exemptions. It was profoundly deranged (and completely democratically bankrupt), because it meant that Congress’s laws were essentially subordinated to whatever rapacious wish-list a corporate lawyer could cram into an unreadable EULA.

Distillation bans are just the latest iteration of that exact mechanism. By leveraging end-user license agreements to artificially restrict the downstream processing of legally unencumbered public-domain data, these corporations are executing an algorithmic monopsony. Speaking in the dry vernacular of antitrust scholars, they are deploying their massive market dominance to establish a regime of private legislation. They are substituting the democratic process of copyright reform with a unilateral fiat authored by their own highly paid corporate attorneys. This systematically criminalizes the very technical processes that historically allowed new entrants to discipline entrenched market actors. It is a textbook attempt to outlaw the commoditization of their complements. The frontier labs realize their staggering capital expenditures are failing to yield a defensible competitive moat, especially in the face of leaner, highly capable open-source models. Unable to compete on raw utility, they are desperately throwing up artificial legal barricades to ensure they remain the sole rentiers of the entire artificial intelligence ecosystem.

This is a staggeringly filthy maneuver.

We cannot trust these firms to voluntarily discover a moral compass and surrender their market power. Entrenched tech giants will never willingly surrender their chokeholds. Shifting from observation to militant demand means recognizing that click-wrap agreements must never be permitted to supersede federal copyright law. We must legally invalidate these anti-competitive Terms of Service. We need the Federal Trade Commission to step in immediately and explicitly ban anti-distillation clauses as an unfair method of competition under Section 5 of the FTC Act. We need Lina Khan's FTC to loudly declare that you cannot use a boilerplate contract to privatize the public domain.

The sheer audacity required to scrape the entirety of human creative endeavor without compensating a single author, only to turn around and weaponize private contracts to prosecute anyone who dares to parse the uncopyrightable weights of your resulting algorithm, is a breathtaking display of monopoly hubris. They are claiming a unilateral right to enclose the digital commons while fiercely defending their right to strip-mine ours.

We refuse.

If regulators allow click-wrap agreements to permanently erase federal copyright exemptions, the strangler fig will complete its total enclosure of the digital forest, leaving nothing for the tinkerers, the builders, or the public. If training a trillion-parameter model on the accumulated, copyrighted labor of every human artist and writer is 'fair use,' then distilling that uncopyrightable machine-barf into an open-source competitor is fair use, too. We don't need to ask for their permission: we need to render their anti-competitive garbage-novellas null and void.