Artificial Intelligence Systems Are Not Inventors… Yet.

Recently, an Artificial Intelligence (AI) named DABUS (“Device for Autonomous Bootstrapping of Unified Sentience”) made international media buzz when it was listed as an inventor on patent applications filed in the United States Patent and Trademark Office (USPTO), United Kingdom Intellectual Property Office (UKIPO) and the European Patent Office (EPO).  While there have been thousands of patent applications for various AI systems, from machine learning (ML) systems, to neural networks, generative adversarial networks (GANs) and beyond, this is at least one of the first AI systems to be listed as the inventor itself.

What is DABUS

Developed by Stephen Thaler, a well-known expert in the field of AI, DABUS is an implementation of a type of generative adversarial network (GAN), where two or more neural networks are used to play against one another in order to improve the outputs of at least one of the neural networks to create something different and new.

In the most common implementation, a GAN comprises two neural networks.  The first is a generator, which is used to generate new data (e.g., images, text).  The second is a discriminator, which is used to determine whether the data generated by the generator is real or fake.

In order to determine fakes, the discriminator is given a (preferably large) dataset of real data to compare the data generated by the generator against.  Like a high-tech game of cops and robbers, the discriminator is trying to identify the fake data provided to it from the generator.  Each time the generator fails to convince the discriminator of the validity of its generated data, it learns.  This process is repeated at the incredible rate that computing does these days.  The results, can be stunning.   In fact, a piece of artwork generated by such GANs recently sold at Christie’s for $432,000.

DABUS operates on this same principal.  Give a dataset to a discriminator as a training model and let the generator try to fool the discriminator.  Now, of course, Dr. Thaler uniquely calls his generator an Imagitron, and his discriminator a Perceptron, but ultimately it is still a GAN, doing what a GAN does.

What did DABUS “Invent”

Without getting into the debate or hype over what was actually “conceived” by DABUS, the actual products that were more aptly described as “designed” by DABUS are: 1) an interlocking food container; and 2) a light that flashes in a particular way that mimics neural activity.  I do not intend to go into the details of these inventions, as they have been discussed ad nauseum elsewhere.

Regardless of what DABUS designed, the inventions themselves were based on datasets and information provided to it.  As was said excellently elsewhere, “[DABUS] was ‘mentored’ by Dr. Thaler over a two month period to produce increasingly complex concepts.”  Put differently, DABUS did not set out on its own to solve the problem of interlocking containers or calming pulsating lights; rather, DABUS was tasked with solving these concerns.

The Patent Applications

The three patent applications were submitted by a group led by Ryan Abbott, a professor of law and health sciences at the University of Surrey in the United Kingdom.  Not unsurprisingly, a novel question of law has been raised by an academic.  And that is not a criticism, but rather an acknowledgement that patent law will have to address certain pressing questions about the patentability and ownership of inventions in a world that is increasingly being influenced by artificial intelligence and its effects.

The question at issue here is whether, given the “creative” output of DABUS qualifies it as an “inventor” under the laws of the patent offices the applications were filed in.  For instance, in the USPTO, the Manual for Patent Examining Procedure defines an inventor as a person who contributes to the conception of the invention.

Here, Dr. Abbott and his team are suggesting that DABUS, through its processing of the data into a useful invention, contributed to its conception in such a manner that it would have to be considered an inventor under that definition.  Others, such as Dr. Noam Shemtov, state that DABUS, and other AI systems are merely tools.  In a 2019 report commissioned by the EPO, Dr. Shemtov writes:

When it comes to a human actor that uses an AI system, whose identity may be inconsequential to the invention process, who simply uses a machine learning technique developed by another, the inventor may be the person who “tooled” the AI system in a particular way in order to generate the inventive output. Hence, under such circumstances the person that carries out the intelligent or creative conception of the invention may be the one who geared up the AI system towards producing the inventive output, taking decisions in relation to issues such as the choice of the algorithm employed, the selection of parameters and the design and choice of input data, even if the specific output was somewhat unpredictable[i].

While I tend to agree with Dr. Shemtov’s analysis of current AI systems as tools that individuals use to provide useful output, that is not to say that it is impossible to conceive that an AI system in the future will able to identify a problem, identify available materials/components/data, and solve the problem on its own, all without human intervention or direction.  And herein lies where Dr. Abbott’s filing of the patent applications with DABUS as an inventor is truly aimed at addressing – AIs as inventors.

AI Inventorship

Dr. Abbott’s true intention appears to be raising the debate over whether current statutory regimes at the various patent offices would allow for an AI to be listed as an inventor, given that even at present, it could be rationalized that the AI is part of the conception of the invention.  However, in the case of DABUS, it seems that there is a strong argument that the conception was at the hands of Dr. Thaler, and DABUS merely reduced the invention to practice.

Regardless, the issues related to having an AI system as an inventor are definitely worth bringing to light.  Dr. Abbott told BBC News, “These days, you commonly have AIs writing books and taking pictures – but if you don’t have a traditional author, you cannot get copyright protection in the US.”  And Dr. Abbott is not incorrect.  The United States Copyright Office published an opinion in 2014 stating that, “[O]nly works created by a human can be copyrighted under United States Law.”

Dr. Abbott continued, stating to BBC News, “So with patents, a patent office might say, ‘If you don’t have someone who traditionally meets human-inventorship criteria, there is nothing you can get a patent on.’ In which case, if AI is going to be how we’re inventing things in the future, the whole intellectual property system will fail to work.”  And again, Dr. Abbott is not incorrect.  35 U.S.C. § 100 defines an “inventor” as the “individual” who invented or discovered the subject matter of the invention.  If an AI system conceived of both the problem and the inventive solution to that problem, it becomes more difficult to say that the inventor of the AI system itself was the inventor of that invention, leaving that IP unable to be secured by patents under the current regime.

Similar inventorship issues arise in both the UK, where the UK Patents Act of 1977 requires an inventor to be a person, and the EPO, where in the EPO’s published opinion written by Dr. Shemtov states, “[I]t has been shown that is is unambiguously implicit that AI systems cannot be identified as inventors.”

All three patent offices where these DABUS patent applications have filed are aware of the issue and are reviewing options and requesting input from the community.  In fact, On August 27, 2019, the USPTO just released a Request for Comments on Patenting Artificial Intelligence on the Federal Register, asking for input on whether current US patent laws need to be revised to take into account inventions where an entity or entities other than a natural person contributed to the conception of an invention.

 Conclusion

While lofty news articles have hyped the filing of these patents in DABUS’s name, let us not fall for the smoke and mirrors associated with the current state of AI systems.  At least in the vast majority of existing AI systems, they are not inventors as we define them for the purpose of patents in the US.  They are advanced tools that assist in the reduction to practice of inventive concepts and inventions themselves.  DABUS is not an inventor for the purpose of US patent law.  However, using the leverage and momentum of a good news cycle, DABUS has been able to push forward the conversation on what will inevitably be questions about ownership and protection of all forms of Intellectual Property generated by AI systems in the future.

[i] Dr Noam Shemtov, “A study on inventorship in inventions involving AI activity” (2019)

Patenting and Protecting Artificial Intelligence in the United States

Advancements in Artificial Intelligence (AI) have been occurring at an ever-increasing rate, impacting almost every field of technology, from medical diagnosis and analysis, to driverless cars, to automated securities trading platforms, all the way to home security[i].  The arms of AI can be felt in every industry, in one way or another. Given the speed of advancement and the very nature of AI itself, it is important to consider the complex landscape around how to protect improvements in the AI space.

At first, it is important to note the types of intellectual property (IP) protection that can apply to inventions in the AI space.  The definition of IP generally comprises the core four – patents, copyrights, trademarks and trade secrets.  In the case of AI, each of these may apply, and each has its own particular usefulness and advantages.  Further, each of these types of IP has its own concerns with respect to the timing of obtaining the protection.  For this article, we will primarily be focusing on the two areas of protection that generally are of the most concern – patents and copyrights.

shutterstock_1096541144.jpg

Patenting AI

AI, in almost all cases, exists as a software component.  Depending on what the use case for the AI is, there may also be a hardware component (e.g., vision system, sensors, actuators), but these hardware components are generally peripheral to the core AI, which is software based.  While AI inventions based in combined software and hardware solutions or standalone software-based solutions may both be patentable, the analysis for whether an AI based system constitutes patent eligible subject matter does take into consideration what is actually involved.

Where there is a corresponding hardware component in use with the software-based AI component, the subject matter eligibility analysis is usually relatively simple, favoring eligibility over not.  For instance, if an AI system is used to automatically control a series of vision systems (e.g., security cameras) and detect intruders, the invention is likely eligible for patent protection from a subject matter perspective, assuming the application is drafted appropriately.

Where the AI system solely exists as a software solution, an examiner at the USPTO will likely give the invention more scrutiny under the subject matter eligibility tests.  We have written more on the patentability of software-based inventions in a separate article that you can find here. However, while patent applications directed to software only inventions may receive additional scrutiny, true AI inventions likely are sufficient to overcome these rejections.  In fact, USPTO Director Andrei Iancu has even discussed the patentability of AI, through discussions of “[H]uman-made algorithms”, during a hearing regarding the oversight of the USPTO in April of 2018.   Again, the critical point in securing a patent in a solely software based AI invention is appropriate drafting of the application.

With respect to patenting AI based inventions, particularly as it relates to the software component of the AI, it is important to note that a utility patent covers the functionality of the software, through defining the software in terms of systems and methods.  What patents do not cover is the actual code.  Source code is covered largely by copyright, which can protect direct copying of the code, but not those who write their own code to perform the same functionality.

In defining what an inventor wants to protect with respect to their software-based AI invention, it is important to look at the invention in terms of a method, or a series of steps.  Considering everything a computer does is generally a series of steps involving processing some data, framing the inventive aspects of the software-based AI invention in such a methodological manner is generally straightforward.

What inventors want to avoid is viewing the invention in the abstract, or very high-level depiction.  For instance, you cannot get a patent on the idea of “an AI based dating platform”, but you could potentially get a patent on the methods performed by the AI in order to find compatible matches (e.g., based on training models and predictive analytics).  So, a focus needs to be on what actual occurs in order to make the invention possible, not solely focusing on a conclusory statement about what problem is being solved.

Another important thing to remember when seeking patent protection for AI inventions, or any invention, is to do so sooner rather than later.  There are two main drivers for this.  First, the USPTO, and most if not all other national patent offices are “first-to-file” for priority on inventions.  What this means is that, even if you get to the market first with your invention, or conceived of the idea before another inventor, if another party’s application gets to the patent office before yours, then the patent rights will be theirs, and you will be prevented from getting a patent on the invention.

The second reason is that your ability to get a patent on an invention, even without worrying about what others are doing, can be jeopardized if you offer for sale or otherwise disclose your invention publicly before filing.  The USPTO gives you one year from making a public disclosure of the invention to file your patent application.  However, the rest of the world is not so nice, with many jurisdictions making it a bar on patentability if you publicly disclose your invention prior to filing a patent application in at least one jurisdiction first.

shutterstock_663863257.jpg

Copyrighting AI

With respect to copyrighting all or portions of an invention based in AI, there are certain aspects of these inventions which are protectable and those which are not.  Copyrights cover artistic works, which includes everything from literary works, to graphical/visual works (e.g., paintings, movies, photographs), to musical works and even choreographed dances.

When considering copyrighting portions of an AI based invention, the focus is generally on copyrighting the source code.  Source code is considered a literary work for the purposes of copyrights, and inventors can receive a federal copyright registration in the uncompiled source code.

What is protected by a copyright registration on source code is the copying of the actual code.  It does not prevent others from creating code of their own that performs the same functions.  However, it does restrict the copying of subsets of the whole code, such as the copying of a module, or a series of functions.

One issue when considering copyrighting source code is how frequently the source code is updated.  Rarely is there a piece of software that is static for very long.   Updates in source code, while they may be considered derivative works of the originally copyrighted code, may not be independently covered by the initial registration.  Inventors should consider at what point they want to secure additional copyrights on later versions of a software-based invention.

It is important to note that while a copyright registration can be done at any time, as the works form in the author upon creation, statutory damages and attorneys fees are generally only available if the copyright registration is filed within 3-months of publication of the work[ii].  Filing your registration after that point will limit damages to “actual damages” (e.g., lost profit), which can be harder to prove.

Separately, more and more we see the question about whether it is possible to copyright the output of AI.  Recently there have been numerous instances of AI generating their own artistic works, such as The Next Rembrandt and Bayou.  The law is currently unsettled as to whether these works would be copyrightable.  For instance, in April of 2018, the Court of Appeals for the Ninth Circuit held that the Copyright Act only provides standing to humans[iii].  The case, involving copyrights associated with a Monkey Selfie, but the same findings would presumably extend to works authored by AI.

Of course, numerous scholarly and legal minds believe that works would be derivative works of the individuals who wrote the code for the AI, and as such those individuals would be the rightful owners of works generated by the AI.  We ultimately will have to wait to see how this plays out in the future.

 

Conclusion

Overall, it is important to understand and analyze what aspects of an AI based invention can be secured early on in the process.  Timing is crucial for both patents and copyrights with respect to being able to secure the rights and receiving the greatest protection available under the laws.  This area of technology is moving quickly, so delay and lack of planning can be devastating. Devoting at least some time to do the analysis may help with providing a roadmap for how and when to protect various aspects of your AI based invention so that you reap the greatest rewards possible.

[i]Check out our client Deep Sentinel: https://www.deepsentinel.com/

[ii] See, 17 U.S.C., 412 https://www.law.cornell.edu/uscode/text/17/412

[iii] Naruto v. Slater http://cdn.ca9.uscourts.gov/datastore/opinions/2018/04/23/16-15469.pdf