Software Patents are Obsolete in the Age of AI
Businesses need safeguards and moats to grow. In order for investors to justify allocating capital to a business, there have to be mechanisms to ensure that value created doesn’t leak out. Traditionally, this value has been safeguarded by a variety of legal mechanisms. For example, in markets where products are scarcely differentiated from one another, branding becomes critical. Tide detergent and Costco generic detergent are likely near identical, but Tide has the advantage of strong branding and packaging. In order to protect this value, Tide can take out trademarks protecting its distinctive packaging and visuals. In the pharmaceutical industry, a company’s critical intellectual property has resided in novel chemical entities. The effort expended in these inventions is protected by patents that provide the right to use the novel molecule for a given purpose. The rule of law ensures that knock-offs can’t ignore these protections and piggy-back on Tide’s brand or Merck’s molecules.
This system has worked powerfully for centuries. After originating in Venice in the 1400s, and being adopted by England in the 1600s, patents and other legal instruments have served as powerful instruments for rewarding intellectual effort and research. Companies have used these protections in order to build and grow. It wouldn’t be an exaggeration to claim that today’s capitalistic society depends critically on legal instruments like patents and trademarks. That said, the patent system is showing its age. While the system of protection offered by patents and trademarks works well for physical products, standard measures of novelty and invention have proven less meaningful for the software realm, with the advent of AI threatening to make them entirely obsolete.
This issue has become increasingly important over the last thirty years, as many companies have started to shift onto software driven foundations. Products including cars, phones, computers, washing machines and more have increasingly come to depend on software to provide critical services. When this transition started, adapting the patent system over to software worked reasonably well. Companies were allowed to patent the application of particular algorithms or computational methodologies to specific problems or devices. Creating an implementation of a new algorithm suited for a new product space took considerable effort. As with physical products, it was hard for other entrants to run afoul of existing patents without considerable focused effort.
In the last twenty years though, the situation has changed considerably. The growth of the open source software movement alongside the rise of the internet and collaborative sites like github and stack overflow have made the construction of new software systems increasingly straightforward. Code for many common tasks has become trivial, with copy-paste and basic configuration scripts allowing for nearly automated design of simple websites and apps. The critical result is that construction of many software systems now requires significantly less novelty or innovation than previously. In this environment, a patent that controlled methods for building websites would likely fail a novelty test, since the techniques have become common knowledge for software engineers.
As a result, software patents have gradually become less meaningful. It’s often possible for companies to duplicate (at least approximately) their competitors software creations with some amount of effort. Consequently, with some high profile exceptions, such as Oracle vs. Google, companies don’t often attempt to exercise software patents. It’s easy to unknowingly violate these patents, and the effort required for enforcement is considerable. Only companies who expect significant legal upside (like Oracle) attempt to prosecute. Nonetheless, it’s still common practice for companies to build a war chest of “defensive patents.” Building out a base of patents for computational techniques lets companies defend against “patent trolls,” entities which seek to extract value from businesses by claiming ownership of relevant patents. More importantly, investors tend to love patents, especially for businesses that are flailing otherwise. The patents can be bundled up and sold off in case the business isn’t doing well. This state of affairs has persisted for some time. Engineers remain skeptical about patents, trolls abuse the system, and investors profit from patent hoards. But is this situation tenable in the long run?
In the last five years, there have been dramatic advances in AI and learning. These techniques allow computational handling of images, videos, speech, molecules, measurements and more with dramatic improvements in accuracy and in ease of use. In particular, the advent of open source deep learning systems such as Tensorflow, PyTorch, and DeepChem have enabled even novices to build considerably sophisticated learning systems using commodity GPU hardware. These improvements have enabled Google to swap out their highly engineered and tuned Google translate system for a new neural machine translation system with only a few months’ effort. The neural implementation led to dramatic increases in accuracy; more powerfully, it’s likely the new Google translate was dramatically easier to construct than the original. Tools like Tensorflow have made the construction of simple neural machine translation system into homework exercises for computer science classes. Production quality systems learning systems require more effort, but a good team could probably build a reasonable translation engine for another company without too much pain.
What does it mean for patents when open source learning tools allow for the construction of systems that would have taken years to construct previously? As deep learning techniques extend their reach to all parts of the software ecosystem, this question will become increasingly important. Once algorithmic systems can learn to duplicate sophisticated hand-tuned software packages with minimal effort, the novelty bar for software inventions will need to change dramatically. Patents for applying learning algorithm X for application Y or system Z should not be granted, since these connections will become (and are already) common knowledge for practitioners.
Unfortunately, such patents have been granted widely already. In one egregious case, Health Discovery Corp. took out patents for the application of support vector machines to biomarker discovery, which it consequently has attempted to enforce and extract royalties for in the last few years. Given that the use of support vector methods for biological data can be performed using two lines of python, the situation is patently ridiculous. Nonetheless, companies continue to amass software patents for both classical methods and learning techniques. Google, despite its strong stance on open source software, has applied for troves of “deep learning architecture X for problem Y” patents. These patents are somewhat silly, given that students with 6 months of AI training can easily create prototype versions of systems covered by these patents, often using tools created by Google itself. While Google has not attempted to enforce these patents, it’s impossible to predict what could happen if Google falls upon hard times financially.
Why does it make sense anymore to offer up patents for applications of these new AI techniques? Violating these patents is straightforward. Nevertheless, companies entering new spaces have to step gingerly to avoid issues. If Google or Facebook or Microsoft decided it were in their interests to enforce the patents they held, the entire machine learning software ecosystem could collapse in on itself. More fancifully, while artificial general intelligence remains distant, it’s not inconceivable that eventual intelligent systems will depend critically on technologies patented today. Will these intelligent children of our minds be born into predetermined slavery, with parts of their “thought” process owned by descendants of today’s corporations?
Fixing this state of affairs requires a reassessment of the role of software patents. When software patents have begun to fail at their intended role of encouraging innovation and business, continuing to obtain and value them threatens the health of our business and societal ecosystems. Companies can take some positive actions today to ameliorate the situation. First, large companies like Google or Facebook should “open source” their patents, relinquishing ownership of the capital for the common good. Releasing patented technologies into the wild will allow for growth of software ecosystems, and will prevent internecine battles over foundational technologies. Second, companies and (more importantly) investors need to stop insisting on patenting basic algorithmic applications. The idea of “deep learning X for Y” is very unlikely to be unique; there are probably dozens of engineers elsewhere who have working prototypes already. Attempting to take control of common knowledge for personal enrichment will only hurt the common ecosystem. Instead, companies should open source code, publish papers, and contribute to a healthy software ecosystem. Seek to build value by creating strong brands, customer loyalty, and physical products that aren’t so easy to duplicate. Building strong software ecosystems and avoiding frivolous patents will only benefit the technology industry and society at large.