The Clean Room Doesn't Exist Anymore

On March 31, 2026, Anthropic accidentally shipped 512,000 lines of Claude Code's production TypeScript inside an npm package. Within hours, a developer named Sigrid Jin used OpenAI's Codex to rewrite the core architecture from scratch in Python: the CLI, query engine, tool execution, session management. Not the full codebase, but the structural skeleton. The project, called claw-code, hit 50,000 GitHub stars in two hours. Anthropic filed DMCA takedowns against direct forks of the original TypeScript. They didn't touch the rewrite.

Most coverage asked whether source code can still be owned. The deeper question is about the legal mechanism Jin invoked, one of the oldest doctrines in intellectual property law, and whether it means anything when a machine does the rewriting in a single night.

What "Clean Room" Actually Means

In 1982, a company called Phoenix Technologies wanted to build a compatible clone of IBM's personal computer. IBM's BIOS, the fundamental firmware that made a PC a PC, was copyrighted. Phoenix couldn't copy it. So they did something clever.

They split their engineers into two teams. The first team studied IBM's BIOS, disassembled the code, and wrote a detailed functional specification describing what the BIOS did, not how it was implemented. The second team never saw IBM's code. They worked from nothing but the specification, building a new BIOS that behaved identically but shared no code with the original. A "clean room," separated by a wall that no copyrighted expression could cross.

IBM sued. IBM lost. The court held that copyright protects expression, not function. The ideas, algorithms, and functional behavior of the BIOS were unprotectable. Only the specific code implementing them belonged to IBM. Phoenix's second team had never seen that code, so their implementation (though functionally identical) was independent creation.

This doctrine became the foundation of the compatible software industry. In Sega v. Accolade (1992), the Ninth Circuit held that Accolade's disassembly of Sega's game cartridge code was fair use, because the purpose was to discover unprotectable functional elements. In Sony v. Connectix (2000), the same court extended that reasoning: Connectix's reverse engineering of Sony's PlayStation BIOS to build a software emulator was fair use, even though intermediate copies were made. The Supreme Court reinforced the principle in Google LLC v. Oracle America (2021), ruling 6-2 that Google's copying of Java API declarations for Android was fair use: the declaring code was "further than usual from the core" of copyright because it was bound by functional constraints.

Clean-room design works because it creates a provable chain of independence. Three elements matter in court: genuine separation between the team that studies the original and the team that builds the new version; substantial effort in the second team's creative choices, proving the output isn't a mechanical translation; and accountable humans, identifiable people who can testify about the process under oath.

For forty years, that was enough. The wall held.

The Machine in the Room

Here is the uncomfortable analogy. An AI language model is trained on billions of copyrighted works: books, articles, code, photographs. It builds a statistical model of patterns, styles, structures, and ideas from that corpus. Then it generates new works that are influenced by those patterns without reproducing them verbatim (most of the time).

This looks a lot like a clean room. The training process is analogous to the specification team: it ingests the originals and extracts functional patterns, stylistic tendencies, structural elements. The generation process is analogous to the implementation team: it produces new works that embody those patterns without directly copying the originals. The statistical model is the wall. Or rather, it's supposed to be.

But the wall has three cracks.

First, the model memorizes. Research consistently shows that large language models can and do reproduce substantial portions of their training data verbatim when prompted correctly. A 2023 study by Carlini et al. demonstrated extractable memorization in production models. In the New York Times lawsuit against OpenAI, the Times showed the model reproducing lengthy passages of its articles nearly word-for-word, what the litigation calls "regurgitation." In a clean room, the specification team extracts ideas and passes them to the implementation team. If the specification contains the original expression itself, the wall is broken. The model's training corpus doesn't just contain specifications. It contains the originals, in full, and the model sometimes leaks them.

Second, the wall is invisible. In a traditional clean room, there are two physical teams, a documented specification, and a chain of custody. A court can examine the process and verify that no copyrighted expression crossed the barrier. In a neural network, the "specification" is a set of billions of floating-point numbers. Nobody can read it. Nobody can verify what it contains and what it excludes. There is no auditable boundary between "patterns extracted" and "expression memorized." The wall exists as a mathematical claim, not a procedural one. You cannot cross-examine a weight matrix.

Third, the speed erases the evidence of independent creation. Phoenix Technologies spent months building their BIOS clone. The time and effort were features, not bugs. They demonstrated that the second team made genuine creative choices. When an AI rewrites a product's core architecture overnight, there's no evidence of independent creative decisions. There's a prompt and an output. The Copyright Office said in January 2025 that prompts alone are insufficient for authorship. If the "implementation" is produced in minutes by a statistical process rather than through months of human creative choices, the traditional evidence of independent creation evaporates.

The Wire Rewrite, Again

We wrote about the wire rewrite in our earlier piece on AI copyrightability. A journalist at a regional paper reads a Reuters dispatch, synthesizes it with other sources, rewrites it in her own voice, adds local context. Copyrightable original work. The same process done by AI isn't copyrightable. Not because the output is different, but because the author isn't human.

That asymmetry was about who owns the output. The clean-room problem is about who can sue over the input, and it runs in the opposite direction.

A journalist who reads the New York Times every morning, absorbs its style and framing, then writes similar articles from memory is doing something perfectly legal. Ideas aren't copyrightable. Style isn't copyrightable. Facts aren't copyrightable. The journalist's output is influenced by but not copied from the source. This is how journalism works. This is how all creative work works. As Justice Breyer wrote in Google v. Oracle, "upholding the copyrightability of the declaring code would risk undermining the basic objective of copyright law."

An AI system that does the same thing at scale, reading millions of Times articles, absorbing patterns, then generating similar articles, is doing structurally the same thing. The Times is arguing it's different. In their complaint, they claim this is "unlawful copying" at an unprecedented scale. The defense argues it's fair use: transformative, non-expressive, and analogous to how humans learn.

In March 2025, the district court declined to dismiss the case, allowing it to proceed to discovery. In May 2025, Magistrate Judge Ona T. Wang issued a sweeping preservation order requiring OpenAI to preserve all output log data. The case is still winding through the courts as of this writing. Nobody knows how it ends. But the question it raises is the same one the clean-room doctrine was built to answer: where's the line between learning from something and copying it?

The honest answer is that we drew that line based on assumptions about how learning works. Assumptions that only applied to humans. Humans learn slowly, forget quickly, and can't reproduce what they've read verbatim without extraordinary effort. AI systems learn instantly, forget nothing, and can reproduce training data with the right prompt. The legal doctrine wasn't wrong for humans. It just wasn't built for machines.

Three Regimes, Three Cracks

Clean-room doctrine lives in copyright law, but intellectual property has three main regimes, and they're all breaking differently.

Copyright protects expression: the specific words, code, brushstrokes. It's the regime where AI causes the most obvious problems, because it's built on the premise that unauthorized copying of expression is infringement, while using ideas isn't. The Thaler v. Perlmutter line of cases (D.C. Circuit 2025, cert. denied March 2026) established that purely AI-generated works can't be copyrighted. But the harder question, whether training on copyrighted works is infringement, remains unresolved in any binding appellate decision. The three key district court rulings from 2025 pointed in different directions, and Reuters characterized the landscape as courts "beginning to draw lines" rather than settling anything. Copyright's "substantial similarity" test, which asks whether an ordinary observer would recognize the accused work as copying the original, is also strained by AI. A model that has read every novel published since 1923 and generates prose that resembles none of them specifically but all of them generally: where does substantial similarity even apply?

Patent protects novel, useful, non-obvious inventions for twenty years. Here AI creates a different problem. In 2020, Stephen Thaler filed patent applications listing his AI system, DABUS, as the inventor. The US Patent and Trademark Office rejected the applications. The Federal Circuit affirmed in Thaler v. Vidal (2022) that only natural persons can be inventors. The same human-authorship requirement that runs through copyright law. The EPO and UK Intellectual Property Office reached the same conclusion. Only Australia's Federal Court briefly allowed AI inventors before being reversed on appeal.

The DABUS cases established that AI can't be an inventor. But AI-assisted inventions, where a human uses AI as a tool, are increasingly common and largely unproblematic under current doctrine. The inventor is the human who conceived the invention. The AI is the tool. The difficulty arises when AI independently generates something novel: can a human who directed the AI then patent the result? The PTO's guidance suggests yes, if the human made the inventive conception. But the line between "using AI as a tool" and "AI inventing independently" is blurry and case-specific. Unlike copyright, patent doesn't have a clean-room problem in the traditional sense. You can't independently invent your way around a valid patent. If the invention is covered by the claims, it infringes regardless of how you arrived at it. But the weaponization of patent law through AI is already happening. Filing thousands of AI-generated patents to create a defensive thicket. Using AI to design around existing claims., and the legal system has no framework for evaluating whether an AI-assisted design-around constitutes genuine independent invention.

Trade secret protects confidential business information that derives value from not being generally known. AI poses perhaps the most immediate practical threat here. A trade secret has no expiration date. It lasts as long as it remains secret. Consider a company with a proprietary pricing algorithm, the kind of thing protected as a trade secret because it can't be reverse-engineered from the product's public behavior. A competitor can't see the algorithm, but they can observe its outputs: prices that change in response to market conditions. If an AI can ingest enough of those observed outputs and synthesize a functionally equivalent algorithm (not by stealing the code, but by independently deducing the logic from public behavior), the trade secret loses its value without ever being "misappropriated" in the legal sense. The Uniform Trade Secrets Act requires misappropriation through improper means. Independent discovery is explicitly permitted. Clean-room reverse engineering of a trade secret is legal if you do it right. AI makes it possible to do it at scale, at speed, and at near-zero cost. The protective moat of "nobody can figure this out" was always a function of human cognitive limits. Those limits no longer apply.

Three regimes. Three different legal frameworks. All built on the assumption that the entity doing the learning, inventing, or discovering is human: slow, limited, and accountable.

The Substantial Similarity Problem

Copyright infringement requires "substantial similarity" between the accused work and the original. The Ninth Circuit has described this as whether an "ordinary reasonable person" would find the works "substantially similar" in their "total concept and feel." It's a subjective standard, applied case by case, and it has always been messy.

AI makes it unworkable.

Consider a model trained on every published science fiction novel. You ask it to write a new one. The output contains no passage that matches any specific novel. The characters are original, the plot is original, the prose is original. But the structure echoes a thousand novels simultaneously. The hero's journey. The first-contact scenario. The unreliable narrator. The twist ending. Every element is unprotectable on its own. But the particular combination, the gestalt, resembles dozens of copyrighted works at the level of "total concept and feel."

In a traditional case, you compare two works. Does this song infringe that song? Does this software infringe that software? With AI, the output isn't similar to one work. It's slightly similar to thousands. No single comparison rises to "substantial similarity." The aggregate effect, the "feel" of the output, is shaped by the entire training corpus. But copyright doesn't have a doctrine for "distributed similarity across ten thousand sources." It was designed for one plaintiff, one work, one accused copy. The test asks whether an ordinary observer would recognize copying. When the copying is spread so thin across so many sources that no ordinary observer could identify any single one, the test returns a false negative. The output is genuinely influenced by copyrighted works in aggregate, but no individual plaintiff can prove it. The doctrine was built for one-to-one comparison. AI is one-to-many. No amount of refining the test fixes a structural mismatch.

Can You Clean-Room Rewrite a Competitor?

Here's the practical question every tech company is quietly asking its lawyers. A competitor has a product you want to replicate. You can't copy their code. That's infringement. But you can use AI to study the product's behavior, generate a functional specification, and have the same AI build a new implementation. Is that a clean-room rewrite?

Legally, maybe. Practically, it's a mess.

The traditional clean room requires separation: one team studies, one team builds, no overlap. When a single AI system does both, the separation doesn't exist. The model that ingested the original also generated the output. The "wall" in a traditional clean room is a procedural guarantee, documented, auditable, testified to under oath. The "wall" in an AI pipeline is a mathematical abstraction that nobody can inspect.

But here's the twist. The claw-code case suggests it might not matter. Anthropic didn't sue over the AI rewrite, even though they aggressively DMCA'd direct forks. Why? Because proving that an AI-generated rewrite infringes requires showing that the output is "substantially similar" to the original. If it's in a different programming language, with different variable names, different abstractions, and different architectural choices, the substantial similarity test is hard to satisfy, even if those choices were influenced by the original. is hard to satisfy. The output looks different because it is different, at the level of expression. It just happens to do the same thing.

That's exactly what clean-room doctrine was designed to permit. The question isn't whether the output is similar in function. It's whether it's similar in expression. And function has never been copyrightable.

So the practical answer, in the short term, is probably yes. You can use AI to clean-room rewrite a competitor's product, and the legal risks are manageable. Which means the competitive moat of "we built it first" is eroding faster than most companies realize.

The Concentration Problem

There's a deeper problem underneath all of this, and it's not about copyright or patents or trade secrets. It's about who can afford the argument.

Clean-room defense is expensive. You need documentation of the process. You need lawyers to certify the separation. You need expert witnesses to testify about the independence of the implementation. You need to survive discovery, which means you need to produce emails, chat logs, internal memos, everything that shows the wall was real.

Big companies can afford this. When Google needed to argue fair use for copying Oracle's Java APIs, they spent an estimated $100 million in legal fees over a decade of litigation. They won. A startup couldn't have made that argument. Not because the law was different for them, but because the cost of proving fair use exceeded their total funding.

AI makes this asymmetry worse in two directions. Companies using AI to produce potentially infringing works need to argue that their process was clean. That the AI's training is fair use, that the output isn't substantially similar, that the functional elements they're replicating are unprotectable. That's a novel legal theory, and novel legal theories are expensive to litigate. Only companies with deep pockets (OpenAI, Google, Meta, Anthropic) can afford to test them.

Companies whose works are being trained on need to argue that AI training is infringement, that the outputs are substantially similar, that the scale of ingestion makes it qualitatively different from human learning. That's also a novel legal theory. The New York Times can afford to make it. Individual journalists, freelance photographers, independent musicians, and small publishers cannot.

The result is a legal landscape where only the largest entities on both sides can afford to litigate the questions that matter. Everyone else operates in legal gray zones, hoping they're not the one who gets sued first. The Yale Law and Policy Review argued in 2025 that an "unregulated AI oligopoly has undesirable economic, national security, social, and political consequences." The FTC has also increased scrutiny of concentration in AI markets. But the copyright dimension of this concentration, who can afford to argue about ownership, receives far less attention.

Clean-room doctrine was supposed to prevent concentration. It was a way for smaller companies to compete with incumbents by building compatible products without infringing. Phoenix Technologies was not IBM's peer. Accolade was not Sega's peer. Connectix was not Sony's peer. The doctrine existed to level the playing field. Now it's a weapon that only the largest companies can afford to wield. And there's a feedback loop: the companies that can afford to litigate set the precedents, and the precedents determine what everyone else can do. The law doesn't come from statutes in this area. It comes from case law. Case law comes from the parties who showed up. When only the deepest pockets show up, the law bends toward the deepest pockets.

What Happens Next

The legal framework is not ready. The Copyright Office's Part 2 report, released in January 2025, concluded that existing law can handle AI copyrightability without new legislation. That's probably true for the narrow question of whether AI output can be copyrighted (it can't). But the broader IP questions (training, substantial similarity, clean-room equivalence) are being resolved case by case in courts that were designed for disputes between humans.

The NYT v. OpenAI case will likely produce the first appellate ruling on whether AI training constitutes fair use. Whatever the result, it will be appealed. The Supreme Court may eventually weigh in. But the Court declined to hear Thaler in March 2026, which suggests it's not eager to confront these questions yet.

Internationally, the landscape diverges sharply. The UK's Copyright, Designs and Patents Act 1988 grants copyright in "computer-generated works" to "the person by whom the arrangements necessary for the creation of the work are undertaken." China's Beijing Internet Court ruled in 2023 that AI-generated images can qualify for copyright when a human made "intellectual contributions." The EU AI Act, which began phased enforcement in 2025, requires general-purpose AI model providers to publish "sufficiently detailed summaries" of training data content under Article 53, creating a new evidentiary source for potential plaintiffs, though still not resolving the underlying IP questions. None of these frameworks address the clean-room problem specifically.

In the meantime, the practical reality is simpler than the legal one. If you can point an AI at a product and get a functional equivalent overnight, and if the output is in a different language with different expression doing the same function, you have something that looks an awful lot like a clean-room rewrite. The traditional tests for infringement are based on comparing expression, and the expression is different. Whether the process was clean is a question that matters in court but doesn't matter in the market. By the time a case reaches trial, the competitor already exists.

We wrote in our earlier piece that "copyrighted human work goes in, uncopyrightable AI work comes out." The clean-room dimension adds a corollary: the legal doctrine that was supposed to protect the process of independent creation is being performed by machines that can't provide the evidence the doctrine requires. Not because the output is infringing, but because the proof of non-infringement (documented human separation, auditable walls, accountable witnesses) doesn't exist when the entire process runs through a neural network in an afternoon.

The clean room didn't fail. It was designed for a world where creation took time, teams were human, and walls were physical. That world is gone.

Sigrid Jin was waiting to board a plane when he saw the news about Anthropic's leak. He pointed an AI at 512,000 lines of code he'd never seen before. By the time the plane landed, the rewrite had 50,000 stars on GitHub. Anthropic sent DMCA takedowns for the forks. They didn't touch the rewrite. That silence is the legal system's answer, for now: we don't know how to handle this, so we're handling nothing. The next case will set the precedent. The companies that can afford to be part of that case will shape it. Everyone else will live with the result.