“We need to make the effort, and not prematurely foreclose the possibility of developing AI to significantly reduce current-day patent transactional costs.”
In response to articles on implementing AI into our patent system, and specifically to the suggestion that we should consider developing AI to replace some aspects of human decision making in the patent space, we have received a number of comments and even objections to the idea.
A common objection: it is likely impossible and impractical for us to advance AI to the point where it can make reliable subjective decisions (e.g., infringement and obviousness), let alone reliably replace human decision making.
At the outset, we challenge the presumption of this argument.
Looking back at the history of human achievement, if we never believed it possible to launch a human past our atmosphere, for a human to survive in the harsh environments of space, and to land safely on the surface of the moon, would we have ever successfully made a moon landing, let alone made the attempt?
Human achievement starts with belief in the possibility. It is only when we believe we can achieve that we can strive to create. This is the bedrock foundation of innovation, and the cornerstone of our discussion here: how to improve the infrastructure of our innovation system.
Reorienting on the Goal of AI
When discussing our patent system and improvement thereto, we must ensure we are speaking from a framework grounded in economic principles.
I emphasize this point because our current thinking around our patent system is not centered around economic principles, but rather character archetypes used for storytelling.
Relying on the economic principles taught by Nobel Memorial Prize co-recipient Douglass C. North, viewing our patent system through an economic framework allows us to understand how a system’s infrastructure creates (1) incentives and (2) actors that manifest within that incentive infrastructure.
The current infrastructure of today’s patent system creates (1) the incentive to engage in patent transactions that are below the exorbitant costs of litigation, and (2) actors that engage in such patent transactions, who are obtusely referred to as “patent trolls” that practice “nuisance-value litigation.”
To create an improved patent system, we must focus on eliminating its core inefficiencies: the exorbitant transactional costs to determine the most basic informational attributes of a patent such as scope, validity, and value.
Compare our patent system to real estate. To purchase real property, the information attributes of the property are easily determined, e.g., obtain an appraisal from a bank. But unlike real property, to learn the value of a patent (e.g., does it read on a product and is it valid), you need to spend millions of dollars and undergo years of litigation in a court of law.
Our discussion here is whether we can incorporate technology to help significantly reduce, if not entirely eliminate, such transactional costs. The technology at the center of attention is AI: can we use AI to eliminate these transactional costs?
As to whether this takes the form of assisting human decision making or actually replacing some aspects of it, we should be open to either possibility.
The State of AI
Developing AI to the point of making subjective decisions such as infringement and obviousness determinations certainly poses obstacles that may seem insurmountable at this point in human history.
For example, AI is still incapable of processing and understanding the intricacies of speech and written word the way a human can. The complexities of natural language processing to date have posed too great a challenge to those in the AI community.
At the outset, this would seem to pose a significant problem if we wanted to apply AI to patents. First and foremost, the most important part of a patent are the claims—words that comprise a sentence that defines the subject matter covered by the patent.
If AI is unable to engage in in-depth natural-language understanding, how are we to apply AI to understand claims and patents, let alone reliably make high-level subjective determinations such as infringement and valuation?
To gain a better understanding, we spoke to Professor Dean Alderucci, the Director of Research for the Center for AI and Patent Analysis at Carnegie Mellon, to learn more about the state of AI and its potential future.
Alderucci acknowledged that at its current state, AI is nowhere near the point of being able to replace the human thinker. He was quite clear in that AI is not good at making decisions, where the decision would be based on AI’s understanding of text.
But Alderucci provided some interesting insights about the capabilities of AI today.
One, he noted that despite AI’s shortcomings, it is good at compiling the information that would be relevant to a human that would be making a decision. As an example, AI may not be good at making an obviousness decision based on its understanding of the meaning of a claim, but it would be good at compiling the references an examiner might consider analogous, and at identifying the specific paragraphs in those references that would be most useful to the examiner in making a decision on obviousness.
Two, Alderucci noted that when it comes to AI understanding technology, it can benefit from the fact that patents are typically targeted in their improvements. For example, if a patent covers a particular technique related to oxidation, AI does not need to understand all of chemistry to understand the technology. The technology can be targeted to oxidation, and perhaps more specifically to particular techniques disclosed in the patent that relate to oxidation. In this sense, AI can be designed and customized to better understand a narrow field of technology like oxidation, which can provide more powerful assistance in analyzing patents, albeit only for patents within the target field of technology. Specialization by technology field can be helpful to, e.g., an examiner or patent attorney whose work entails analyzing a large number of patents in a given field.
Three, AI is also good at discerning between patent elements that are “boilerplate,” and elements that are relevant to the “inventive concept.” This means AI may not need to be an expert at the entirety of a claim but may focus its analysis on the patent claims that capture the “inventive gist” relating to particular technology areas. Among other things, this technology allows the reader of a patent to selectively hide low-information paragraphs when a quick summary is desired.
Taking the above to be true, machine learning may help us bridge the gap left by weak natural-language understanding.
For example, if we take patent applications related to a particular aspect of oxidation and we use AI to isolate elements that capture the “inventive concept,” and if AI presents potentially analogous references to an examiner, it can begin to collect data as to which references an examiner determines to render particular oxidation-related claims obvious or not.
Applying this process to all pending oxidation-related patent applications, could this be used to create a database where AI can begin to apply pattern recognition to determine commonalities between references that render particular inventive claim elements obvious?
Are particular areas of oxidation more prone to obviousness rejections? Do claims with more inventive-gist elements have a lower likelihood of being rejected? Is there a higher or lower rejection rate depending on the patent’s focus of oxidation?
Further, with a large enough dataset, can AI begin to see patterns of particular examiners? Do some examiners apply obviousness rejection more liberally than other examiners?
Can this allow us to begin to build a foundation for AI to develop pattern recognition to help us determine the likelihood patent claims would be rendered obvious or not?
Maybe. Alderucci at least acknowledged this could be a start.
We Need to Make the Effort
The above approach may or may not work when it comes to developing AI’s ability to help make high-level patent decisions.
The point is we need to make the effort, and not prematurely foreclose the possibility of developing AI to significantly reduce current-day patent transactional costs.
Today’s patent system and its fundamentally weak economic underpinnings do not give us the latitude to pontificate what may be possible or not. We need to make a directed and sincere effort to enhance AI’s applicability to “improve” our patent system.
It starts with the belief that we can.
Image Source: Deposit Photos