On March 15, 2016, AlphaGo—a machine designed by researchers at DeepMind, an artificial intelligence laboratory owned by Google—defeated Lee Sedol, the best Go player in the world. Go is a popular board game played by two people; the goal is to surround more points on the board than your opponent. For years, Go had posed a unique challenge to computer scientists. The rate at which possible moves in the game increase was so significant that most methods of calculating the optimal move required resources beyond the reach of a computer program. When humans play Go, they do not run through every single possible outcome, but they can still select a good move. Computers, however, cannot divine the optimal move. To decide how to move, AlphaGo learned how to play Go through techniques such as “deep learning.” These complex processes enabled a computer to master a game previously thought to be solely within the province of human skill.
After AlphaGo won, not only did people hail the machine as a dramatic advance in artificial intelligence, but many were quick to suggest that this heralded the decline of human’s advantage over machines in a wide variety of fields. Even though people have suggested that machines will replace humans before, AlphaGo, and machine learning more broadly, suggest that machines may soon start replacing people even in more creative industries, such as the arts or engineering. This may raise significant issues for the current intellectual property regime: if a computer invents something, who owns the patent? The current intellectual property regime in the United States does not provide a ready answer to this question, at least as to patents. Given the challenges of defining ownership, inventions created purely by artificial intelligence should not be patentable.
Machine Learning’s Promise
Machine learning aims to give computers the ability to solve problems and learn without any explicit programming that directs the computer. From a technical perspective, machine learning covers an extraordinarily wide range of different approaches, including decision tree learning, artificial neural networks which try to mimic the human brain, and Bayesian networks which rely on probabilities. These tools have enabled an even broader range of products: self-driving cars, speech recognition software, and more effective web searches.
Many have proposed even broader applications of machine learning. Coupled with data mining techniques and tremendous amounts of information—“Big Data”—machine learning algorithms may be able to help detect and respond to threats ranging from terrorism to natural disasters. Likewise, these algorithms have been applied to a wide variety of legal tasks, ranging from document review to predicting the outcome of a particular case. Machine learning, then, is poised to revolutionize fields previously thought immune to automation.
Challenges for the Patent System
Some companies have already begun trying to use advanced computing power as a tool to patent new inventions. Back in 2014, a company called Cloem began providing software that would manipulate the text of a patent claim and replace key words with synonyms or reorder particular steps in the sequence. Cloem claimed that these new patents would satisfy the inventiveness requirement and therefore could be patented. While Cloem’s usefulness in creating novel inventions remains questionable, others have patented new inventions relying on computer-based designs.
Combining these existing processes with more advanced machine learning tools raises the possibility of inventions created purely by machines. In theory, computer-generated patents should be patentable even if no human is involved beyond the initial coding that develops the machine learning algorithm. 35 U.S.C. § 103 states that “patentability shall not be negated by the manner in which the invention was made.” However, 35 U.S.C. § 100 defines “inventor” as an “individual,” suggesting that the invention must be made by a person. One article has suggested that, if the technology used is similar to Cloem’s software, then the resulting patent should belong to the inventors of the “seed” patents—the patents which provided the base language or information which the software then examined and manipulated. The authors likewise conclude that inventions that are computer-generated should be patentable because they have the potential to “promote the Progress of Science and useful Arts.”
However, these conclusions assume that the computer generating the patents will be running a Cloem like program, not using machine learning to create actually creative works unrelated to or only tangentially related to any initial inputs. The whole premise of machine learning is that the computer can generate outputs that are not explicitly programmed and, in that sense, are created by the computer. Because of the language of the patent statutes, no human could claim this invention because they will not have contributed to the invention. While the software itself may be patentable, any resulting outputs would not be.
This is not necessarily a bad thing: companies that have the skills and resources to build patent-inventing software using machine learning still have an incentive to do so. They can patent the software and sell it to others or possibly rely on trade secrets protection while they work on developing any ideas generated by the machine into fully-formed products. Alternatively, U.S. patent law may need to be amended to account for inventions created by computers. While this may seem like science fiction, achievements like AlphaGo, which came years earlier than anticipated, remind us that such technologies may appear sooner than we expect.