AI inventors may find it difficult to patent their tech under today's laws
Why newfangled algorithms could be a headache for examiners
Comment Future AI could be a challenge for US Patent and Trademark Office (USPTO) officials, who need to wrap their heads around complex technology that's perhaps not quite compatible with today's laws.
Under the Department of Commerce, the USPTO's core mission is to protect intellectual property, or IP. Creators file patent applications in hope of keeping competitors from copying their inventions without permission, and patents are supposed to allow businesses to thrive with their own novel designs while not stifling wider innovation.
Fast evolving technologies, such as deep learning, are pushing the limits of today's IP policies and rules. Clerks are trying to apply traditional patent approval rules to non-trivial machine-learning inventions, and bad decisions could result in a stranglehold on competition among public and private AI creators. We all know how overly broad patents on software and other technology can make it past USPTO, causing headaches for years to come.
"AI is already impacting most industries and many aspects of our society," Kathi Vidal, the agency's director and a former engineer, said during the inaugural meeting of the AI and Emerging Technologies (ET) Partnership Series held virtually last month.
"AI and emerging technologies have the potential to dramatically improve our day-to-day lives. They will provide countless and unpredictable benefits to our social well-being not just here in the United States, but around the world. But the bottom line is, we need to get this right.
"We need to make sure we're setting laws, policies and practices that benefit the US and the world."
Publishing patents disseminates valuable knowledge, giving engineers and scientists ideas on how to advance technologies or invent new ones. Inventors have to meet a list of criteria in order for their applications to be considered. Not only do they have to demonstrate their invention is novel, non-obvious, and useful, they have to describe their work in a way that someone skilled in the same field can understand and reproduce it.
And here's the rub.
Neural networks aren't easily explainable. The number-crunching process that seemingly magically transforms input data into an output is often opaque and not interpretable. Experts often don't know why a model behaves the way it does, making it difficult for patent examiners to assess the nitty-gritty details of an application.
Furthermore, reproducibility is notoriously difficult in machine learning. Developers need access to a model's training data, parameters, and/or weights to recreate it. Providing this information in a patent application may satisfy examiners, but it may not be in the interests of the inventors or the wider public.
Medical data taken from real patients to train an algorithm that can detect tumors, for example, is sensitive and opens up all sorts of risks if it is handed over for government agency workers to process, publish, and store. Full disclosure of the system may also reveal proprietary information. It may be easier in some cases to not patent the technology at all.
The USPTO previously hit a stumbling block when it came to applying patent law to AI inventions. Mary Critharis, USPTO's chief policy officer and director for international affairs, noted the acceptance rate for AI patents dropped in comparison to non-AI inventions in 2014 following the US Supreme's Court decision [PDF] in the Alice Corp vs CLS Bank International case. Justices ruled CLS could not have infringed Alice's financial computer software patent, because it was too abstract.
Like laws of nature and natural phenomena, abstract ideas can't ordinarily be patented. The Supreme Court ruling may therefore have had a chilling effect on AI patent applications and acceptance, as they too may have been assumed to be too abstract, at least until further guidance was issued to patent examiners on how to deal with abstract designs.
"[The data] provides some suggestive evidence that the Alice decision impacted AI technologies," said Critharis.
"The allowance rate stayed below the non-AI application rate until about 2019. The reason for this was that in 2019, the USPTO had issued revised subject matter eligibility guidance," she continued, referring to the advice discussed here [PDF].
"I think this is the reason why we're seeing an increase in allowance rates, but there was definitely an impact of the Alice decision on AI related applications."
As machine learning evolves, and more patents are applied for and picked apart in court, we could see another dip in allowance rates.
Last year, a group of US senators said there is "a lack of consistency and clarity in patent eligibility laws," and asked the USPTO to clarify what inventions are patentable and why. "The lack of clarity has not only discouraged investment in critical emerging technologies, but also led the courts to foreclose protection entirely for certain important inventions in the diagnostics, biopharmaceutical, and life sciences industries," they wrote in a letter.
Clear guidance from the USPTO is helpful in encouraging inventors to file patents more successfully. But advice only goes so far. US courts, ultimately, have the final say in these matters.
- Vodafone picks Google for MLops project that self-serves models in minutes
- IBM buys Databand to keep bad data from tainting your AI
- Meta's AI translation breaks 200 language barrier
- Health trusts swapped patient data for shares in an AI firm. They may have lost millions
And, separately, it's not clear if and how AI-generated technologies can be patented. Who owns the IP rights of art, music, or writing created using generative models? These creations riff off existing content and can mimic certain styles. Do they violate copyright?
Can these models be listed as inventors if they create content? Current US laws, at least, only recognize IP produced by "natural persons" much to the chagrin of one man. Stephen Thaler sued Andrei Iancu, the former director of the patent office, when his application listing a neural network system named DABUS as an inventor was rejected.
There hasn't been a significant commercial application of these technologies in a way that will precipitate what will be the next patent war in the sense that there was the sewing machine patent war
It could get interesting if, as some legal experts believe, people start filing patents for inventions devised and optimized by automated machine-learning algorithms. These inventions may not be entirely novel but the way in which they were produced was; will these be accepted, or is it an obvious rejection?
The USPTO cannot definitively answer all these questions; some of these issues will have to be tried and tested in court.
"There haven't been a lot of court cases on AI yet," said Adam Mossoff, Professor of Law at the Antonin Scalia Law School at George Mason University, during a panel discussion.
"There hasn't been a significant commercial application of these technologies in a way that will precipitate what will be the next patent war in the sense that there was the sewing machine patent war, and there was the patent war over fiber optics, and there was the patent war over disposable diapers and everything else. And when that happens, I think we're going to see a real concern here."
The UPTSO has asked the public to comment on current policies that describe what inventions can or cannot be patented.
Some people thought the agency was effective at issuing patents and helping protect inventors against patent trolls, while others disagreed and said the agency's framework stifles innovation for small businesses and startups.
A recent report [PDF] from the agency concluded that everyone did agree on one thing: "The standard for determining whether an invention is patenting should be clear, predictable, and consistently applied." ®