Future of Software Patents with AI


With significant changes from and lack of clarity around software patent viability from the US PTO, how do you think that in an AI focused world (where much of the key technology is a trained CNN versus the algorithm to train it) patents will be utilized?

Artificial Intelligence
Patent Applications
Darin G
56 months ago

5 answers


Software patents are difficult to get. Usually, the software needs to be coupled with a system or machine or physical mechanism before a patent can be considered. Algorithms are not patentable.

The words "AI" today in the media usually refers to a dozen or so commodity machine learning algorithms that can be applied to various data sets. These algorithms are well known (See O'Reilly's "Hands on Machine Learning" book), there are people out there offering machine learning services for the same hourly wage as a Java programmer, so no novelty here. And while machine learning algorithms are highly useful for the right types of data sets, these, surprisingly, are not smart algorithms. They are nonlinear statistical algorithms that match patterns to data, somewhat more sophisticated than regression. Simply put, machine learning is not rocket science. Huyndreds of thousands of people known and understand how to use it. Don't be duped by the media.

Other AI algorithms are smarter - like knowledge based systems (not novel), and heuristic search (also not novel, unless you consider context), machine vision (really just sophisticated image processing), and natural language processing (almost a commodity). Harder problems of perception, consciousness, and and true planning are not well solved yet.

What I am saying is that Ai algorithms are mystic until they work, then they are commodities because everyone now understands them. A patent can only be had if these algorithms are coupled with data and a machine, device or system to solve a problem in a novel way that is not obvious. Quite hard, not impossible.

It is a double edged sword. For example, the application of machine learning to drug discovery has caused the USPTO to consider that the drugs designed this way are no longer patentable because they were obtained by an "obvious" means. So there is danger that AI will make things less patentable, because as AI becomes obvious, the novelty is created by an obvious means.

Ed A
56 months ago

Patents focus on physical technologies. Software technologies are difficult for the patent process. If the server sits in another country, it can get very tricky to decide what patents are being infringed. Add to that the fact that most countries do not permit the patenting of mathematical, algorithmic or software patents. In patenting, you generally have to make quite a lot of your IP public, which really only helps the infringer. Obtaining a patent is a slow process and AI develops fast. By the time you've patented your cool new thing, it's already outdated and the 30 year right is of no use to you. For patents to remain relevant in a digital world, we would need to overhaul the entire process of obtaining and enforcing patents. In the present scheme, I believe patents are more harmful than helpful to innovations relating to AI.

Patrick Bangert
56 months ago
Exactly. The IP of an AI company is the resulting CNN that they trained. It's not a process...its an algorithm I suppose, if you get down to it...just one that no human created. Does that mean that there will be less patents and more trade-secrets? And then how does one prove (the good side of patents...see Waymo vs Uber) that someone has infringed or stolen that trade secret w/o disclosures? - Darin 56 months ago
Let's suppose that you trained a CNN and that this is your IP. If you want to patent it, then you would have to publish the matrix coefficients. I could type them into my software and you could never prove it. Even worse, what if I changed a single coefficient in the sixth decimal place ... my CNN is not the same as yours. - Patrick 56 months ago

I’m expecting the patents to be more application oriented.

Mohammad M. Aref
56 months ago

IP assets are growing globally, WIPO reported almost 8% from 2014 to 2015. For example I read that IBM registered and then sold more than 15K patents since 1991 having Google as the biggest customer.

Paolo Beffagnotti
56 months ago

A patent gives the patent owner the right to exclude others from making, using, selling, or importing the invention. So, like all other patents, that is how software patents may be used (utilized). In practice, patents may be licensed or cross-licensed. I expect that the later will a big part of the strategy of any AI patentee.

Meg B
56 months ago

Have some input?