At Google’s cloud computing conference in San Francisco last week, CEO Sundar Pichai mused on his company’s commitment to openness in artificial intelligence.
“We create open platforms and share our technology because it helps new ideas get out faster,” Pichai said. Then he name-checked TensorFlow, the machine-learning software Google developed and uses internally. The company open-sourced the code in 2015, and it has since been downloaded more than 15 million times. “We created TensorFlow to make it possible for anyone to use AI,” Pichai said.
Such homilies to openness have become standard from the large tech companies competing intensely to develop AI technology. Facebook, Amazon, and Microsoft have all, like Google, released software their own engineers use for machine learning as open source. All, including to some extent famously secretive Apple, encourage their AI researchers to publish their latest ideas—helping the companies recruit the brightest faculty and grad students from universities.
At the same time, these proponents of AI openness are also working to claim ownership of AI techniques and applications. Patent claims related to AI, and in particular machine learning, have accelerated sharply in recent years. So far, tech companies haven’t converted those patents into lawsuits and legal threats to thwart rivals. But should AI patents become corporate weapons, the current openness around AI research and ideas could end, likely hampering research.
A National Bureau of Economic Research study released this month shows US filings related to machine learning, the technology driving the current AI boom, increasing rapidly. “We’ve seen a huge explosion of patenting activity in AI and machine learning, and I see this exponential growth continuing,” says Michael Webb, a Stanford researcher and coauthor of the study.
In 2010, there were 145 US patent filings that mentioned machine learning, the study says. In 2016, there were 594—a figure that’s incomplete, since the US Patent and Trademark Office only makes filings public 18 months after they have been registered. (Webb and his colleagues gathered their data in February.) Patent filings mentioning neural networks, a machine-learning technique, climbed to 485 in 2016, from 94 in 2010.
Google itself exemplifies the trend. In 2010, only one Google filing mentioned machine learning or neural networks in its abstract or title, according to a search of the USPTO database. In 2016, there were 99 such filings from Google and other Alphabet companies. Facebook filed for 55 patents related to machine learning or neural networks in 2016, up from zero in 2010. IBM, which has been granted more US patents than any other company for the past 25 years running, boasts that in 2017 it won 1,400 AI-related patents, more than ever before.
It’s not surprising that patent filings related to AI are increasing. In 2012, neural networks suddenly became a hot topic of interest from tech companies, after they enabled big improvements in speech and image recognition. But the moves to lock up technology contrast with the emphasis on openness in companies’ public discussion of their AI strategies.
The patent-filing surge recalls the fierce battles over intellectual property in the last big tech revolution, around smartphones. Apple and Samsung have fought at least 50 lawsuits over technology and designs for smartphones, according to the NBER paper; Apple and Google tussled in about 20.
More patents filed in a particular area make lawsuits more likely, says Richard Abramson, a Stanford lecturer who previously was general counsel at independent research institute SRI. “If you give everybody a gun you can pretty much bet the incidence of shooting is going to rise,” he says.
Litigation over AI could harm the open progress tech giants say they want. Twenty-five years ago, patent suits were mostly disputes between companies fighting to use specific technology in their products, Abramson says. Today many are brought by companies, often dubbed “trolls,” holding patents they don’t plan to use for anything but extracting compensation. “Now companies are freaked out by patent troll activity, and a lot of them stockpile patents in order to have something to shoot back,” he says.
There’s no indication yet that any of the leading AI companies are working to leverage their AI patents. Spokespeople for Google and DeepMind both said that their companies hold patents defensively, not with the intent to start fights with others. Google’s spokesperson also noted that the company accounts for a small minority of recent AI-related filings. A Facebook spokesperson said its filings shouldn’t be read to indicate current or future plans. IBM’s chief patent counsel, Manny Schecter, said the company’s patent horde reflects its investment in fundamental research.
Those statements leave room for future changes in policy. Given the recent history of fights over tech patents, some researchers still worry the AI patents piling up could be used in ways that could stifle progress. Assessing the value and scope of patents is complex, and even expert interpretations can vary. But some of the applications filed by Google and others appear to describe fundamental techniques with broad applications in research, says Miles Brundage, who researches trends in AI development at the University of Oxford. “It’s not had an impact yet, but this might be a ticking time bomb,” he says.
DeepMind, the Alphabet unit behind the AlphaGo software that defeated a world champion at Go, has filed applications including on DQN, an extension to a learning algorithm originating in the 1980s that helped DeepMind software master Atari games. Since DeepMind published academic work on DQN, researchers elsewhere have explored and extended its insights.
Google has a patent pending on dropout, a now-standard technique used to help neural networks generalize to new data they were not trained on. One Facebook application covers an approach to designing neural networks dubbed memory networks, which enhance a conventional machine-learning system for processing text with a kind of short-term memory.
Mark Riedl, a professor at Georgia Institute of Technology currently on leave to work at Salesforce’s AI research group in Palo Alto, says patents on algorithms and other fundamental machine-learning techniques make him uneasy. The patents filed so far haven’t yet caused problems for researchers, but assigning legal ownership of relatively abstract ideas doesn’t fit with the open progress that has lately made machine learning so exciting, he says.
Not all the patents recently filed on AI ideas and techniques will be awarded. Patents on software have become harder to get since a 2014 Supreme Court ruling that merely implementing an idea on a computer isn’t enough to make it patentable. And last year, USPTO significantly expanded the number of examiners dedicated to scrutinizing AI patents, something expected to screen out more applications.
Big changes in what kinds of AI ideas can be patented seem unlikely, though. “The companies filing a lot of applications in this space are a big portion of the economy,” says Joe Holovachuk, a patent attorney with the firm Pepper Hamilton. That means they can pay for lobbyists and lawyers to push lawmakers or courts to support their favored approach—which seems to be making all kinds of AI techniques broadly patentable.
In what could be music to the ears of tech companies, the director of the Patent and Trademark Office, Andrei Iancu, has signaled that he’s been thinking about AI patents. In April he told the Senate Judiciary Committee that he believes recent court decisions have muddied the question of whether algorithms can be patented. Iancu thinks algorithms, including in AI, generally always can be. “We have to make sure our policies, including IP, are highly focused on incentivizing that type of innovation,” he said.