In my previous article, I argued that existing (and foreseeable) artificial intelligence (AI) or machine learning (ML) systems do not exhibit creativity or inventiveness, and are incapable of anything that could reasonably be described as ‘invention’. I acknowledged, however, that some such systems have generated results that may qualify as patentable inventions. I therefore concluded with a question: if computers cannot invent, and yet the outcome of running a computer program can be an invention, then who – if anyone – is the inventor? In addressing this question, it is important to understand that ML systems do not autonomously or independently generate novel outputs. In my view, this is a fundamental error of understanding in Professor Ryan Abbott’s paper ‘I Think, Therefore I Invent: Creative Computers and the Future of Patent Law’, Boston College Law Review, Vol. 57, No. 4, 2016 (also available at SSRN), which I discussed in the first article of this series. Abbott contends, in particular, that ‘machines have been autonomously generating patentable results for at least twenty years and that the pace of such invention is likely increasing.’
This, I have argued, is simply wrong. The difference between a computer that is programmed to play the board game Go, and one that is programmed to learn to play Go is, of course, significant. The former can only make moves that are determined in accordance with its explicit programming, whereas the latter may appear to ‘invent’ new strategies in response to patterns occurring in its training data that have not previously been recognised by human players. But the appearance of invention is not the same thing as actual invention. The ML player is still doing nothing more than following the instructions devised by its programmers. The Google DeepMind AlphaGo system has become the world’s best Go player as a result of years of development, trial, experiment, and experience on the part of its designers. AlphaGo plays as well as it does simply, and only, because that is what it was designed to do. In this sense it is no more ‘autonomous’ than any other computer program.
In this second article I will explain why I believe that in the case of all existing (and currently foreseeable) ML systems which may generate inventions as output, there is always a human inventor. This is consistent with the history and current state of patent law, as well as with the practical and technical reality of ML systems.
The Human Mind is the Source of Patentable Inventions
While there are variations in the patent laws of different countries, it is axiomatic that any patentable invention is ultimately derived from conception within one or more human minds. No patent application can be accepted, or patent granted, anywhere in the world, without the existence of at least one human inventor. There is no scope for non-human inventorship of any kind. Article 4.A.(1) of the Paris Convention for the Protection of Industrial Property (first adopted in 1883, and to which 177 countries are now party) states that ‘[a]ny person who has duly filed an application for a patent … in one of the countries of the Union, or his successor in title, shall enjoy, for the purpose of filing in the other countries, a right of priority…’ (emphasis added). Article 4ter further provides that ‘[t]he inventor shall have the right to be mentioned as such in the patent.’ It is inherent in these provisions that inventors are people, and that non-human parties, such as corporations, acquire rights in relation to patents by legal succession, e.g. employment or assignment.In the United States, 35 USC 101 states that ‘[w]hoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor…’ – a provision that is ultimately authorised by Article I, Clause 8, Section 8 of the US Constitution, which grants Congress the power ‘[t]o promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries.’ Article 60(1) of the European Patent Convention provides that ‘[t]he right to a European patent shall belong to the inventor or his successor in title.’ Section 15 of the Australian Patents Act 1990 stipulates that a patent may only be granted to ‘a person who … is the inventor’, or persons (natural or legal) who legitimately derive the right from the inventor.
The potential for non-human authorship of copyright works has occasionally been raised in the US. In the famous ‘monkey-selfie’ case, People for the Ethical Treatment of Animals (PETA) argued that a male crested macaque named Naruto should be recognised as the author of photographs he took of himself using a camera set up by photographer David Slater. However, a US Federal Court judge found that ‘the Copyright Act does not “plainly” extend the concept of authorship or statutory standing to animals. … The Supreme Court and Ninth Circuit have repeatedly referred to “persons” or “human beings” when analyzing authorship under the Act.’
With regard to machine authorship, a 1986 report by the US Congress Office of Technology Assessment (OTA) [PDF, 4.1MB] raised the question of ‘whether machines or interactions with machines might produce a pattern of output that would be considered creative or original if done by a human being’ and noted that ‘if machines are in any sense co-creators, the rights of programmers and users of programs may not be easily determined within the present copyright system’ (page 72). However, in over 30 years since that report was published, ‘machine authorship’ has failed to emerge as a serious legal concern.
Thus there remains strong historical precedent, and widespread contemporary agreement, that machines cannot be ‘creative’ or ‘inventive’ in the way that humans can, and that the human mind is uniquely the origin of subject matter that may be protected under intellectual property laws. There is no question, however, that computers and software can be – and, increasingly often, are – used as tools to assist people in these processes. In everything from word processing software, to versatile electronic musical instruments and digital recording systems, to sophisticated computer-aided design and engineering packages, computers now assist authors, artists, designers, inventors, scientists, and engineers in their creative work. Computer programming is itself a creative activity, and computer programs are accordingly protected by copyright (whether or not the author wishes it to be so) and, in appropriate cases, inventive applications of computers and software can be protected by patents.
The Legal Concepts of Invention and Inventor
The patent law will only recognise humans – and not machines – as inventors, and for every patent issued there must be at least one inventor. However, the law does not necessarily require that a named human inventor actually engage in any activity that the public-at-large would universally recognise as ‘inventive’, in a colloquial sense of the word. More specifically, although we might think of an invention arising as a result of diligent intellectual effort and/or a flash of inspiration in the mind of an inventor, the legal concept of an ‘inventor’, for the purposes of patent law, is not tied to the legal concept of an ‘invention’ in this way.Setting aside, for a moment, the requirement for an inventor, the legal definition of a patentable invention includes (among other requirements) that there be an inventive step or, equivalently, that the invention not be obvious to a person having ordinary skill in the art (i.e. field of technology or endeavour) to which it pertains. Notably, something may be non-obvious, according to the legal definition, without it necessarily having been ‘invented’, in the colloquial sense of the term. The US patent law is particularly explicit about this. I have already quoted 35 USC 101, which states that a patent may be granted to ‘[w]hoever invents or discovers…’, while the non-obviousness provision, 35 USC 103, expressly says that ‘[p]atentability shall not be negated by the manner in which the invention was made.’ In other words, it makes no difference whether an invention comes about as a result of years of hard work, an instantaneous flash of genius, pure dumb luck, or the use of a computer to search an immense space of possible solutions to a problem. If the result satisfies the legal requirement for non-obviousness (along with other conditions of patentability) then it can be the subject of a patent.
In this context, an inventor is simply a person – or a team of people, in the case of joint inventorship – who independently recognises the significance and usefulness of the invention, regardless of the manner in which it was made.
If you know a bit about patent law, you might object at this point that the above statement seems to be at odds with what the courts in various countries have said about inventorship. In the United States, for example, conception has been called ‘the touchstone of inventorship, the completion of the mental part of invention’ (Burroughs Wellcome Co. v Barr Laboratories, Inc. 40 F. 3d 1223, Fed. Cir. 1994, at 1227-8). Conception ‘consists in the complete performance of the mental part of the inventive act’ and is ‘the formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention as it is thereafter to be applied in practice’ (Townsend v Smith 36 F. 2d 292, CCPA 1929, at 295). Conception is to be distinguished from reduction to practice, which is the process by which a fully-conceived invention may be put into a effect by a person having ordinary skill in the art ‘without unduly extensive research or experimentation’ (Sewall v Walters 21 F. 3d 411, Fed. Cir. 1994, at 415).
While this notion of ‘conception’ may create an impression that something more than discovery or recognition is required, that would be an incorrect interpretation. Indeed, Townsend itself was a case in which the invention came about by accident, as a result of a gear in a machine for cutting screw threads having been manufactured with the wrong number of teeth. Consequently, the invention was, in effect, reduced to practice before it was conceived, i.e. recognised as a potentially useful intentional configuration, by Townsend.
It is also important to keep in mind that where the courts have been required to consider processes of invention, it has most commonly been in instances where inventorship was in dispute. In Australian (and earlier UK) case law, for example, questions of joint inventorship have often arisen. In such cases, the courts have sought firstly to identify the relevant ‘inventive concept’ and then to assess whether each alleged co-inventor ‘made contributions that had a material effect on the inventive concept’, which involves an inquiry ‘as to a person’s contribution to the conception of the invention in order to determine whether there is co-inventorship’ (Kafataris v Davis [2016] FCAFC 134, at [62], [65]; see also University of Western Australia v Gray [2009] FCAFC 116, Polwood Pty Ltd v Foxworth Pty Ltd [2008] FCAFC 9, at [60]-[61]). Generally, where co-inventorship is disputed there has been some contribution by each alleged inventor, and the task of the court is to determine which of those contributions are material to the invention. It is not surprising that the process of invention in such cases involves more in the nature of human exertion (intellectual and/or physical) than in recognition or discovery alone. This is not to say, however, that this is the only way in which invention occurs.
An Alleged Machine Invention
With the above discussion of inventorship in mind, I will now take a closer look at one of the examples of supposed ‘computational invention’ identified in Abbott’s paper. As I discussed in my previous article, Abbott contends that the USPTO granted US patent no. 6,847,851 on an invention that was ‘created by the “Invention Machine”— the moniker for a GP [i.e. Genetic Programming]-based AI developed by John Koza.’ Abbott’s basis for this assertion is essentially that Koza says so, notwithstanding that the patent actually names two other co-inventors along with John R Koza, neither of whom appears to be a computer.The invention disclosed in the patent relates to improvements in proportional-integral-derivative (PID) controllers. PID controllers are extremely common and well-understood systems, used to automatically apply control signals to industrial and other systems requiring continuous adjustment in order to maintain some target ‘output’ value. A common example – mentioned in both Koza’s patent and the Wikipedia article on PID controllers – is automobile cruise control. The principles of PID controller analysis and design have been understood since the 1920’s, and are routinely taught in undergraduate engineering courses… or, at least, they were when I was a second-year electrical engineering student in 1987.
A basic PID controller has three main parameters, which can be thought of as numbers that are used to multiply the values of signals within the controller. The ‘trick’ of controller design is generally to choose these values so as to optimise performance. Usually what we mean by this is that we want the controller to cause the system to move as rapidly as practicable to the set output (think about changing the speed setting on your cruise control), without overshooting, and then to remain stable in the presence of external disturbances that may result in changes to the output (think about what needs to happen to maintain speed when your car starts to climb a hill).
Presumably, what Koza and his co-workers did was to develop a GP system to ‘evolve’ values of the PID controller parameters, using a measure of ‘fitness’ based upon the performance characteristics noted above. Such a system does only one thing – it generates equations, based upon predetermined system variables, for computing parameters that are then ‘plugged in’ to a PID controller having a particular architecture. Optimising these parameters for a particular application is sometimes called ‘tuning’ the controller. As the Wikipedia article on PID controllers explains:
PID tuning is a difficult problem, even though there are only three parameters and in principle is simple to describe, because it must satisfy complex criteria within the limitations of PID control. There are accordingly various methods for loop tuning, and more sophisticated techniques are the subject of patents….
A Narrow Machine-Generated Claim
Koza’s approach to this difficult problem was essentially to throw a whole lot of computing power, and some GP techniques, at it. Claim 1 of the Koza patent exemplifies the result of such a process:
1. A proportional, integrative, and derivative (PID) controller comprising a proportional element, an integrative element, and a derivative element coupled together and responsive to a reference signal to generate a control signal in response thereto to cause a plant to generate a plant output, wherein the proportional element has a gain element with a gain being substantially equal to
where Ku is the ultimate gain of the plant and Tu is the ultimate period of the plant.
In case it is not apparent to readers without an engineering background, everything in that claim except for the equation is entirely conventional. The ‘invention’ is the equation. What is more, it is an equation that includes a number of seemingly arbitrary constants, two of which have been quoted to ten significant figures. Now, I am not sure how much ‘wriggle room’ a court would ascribe to ‘substantially equal’, but I am guessing that when ten figures are quoted, not very much! This is, in short, an incredibly narrow claim.
There are similarly specific equations for computing other parameters of the ‘inventive’ PID controller in other claims of the patent. None of these equations is obvious, in the sense that they do not correspond to any established standard formulation of the parameters of a PID controller, and no control engineer would naturally (without computer assistance) pursue such complex forms.
And yet the computer has not done anything we would regard as ‘inventive’. It has simply followed its programming, albeit including some random elements and with no set notion of the exact form of the output, to find equations that, out of all of its trials, most closely achieve a predetermined optimum controller performance. The real contribution made by Koza and his co-inventors was in programming the computer to execute their chosen genetic algorithms such that the equations would evolve in the general direction of better solutions.
Conclusion – Humans Invent, Machines Assist
The process followed by Koza – and others claiming to have developed inventive machines – is most appropriately described as ‘machine-assisted inventing’. It does not differ from more traditional machine assistance, such as through the use of computer aided design, engineering and simulation tools, in any way that I would regard as changing the fundamental nature of the contribution made by the machine. Just as with more conventional software, machine learning systems still require an operator or programmer who understands the nature of the problem at hand, and can define the parameters of the required solution. With suitable choices of algorithms, inputs, and performance evaluation functions, a machine learning system will produce a solution. That solution might be new, it might be an improvement over existing solutions, and it might be inventive – in the patent law sense of being non-obvious to a person of ordinary skill in the art.However, the machine is completely indifferent to these characteristics of its output. All it ‘knows’ – i.e. all it has been programmed to do – is to improve or optimise its performance in a specific task relative to some objective function. Whether a solution found by a machine learning system is actually new and/or superior to existing approaches is something that can only be appreciated, in a meaningful sense, by a human programmer or operator who is, therefore, the inventor in accordance with existing and long-established principles of patent law.
If a machine gives the impression of ‘inventing’ – by producing an invention as output – it is only because that is what it was programmed to do. The professional and commercial motivation for people such as Dr Koza to characterise their software as ‘inventive’ is understandable. However, the motivation for others, such as Professor Abbott, to accept such claims at face value and to ascribe a degree of agency and inventive capacity to machine learning systems is less clear. It could, however, be as simple as a lack of technical understanding of how these systems work, what they do, and – just as importantly – what their limitations are. At one point, Abbott’s paper states (page 1095):
If a computer scientist creates an AI to autonomously develop useful information and the AI creates a patentable result in an area not foreseen by the inventor, there would be no reason for the scientist to qualify as an inventor on the AI’s result.
Even assuming this to be a valid statement, such an ‘AI’ does not exist, nor is it reasonably foreseeable based on the current state of technology. Machine learning systems do not produce such unforeseen results. AlphaGo is not going to spontaneously take up Scrabble. Koza’s PID controller optimiser will not design a vehicle transmission system. IBM’s Watson is incapable of suddenly deciding to write a Broadway musical. Truly autonomous and independent ‘AI’ of the kind here envisaged by Abbott is not the stuff of science, but of science fiction.Over two articles, I have argued that there can be inventions generated by machine learning systems without the need to invoke computer inventors, and that established patent law is perfectly capable of recognising the relevant human inventors in such cases. This does not mean, however, that there may not be cases in which inventorship is unclear cases of machine-assisted invention. I will address the issue of identifying (human) inventors in the third, and final, article of this series.
0 comments:
Post a Comment