Whether artificial intelligence software is patentable gets to the heart of what an invention is.

AI was one of the hottest buzzwords at the recent RSA Conference in San Francisco. Every second vendor claimed to offer AI as part of its solution. More generally, AI is playing a central role in the value of an ever-increasing number of companies throughout the technology sector.

At the same time, there is growing concern about whether it is or should be patentable. Much of this concern rests on the multiple ways that the concept of algorithm is being used.

A recent hearing of the US Senate Judiciary Committee addressed these issues. Senator Kamala Harris (of California) pointed out that artificial intelligence has been applied to a number of important problems, particularly in medical research of everything from cancer to autism. It is economically important and the US should be incentivizing its development. She was concerned that some recent court cases, including the Alice and Mayo cases raised uncertainty about whether the algorithms, which were the basis of artificial intelligence could receive patent protection.

The director of the US Patent and Trademark Office, Andrei Iancu, responded to Senator Harris that software patentability and artificial intelligence were important issues in the context of economic development and international competitiveness. He noted that the rules that govern what is patentable, 35 U.S. Code § 101, have not been amended at all since 1952 and were essentially the same as they were in 1793. He said that new rules might be necessary to provide the proper incentives for continued progress in this area.

Senator Harris asked (with a knowing smile) whether something like Einstein’s familiar E = MC2 formula could have been patented. Iancu said that it could not because it was the kind of algorithm that reflected phenomena or laws of nature and these were not patentable.

Einstein’s formula is a mathematical representations of the laws of nature. But other kinds of algorithms that were the product of human invention should be patentable. This kind of algorithm, he said, was very different from a law of nature and from mathematical objects like E = MC2.

Part of the disagreement between Senator Harris and Director Iancu, on the one hand, and the court, on the other, is ambiguity concerning the word ‘algorithm.’ The word ‘algorithm”’ has a formal definition, which we will get to in a minute. But the word is also used colloquially to refer to any computer process. This ambiguity leads to confusion.

An algorithm is formally a step-by-step procedure for accomplishing some task. There is an algorithm, the sieve of Eratosthenes, for finding prime numbers, there is an algorithm for finding the square root of a number, and an algorithm for multiplying numbers expressed as Roman numerals. The equation XVII x XXXII = DCCXIV is not an algorithm nor is the equation 17 x 42 = 714. Rather, we need an algorithm to solve equations like these.

Here is an algorithm for multiplying two numbers expressed as Roman numerals:

  • Make a table with two columns, with one of the numbers to be multiplied at the head of each column.
  • Divide the number in the first column by 2 (ignore any remainder), multiply the number in the second column by 2. Write the results in the corresponding columns of the next row.
  • Repeat until no more divisions are possible.
  • Cross out the rows where the left number is even.
  • Add the numbers in the right-hand column that were not crossed out.
  • That sum is the result.

Someone had to invent algorithms like these. They were not merely discovered.
A computer process is not typically an algorithm. You don’t need a computer to execute most algorithms.

The 19th Century view is that algorithms are not patentable because they are discovered and not invented. Alternatively, they are not patentable because they don’t DO anything, they merely provide instructions. Some countries (such as India) argue that computer programs can be written down, just as the algorithm for multiplying Roman numerals can be written down. But program listings are merely descriptions of the computer program, they are not the program itself. Equations are ideas, algorithms are method descriptions, but computer programs actually do something. Computer programs may make use of many algorithms, but they are themselves, not algorithms in the formal sense.

Artificial Intelligence is not an algorithm, though it may use algorithms. For example, stochastic gradient descent is a common algorithm for finding the solution to an optimization problem that is widely used in artificial intelligence and machine learning. It is invented (I don’t believe that it is merely discovered), but it is an algorithm in the formal sense.

Artificial intelligence, as practiced today, requires on a representation of the problem, an evaluation method, and an optimization method (stochastic gradient descent is one optimization method). The computer implementing this process operates on the representation using an optimization method in order to reach the goal condition specified by the evaluation method. The invention in artificial intelligence comes mostly from finding a clever way to represent the problem and to a lesser degree from finding clever ways to focus the optimization.

Computers can now play the game ‘go’ at a high level because the folks at DeepMind came up with clever ways to guide the search through a tree of move choices. Without their insight, the tree is simply too complex to be analyzed fully in a reasonable amount of time. The idea of representing the game in terms of a tree is an old invention. The DeepMind designers were also clever in setting up competitive systems to play against each other. Finally, they mobilized a huge number of computers to implement their system.

Competitive learning, having one computer play against another, improving both, may be a significant invention, but it has been public for too long to be patentable. Representing a problem as a tree may have once been a significant invention, but it may also have been too abstract an idea. Heuristics for limiting tree search may be significant inventions.

These heuristics may or may not be patentable by themselves. None of these is inventions is an algorithm, nor is their combination. On this analysis, a go-playing computer system might be patentable because it provides more than just a description of a process. It plays go.

Go-playing computers are a significant advance in artificial intelligence, but the ultimate goal of artificial intelligence is to create what is called “artificial general intelligence,” that is, a general-purpose system with intelligence equal to or greater than that of a human mind. Current artificial intelligence applications are each savants in some particular task. DeepMind’s AlphaGo can play ‘go’, but it is useless for anything else. Although some people are extremely worried about the prospect of creating an artificial general intelligence, we are very far from achieving such a thing.

Artificial general intelligence will not be achieved using our present methods. It will require some real (in every sense) invention if it is ever going to be achieved. Using current methods, computers can estimate or select in a context specified by their representation, they cannot create their own representations, but that is arguably where the actual intelligence resides.

There are whole groups of problems that people can solve that computers cannot yet address (such as insight problems, for example). These problems require the person to come up with a novel representation. No one has yet invented a process that would allow a computer to invent its own representation. Until that happens, there is no chance of creating anything more than collections of special-purpose systems. The person who can solve that kind of problem may well deserve to have a monopoly on it for a few years. Whatever that process is, it will definitely not be an algorithm in the formal sense.

When we built machines from pulleys and gears, the idea of a pulley or a gear was not new. The innovation came from how they were combined and used. Now we use computer programs rather than pulleys and gears. Why should that kind of innovation be any less valuable than one using mechanical parts?

The waffling definition of algorithm used in the discourse on this topic means that people, particularly the courts and the patent office, are talking past one another. Algorithms (in the formal sense) are not now and probably never will be patentable, nor will equations or physical principles, such as the Bernoulli principle. But airplane wings can be patented.

For example, see: US Patent 3,860,200. The conversation on patentability of software uses a lot of sloppy language to make what would otherwise be a reasonable point. Artificial intelligence systems are definitely not algorithms (in the formal sense), though people may call them that. They are processes, not abstract methods and not mathematical entities.

At Mimecast, artificial intelligence and machine learning play an increasingly important role in our arsenal of tools that we use to identify malware and other important risks. We see these tools, not as magic, but as just as ordinary computational methods that make up the modern approach to information security. Nevertheless, we are also working on several highly innovative applications of artificial intelligence and we will undoubtedly be seeking patent protection for some of them.

Register or login for access to this item and much more

All Information Management content is archived after seven days.

Community members receive:
  • All recent and archived articles
  • Conference offers and updates
  • A full menu of enewsletter options
  • Web seminars, white papers, ebooks

Don't have an account? Register for Free Unlimited Access

Herbert Roitblat

Herbert Roitblat

Herbert Roitblat is principal data scientist at Mimecast, an international company specializing in cloud-based email management for Microsoft Exchange and Microsoft Office 365.