Creativity is human

The reason we may first treat patents filed by computers as a joke is that we were brought to believe that creativity is human, and so only humans can be creative. As unorthodox as I try to be, I still hold strong to this opinion.

Computers running algorithms of machine learning (ML) “invent” by finding patterns that humans don’t find. They reveal and utilize hidden logic in the input data that may be buried so deep, and which is so well hidden from the human mind (which is limited in its data processing capabilities, and so “falls-back” to intuition), that its outcome is sometimes astonishing; but it’s not creative.

Invention vs. creativity

Referring to this ML generation process as “invention” either makes sense or not, depending on what you consider the definition of “invention” to be. Humans come up with revelations using intuition that (also) covers for low computational abilities. Computers “come up with” revelations by numerical brute-force. People think outside the box. Computers think only inside the box, but their computational abilities (in terms of speed and size of state data) make this box very large.

If we judge whether an act is one of “invention” by the outcome of that act, then we can easily accept that computers do invent, even if they do so without exercising creativity. For illustration, imagine we have an algorithm that tries all possible combinations of gluing together 3 to 10 plates of wood (or any other firm substance), in a way that the constellation is such that can tolerate high vertical pressure across a horizontally flat surface without collapsing. When you see an output of this algorithm, you would say that this algorithm created (or: invented) a chair, but you will not consider the computer to have been “creative” in the process. This algorithm could have run on an old computer of decades ago. Today, machines are more powerful, and so are the algorithms they run, allowing them to find solutions to more of our problems, but the paradigm of differentiating between inventor-ship (which could be mechanical) and creativity is the same.

Once we realize that inventor-ship does not require creativity, we can have a constructive discussion about the viability of patents for ML “inventions”. We are used to revelations being made by people with the right orientation, knowledge, mindset, and luck, not as much to revelations made by computers finding patterns in data. Nevertheless, the outcome is the same, hinting that it’s high time we separated invention (the goal) from creativity (the way to the goal, or from now on — a way to the goal).

What are patents good for?

The primary purpose of patents is to restrict the terms by which people can exploit an existing invention, so to give its inventor a temporary head-start in exploiting his invention, in return for teaching his invention to the public. Such publication serves the public’s best interest, and hence governments offer this deal to the inventor. The patent system, as it turned out, does have its areas for improvement, but its purpose is surely pragmatic and appreciable.

The patent is thus a practical method of balancing interests between an inventor and the public, making it worthwhile for the inventor to invest in research and to publish results that can be put to use in products and services. What is a patent not? — It is not a creativity award.

An invention created by a machine-learning algorithm is not a result of creativity, but it is still an invention that may benefit from the protection provided by the patent system, balancing the interest of the legal owner of the system that produced the invention and the interest of the public in having access to the invention. In this regard, ML-generated inventions are no different than man-made ones. One thorny issue is that of ownership.

Ownership

Once we agree that a patent can be invented by a piece of software, we come to the more interesting question of assignment. A patent usually has a name of an inventor (one or more), but this name really means not much more than a little trophy. The more interesting name on the patent is that of the assignee. This is the legal entity (person or company) who owns the rights on that patent. Who should the assignee be for patents on inventions created by an algorithm?

It is probably widely agreeable that the assignee of a patent shall still be a legal entity. Otherwise, the patent cannot serve its basic primary purpose of protecting the interest of someone who should be entitled for commercial interests. As long as software has no bank account that it legally owns, there is no point in considering it as a viable assignee. The remaining contestants are the maker of the ML system (software, hardware, etc.), the one who operates it, the one on who’s behalf it is operated and who may sponsor it, the one who owns the inputs, and just about anyone else in the value-chain of the system and of its data.

The cleanest approach, in my opinion, is to assign the patent by default to whoever files it, that is, to the entity that first possesses the patent contents (description and claims) and that files them (the “first to file” principle). Then, we allow traditional commercial laws to define the system by which parties agree to pass such assignment between them, as part of standing agreements between them. This is not substantially different than how patents are handled today when their (human) inventor is not their lawful assignee, for example, when the invention is done for hire. In such cases, the typical flow is that the inventor signs a transfer of assignment to the company that hired him to carry out the work, usually honoring a clause in the engagement contract which prescribes this change of assignment.

This approach will keep the patent filing and ownership system clean and independent of the complexities of relationships between the stakeholders along the value-chain of the ML system that created the invention. The contents of the filed patent are first available to one entity, typically the one operating the system. This entity has the procedural ability to file the patent (the equivalent of a human inventor in a non-ML patent), and this entity that approaches the patent office and files, will get the assignment by default. If this assignment better belongs elsewhere (for example, if the entity that got first hold of the output invention was merely running an ML system on behalf of some other entity that should own the rights), then let those entities settle the transfer of assignment later.

Summary

Our discomfort with the notion of patented matter created by machine learning algorithms stems from our traditional binding of inventions with creativity. Machines have always helped us solve our problems, and at this day and age they are powerful enough to solve complex problems in ways that may result in patentable matter. This shall not be confused with creativity, but it is still inventor-ship that can benefit from the protection granted by the patent system.

Ownership can be handled by the traditional approach of: whoever has it, files it, and assignment is independently transferred later. This leaves the patent system oblivious to the complexity of the value-chain that might be involved in building and running the ML system that created the invention.

Edited to add (2021-10-27)

The BBC News published the piece: “AI cannot be the inventor of a patent, appeals court rules”, where Lady Justice Elisabeth Laing wrote in her judgement:

Only a person can have rights. A machine cannot

This, in my opinion, summarizes a fundamental source of the dispute: the ‘right’ is not in the inventor-ship but it is in the assignment of the patent. The inventor is merely the source of the protected invention. Traditionally, it was a person, but inventions coming from machines could also be eligible for patent protection for the same motivations. The rights would still belong to the legal entity that files those inventions, not to the machine.