Your AI overlords demand to be recognized...

Oh, inventorship and AI. As much as Crispr has been preoccupying the IP agendas of IP conferences the last couple of years, inventorship and AI looks to be taking its place. The IPKat has also been a fan of the topic (see Kat posts here). For the AmeriKat's part, she thinks that, for the time-being, AI inventorship is (for the most part) a non-event, but not everyone agrees.  Indeed some issue warnings about not recognizing the role of AI in inventorship.

"Last week the European Patent Office stated that an inventor must be a ‘natural person’. This provides the basis for the EPO’s decision in December to reject two patent applications which had named the artificially intelligent system Dabus as the sole inventor. The patent applications were made by Dr Thaler’s team at Surrey University and described a flashing beacon light and plastic food container.

Dr Thaler’s failed ‘AI inventor’ applications have raised an important question - can AI be inventive? However, in this article, I shall not dwell on the specific merits of Dr Thaler’s applications but instead focus on Google DeepMind’s artificially intelligent system AlphaGo.

In 2016, AlphaGo was capable of reading and understanding the prior art of the game Go, finding novel and inventive solutions to defeat its opponent Lee Sedol (the world champion) 4 games to 5 and effectively communicating that solution to Aja Haung, the DeepMind team member who placed stones on AlphaGo’s behalf. However, the technology behind AlphaGo is not limited to board games, it also extends to sectors with patentable subject matter such as telecommunications and pharmaceuticals.

This article argues that AI can be inventive, that AI differs from other platform technologies and for the sake of transparency, AI inventorship should be recognised.

AlphaGo

In 1997, Garry Kasparov became the first chess world champion to lose a match to a computer which had been programmed by expert chess players. In 2016, Lee Sedol became the first Go world champion to lose to a computer that had taught itself - AlphaGo.

Prior to Sedol’s defeat it was thought that humans would retain the upper hand in Go because in any one position there are 200 possible moves, whilst in chess there are 20. To put this into perspective the number of possible configurations of a Go board is greater than the number of atoms in the universe. In fact, even if you took all the computers in the world and ran them for a million years, this would still not be enough computing power to calculate all the possible variations. Therefore, AlphaGo cannot simply compute every possible move, it must rely on something else – intuition.

In the second game AlphaGo made Move 37 and gave us a glimpse of a future shaped by computer intuition and inventiveness. However, Move 37 appeared at first to be a mistake because no competent human player would have made it. In fact, AlphaGo knew that there was only a 1 in 10,000 chance a human would have made Move 37, yet despite this, it ignored the games received wisdom or prior art and made an inventive move.

Post-match analysis concluded that Move 37 was the pivot on which the game had turned in AlphaGo’s favour. Lee Sedol later described it as a “really creative and beautiful” move. The lessons that AlphaGo has taught and continues to teach human Go players amounts to an improvement in the game’s prior art and if Move 37 were patentable, it would be an inventive step.

How AI differs from other platform technologies e.g. mouse models

Aja Haung was the DeepMind team member who placed stones on the Go board for AlphaGo in its victory over Lee Sedol. However, it is unreasonable to suggest that Haung was the creative mind behind Move 37 because at no point did he exercise choice. Move 37 would have occurred irrespective of whether it was Haung or this author placing the stones.

Mouse models on the other hand still require scientists to exercise choice either when selecting the antigenic challenge or in determining which antibody therapeutics may be suitable for use in humans. The scientist in making these choices must rely on their experience and intelligence, which is why the scientist rather than the mouse is the named inventor. A useful antibody therapeutic would not exist irrespective of whether a scientist skilled in the art or this author were conducting the experiment.

I have used AlphaGo as an example because the human operator is an automaton but mouse models do pose an interesting problem; namely making a choice based on a menu of limited options does not appear particularly inventive. Is there a difference between a mouse model and an artificially intelligent system that still requires the human operator to exercise choice?

The difference is that the immune system of a mouse is static, in that it does not get better over time at identifying antibody therapeutics for use in humans. Artificially intelligent systems on the other hand are dynamic in that over time they do get better at their given task. For example, DeepMind went onto develop AlphaGo Zero in 2017, which with no prior knowledge of the game Go and only the basic rules as an input had at three days surpassed the abilities of AlphaGo Lee, the version discussed above. It had achieved this entirely from self-play, without human intervention or historical data. In just a few weeks AlphaGo Zero had accumulated thousands of years of human knowledge but it had also discovered new knowledge. It developed unconventional strategies and creative new moves that echoed and surpassed Move 37. Mouse models will always require an intelligent human scientist but artificially intelligent systems will over time make their human operators increasingly redundant. How will we know when humans have become redundant in the inventive process if an AI cannot be named as an inventor?

Furthermore, Google recently announced that its Sycamore quantum processor had performed a specific task in 200 seconds that would have taken the world’s best supercomputer 10,000 years to complete. One of the major applications of quantum computing is in artificial intelligence and if successfully applied will make AlphaGo Zero appear prehistoric in comparison.

The dual purpose of patents - ownership and transparency

In return for the grant of a twenty-year monopoly society demands transparency from the owners of patents. But in denying the existence of ‘AI inventors’ society risks losing this transparency because it encourages human inventors to either down play the extent of AI assistance in a patent application or alternatively to rely on trade secrets. In either scenario transparency is lost but does this matter?

The creative outputs of both Aja Haung and AlphaGo are owned by Google but that ownership should not make it irrelevant which of the two’s creativity was responsible for defeating Lee Sedol. Without the legal fiction of AI inventorship we are headed for a society in which Aja Haung became the Go world champion in 2016. "

The EPO generally likes applicants to abide by the Articles and Rules of the EPC. It is therefore not surprising that the Receiving Section of the EPO has refused the recent European patent applications designating an AI inventor for contravening the provisions on inventor designation (Article 81 and Rule 19 EP).  The refusal followed a mere 16 minutes of deliberation at an EPO hearing in Munich.

IPKat reported on the AI inventor patent applications earlier this year (IPKat: The first AI inventor - IPKat searches for the facts behind the hype). Dismissed as a publicity stunt by some, the team behind the European, UK and US applications have defended the applications as an attempt to address the moral issues arising from inventions for which no human inventor can be ascribed. However, it is clear from the Minutes of Oral Proceedings (and preliminary opinion), that the EPO was not convinced that the human applicant (Dr Thaler, inventor of the AI) had satisfactorily demonstrated how he had derived the right to the applications from the AI. Important questions also remain as to whether the AI can truly be said to have invented, and whether the whole debate is decades too premature.

Ownership is what matters in European patent law

One crucial issue of AI inventorship is how the applicant can be said to derive the right to a patent from the AI. How can an AI assign such a right? The Receiving Section stressed that the EPO does not investigate whether the designation of an inventor is correct (Rule 19(2) EPC). However, an applicant for a patent who is not the inventor is required to state how they derived the right to the invention from the inventor (Article 81 EPC). The team behind the AI inventor applications have always maintained that the applications are owned by the inventor of the AI, Dr Thaler, and not the AI itself. But how does Dr Thaler derive this right?

Dr Thaler, initially indicated that he derived the rights to the invention from the AI inventor as the AI's employer (a claim that raises certain philosophical questions in and of itself). The EPO did not accept this as a valid derivation of right, and Dr Thaler tried again by claiming that he was the successor in title of the AI as the AI's owner. However, as the Receiving Section noted in the preliminary opinion: "machines do not have legal personality and cannot own property...a machine cannot own rights to an invention and cannot transfer them within a employment relationship (as proposed by the applicant on 24.07.2019) or by succession (as suggested by the applicant on 02.08.2019)".

Chappie: Artificial General Intelligence employee (and most underrated movie of 2015)

It seems that the EPO can therefore neatly dismiss any need for detailed discussions on the morality of AI inventorship, so long as the AI inventor is not also claimed to be the owner of its creations. It has always been this Kat's view that the whole exercise of the AI inventor applications is slightly irrelevant given that what primarily matters from a legal standpoint is who owns an invention, and who should therefore be awarded the patent.
The written decision is expected early in the new year. IPKat has learned from Professor Ryan Abbott (University of Surrey and key member of the "AI inventor" team) that Dr Thaler plans on appealing the decision.

Has the AI actually invented?

As already mentioned, the EPO does not investigate whether the inventors named on a patent application really are the inventors (Rule 19(2) EPC). The question of whether the AI really did invent the fractal food container and fractal light signalling inventions that are the subject of the applications is therefore not one that the EPO will attempt to answer. However, it is worth remembering that the team's claims about the AI inventor are extraordinary.

An AI that is truly capable of inventing will be capable of reading and understanding the prior art, finding a problem to be solved, finding a novel and inventive solution to the problem and then communicating that solution in a way that is understandable to a skilled person. An AI capable of such a complex collection of tasks would have a close to a human level of intellect, i.e. an artificial general intelligence (AGI). An artificial general intelligence is a 20 year goal of companies like Google DeepMind and Google Brain. Dr Thaler has, in fact, argued that his algorithm "paves the way for sentient AI since it teaches how machines may generate the equivalent of subjective feelings...It is expected to be the successor to deep learning and the key to achieving human level machine intelligence. It will be used to build highly transparent and self-explanatory synthetic brains to achieve so-called “Artificial General Intelligence” (AGI)" [Merpel: someone needs to tell Demis Hassabis...]. A summary of the incredible abilities of Dr Thaler's AI can be read on his website.

So far, however, the precise details of how the AI performs the inventive act have not been published and remain distinctly vague. It seems likely (at least to this Kat) that Dr Thaler's AI may have more limited abilities than are enthusiastically propounded by the team. This raises the question of whether the AI is, in fact, no more than a tool whose abilities can be equated to other types of technology.

At the recent Life Science Patent Network (LSPN) London conference, this Kat had the pleasure of participating in a panel discussion with Professor Ryan Abbott on the topic of AI inventorship. Professor Ryan Abbott made a plea for the importance of the moral and social issues surrounding AI inventorship. However the Professor did not seem to address the question of how the AI could be said to be different to other types of platform technology, e.g. mouse models.

The central claim of the AI inventor team is that there was no human input in the generation of the invention: "the machine only received training in general knowledge in the field and proceeded to independently conceive of the invention and to identify it as novel and salient". However, the same could be said of a mouse model used to identify novel, inventive and useful antibody therapeutics. The human experimenter provides the antigenic challenge, and the mouse's immune system produces structurally unique, non-obvious therapeutics that are potential suitable for use in humans. The "invention" of the antibody structure could not have been derived by the human experimenter themselves. However, no one is arguing that mice should be inventors.

Are the AI's inventions even novel?

As a side issue, a perusal of the file reveals that the AI inventor may also not be as good at inventing as had been first thought. The two applications claimed a fractal food container (EP 18275163) and fractal light signals (EP 18275174). The team behind the AI have been previously eager to point out that a search by the UK IPO found the claims of the UK application to be novel and inventive. The EPO did not agree with the UK IPO. Both the Food Container and Light Signals invention, as originally claimed, were found to lack both novelty and inventive step (European search opinion). The Examiner cited, for example, US 5803301 as disclosing a beverage container having all of the features of the one allegedly invented by the AI.

Prior to the EPO's decision to refuse the applications because of the inventorship issues, the applicant had submitted amendments and arguments in response to the substantial objections. However, the issue of novelty and inventorship have become redundant in view of the formality issues with the applications.

Final thoughts

If Dr Thaler's appeal of the Receiving Office decision does indeed go ahead, IPKat will await the Statement of Grounds with interest (deadline for filing the appeal will be early 2020). Based on their submissions to the Receiving Office, it seems that the AI inventor team are lacking the legal arguments to overturn the decision. Moral and social arguments are unlikely to convince the Boards of Appeal. A request for a referral to the Enlarged Board can probably also be expected, but is similarly unlikely to be granted. None-the-less, the team have undoubtedly been successful in one goal, bringing attention to Dr Thaler and the team. Notably, we are only able to see the file history because early publication of the applications was requested. However, in this Kat's humble view, the whole argument surrounding AI inventorship is premature until the existence of an AI truly capable of a inventive act has been proved.

 
The AI inventor, named “DABUS” by its creator Stephen Thaler, relies upon a system of many neural networks generating new ideas by altering their interconnections. A second system of neural networks detects critical consequences of these potential ideas and reinforces them based upon predicted novelty and salience.

"what are u up to inventor-robot"?

 “Powerful AI systems could hold the key to some of the mega challenges facing humanity – from the cure for cancer to workable solutions for reversing climate change. But if outdated IP laws around the world don’t respond quickly to the rise of the inventive machine, the lack of incentive for AI developers could stand in the way of a new era of spectacular human endeavor.”