No Man’s Sky, a highly anticipated video game released this summer, allows players to explore a massive game world consisting of eighteen quintillion planets. Creating this much content would have been impossible for human game designers, so Hello Games used a technique called procedural generation. This technique relies on a computer program to design much of the game content, with the human developers creating a framework around which the AI designs the content. Legally, however, there is a catch to procedural generation. Generally, only content created by humans can receive copyright protection, and U.S. legal authorities do not agree on whether copyright in content can be vested in the creator of the computer program that creates the content. The Copyright Office takes the position that it cannot, but the Second Circuit, at least for relatively predicable programs, has taken the position that it can.
There are other problems with granting copyright for procedurally generated content aside from the immediate author lacking legal personhood. Programmers can easily create algorithms that would fence off vast amounts of creative real estate. A useful example is the Library of Babel. This website contains every possible combination of 3200 characters, including this very article. Jonathan Basile certainly does not own all future one page writings. So any doctrine that allows copyright in procedurally generated works would have to contain a limiting principal that would exclude that outcome.
With the law uncertain, game developers might take the cautious approach and assume that they cannot get copyright protection in their procedurally generated content. Even under this assumption, the scope of the rights lost by employing procedural generation can be very narrow. In No Man’s Sky, for example, the human developers designed a large library of interchangeable body parts. They then let the AI loose to play Mr. Potato Head with these body parts. The developers can own copyrights in the all of the in-game objects they designed. It is only the mixing and matching choices that are potentially uncopyrightable. So game developers like Hello Games remain fundamentally protectable when they create games using procedural methods.
Other AI researchers, however, may not be able to shield themselves from the legal risks associated with computer generated content. The role of AI in some applications is much more fundamental. Google’s AlphaGo beat 9-dan Go player Lee Sedol by analyzing a database containing thirty million moves and board states from professional Go matches, playing thousands of games against itself, and designing its own algorithm for maximizing its odds of winning. If code written by AI is uncopyrightable, unlike in procedurally generated video games, which lose protection for cosmetic choices, Google would lose protection for the core operational features of AlphaGo. Extreme choices could lead to a system in which Jonathan Basile owns almost all new content, or one in which no one owns the vast majority of new content. Currently, it is difficult to discern who will own what once machines are designing most of our machines.