Clean Room

When a company wants to recreate someone else’s product, it faces a legal problem. Anyone who copies the original text of a program infringes copyright. A clean room is a working method that gets around this problem. Two separate teams work on it: the first examines the original and writes down what it does. The second team builds the recreation without ever having seen the original. It only receives the description from the first team – and therefore cannot copy anything at all.

Why recreating is allowed, but copying is not

Copyright protects the concrete formulation, not the idea behind it. A cookbook is protected; the recipe “pasta with tomato sauce” is not. The same applies to software: the written program code belongs to its author. The function it performs belongs to no one. Anyone who achieves the same function with their own code does not infringe copyright.

In practice, this is hard to prove. A court will ask whether the developers knew the original code. If so, there is a strong suspicion that they were at least unconsciously guided by it. This is exactly where the clean room helps. It creates a seamless documentation of who had which information and when. In court, this is worth far more than simply asserting that one did not copy.

Economically, a great deal is at stake here. In the 1980s, the company Phoenix Technologies used this approach to recreate the BIOS of the IBM PC, i.e. the computer’s startup software. Only this made it possible for other manufacturers to build compatible PCs. Without this method, the market for PC clones might never have emerged.

The two separate teams

The first team is often called the “dirty” team. It is allowed to look at everything: the code, the manuals, the running program. Its task is to describe the behavior of the original. In other words: which inputs lead to which outputs? What interfaces exist? This description must not contain any wording from the original.

Before the description is passed on, a legal department usually reviews it line by line. Anything that sounds too close to the original is struck out. Only then does the second team, the “clean” team, receive the document. These developers have never seen the original and are not allowed to look at it either. They program solely on the basis of the description.

The entire process is logged. Who was on which team, who had access to which file, which version of the description was released when? A related term from banking is the “Chinese Wall”, i.e. an information barrier between two departments of the same company. The principle is the same: knowledge is deliberately not passed on.

Clean rooms in the dispute over AI training data

Today, the term appears mainly in news about legal disputes. When a company recreates a competitor’s product, it likes to emphasize that this was done “using the clean-room method.” This is a signal to investors and courts: we have covered ourselves. The method also became known through projects such as ReactOS, an operating system intended to run Windows programs without containing any Microsoft code.

In the AI industry, the term takes on a new meaning. Many language models were trained on text and images from the internet whose rights belong to others. This has given rise to numerous lawsuits. Some companies therefore deliberately build models only from licensed or public-domain data and advertise this clean provenance.

A note on the risk of confusion: in the chip and pharmaceutical industries, “clean room” means something entirely different, namely a dust-free production space with filtered air. Both meanings share only the basic idea – keeping out something that would contaminate the process. In one case it is dust particles, in the other it is foreign knowledge.