Patenting Artificial Intelligence
Under today’s patent law, could Star Trek’s Dr. Soong receive a patent on Data’s positronic brain? Dr. Frankenstein, a patent on his method of making a monster? Dr. Alfred Lanning (of Isaac Asimov’s I, Robot) on his child-like robot Sonny?
Companies Seek Broad Patent Protection for AI
Artificial intelligence (AI) has already made the leap from fiction to non-fiction. Alexa can place an order for groceries, while Siri can answer your questions. One AI system uses human-like vision to diagnose cataracts in children with as much or more accuracy than actual ophthalmologists. And it is rumored that Google’s own Neural Machine Translation system was shut down when it spooked its handlers by creating its own internal language that couldn’t be decrypted. In short, everyone from start-up companies to Google is racing to create the best, most efficient, most intelligent AI. And when they create it, they patent it.
Currently, AI patents and applications generally fall into two categories: (a) AI algorithms themselves; and (b) use AI for specific applications – e.g., self-driving cars and medical diagnosis. The U.S. Patent and Trademark Office (USPTO) contains numerous classifications for AI-related inventions, ranging from fuzzy logic hardware to machine learning, neural networks, knowledge processing systems, and applications using AI.
Google has filed for patents covering AI such as neural networks, parallel convolutional networks, natural language processing, classifying data objects, word embedding, and a method for neural network learning. One Google-developed algorithm identifies skin cancer in patients by comparing samples with over 130,000 high-resolution images of skin lesions representing 2,000+ diseases, but rather than covering the narrow scope of what they are meant to do, the Google patents are broad, seeking to maximize their scope of protection. They seek to cover the fundamental machine learning techniques that may eventually provide the underpinning for more refined and advanced work.
Impact on Start-Ups and Innovation
Smaller start-ups therefore worry that anything they create may fall under these wide-ranging patent claims if they are eventually granted. Conversely, they have no other choice but to rush to patent the AI they do invent – and to likewise file broadly. It’s an AI arms race . . . and to the victor go the artificial spoils.
What if, dozens of years ago, the first patent granted on a computer was so broad that it could have covered any computer in existence today? Would there be as much innovation in the computer space if innovators were afraid of infringing overly-broad patents on the first computers? Google received a patent on a common machine-learning technique called dropout, and can therefore prevent anyone else from using this technique until today’s preschool kids graduate from high school in 2032. Microsoft has another one pending on active machine learning. Will today’s broad AI patents and fear of infringement, stifle innovation as we dive deeper into the art of the mechanical brain? The USPTO and the courts are becoming increasingly careful about granting broad software patents, but the damage may already be done.
Additional Methods for Protecting AI
There are alternative methods of protecting AI, but none provide as strong protection. For example, trade secret is the inverse of patent law. Instead of filing for a patent and making your invention public in exchange for exclusivity, trade secret allows for keeping the AI under wraps. But a company must then protect these secrets, and if they are disclosed, or if Company B applies for a patent on their version of the invention while Company A is busy keeping their version secret, Company A may not be able to continue to use their invention without infringing on Company B’s patent.
The second option is copyright protection. Programmers can file copyrights for all written codes, much like authors file copyrights for their books. However, the copyright does little else besides prevent third parties from replicating the exact same code, or near-identical versions. A copyright will not stop a second programmer from creating a different code that accomplishes the goal but using different enough language to avoid infringement. Ultimately, what matters most in terms of AI and computer code, is not so much the actual computer language and code but the functionality of the software, and patent protection is usually the most effective way to protect that element of inventorship. Whether the pattern of broad patents will continue, whether law and statute will force them to become more narrow, and whether future AI refinements will be encouraged or discouraged by existing patent monopolies, remains to be seen.
If you need more information about patents for your inventions, please send us a message. Our attorneys have over 100 years of combined experience and are licensed in Arizona, Connecticut, and New York. We can handle federal intellectual property matters in any U.S. state and assist with international matters. For even more information, sure to connect with us on Facebook.