What Can Be Patented?
The influx in AI and ML patent applications begs the question: what can and cannot be patented?
Inventions come in a variety of forms, and the USPTO responds uniquely to the distinct industry that a patent falls within. Generally speaking, a patent can fall into any one of three categories:
Utility patent: The most broad type of patent, utility patents encompass any new and useful processes, machines, articles of manufactures, or composition of matter. This can also include the improvement of such inventions.
Design patent: These patents are used for any new, original, and ornamental design used for an article of manufacture.
Plant patent: These may be granted to any individual who invents, discovers, or asexually reproduces a distinct or novel variety of plant.
While the Constitution uses the word “discovery” within the statute, it is important to note that an individual cannot patent a mere discovery. Rather, the idea or discovery must take physical form or be implementable for it to be granted a patent. This is an important caveat of patent law–the physicality of the discovery. Abstract ideas and mathematical formulas are not patentable under the law, nor are natural phenomena or laws of nature.
So, under federal patent law, for an patent to be granted, an individual must have contributed a new and useful:
Process: Most commonly, this refers to industrial or technical processes, acts, or methods.
Machine: This contribution is the most straightforward, encompassing new fabricated machinery.
Manufacture: It is helpful to think of this as articles that are made.
Composition of matter: This includes chemical compositions, mixtures of ingredients, and new chemical compounds.
Notice that all of these contributions are new, useful, and can be represented physically. These factors are all important, and hinge on the novelty of an invention. This is because the Constitution calls for “novelty and utility” in all granted patents. With this in mind, what does it mean for an invention to have novelty and utility, especially regarding artificial intelligence and machine learning?
Novelty And Utility In AI Patent Applications
Andrei Iancu, Director of the USPTO, puts it well: “Unlocking the potential of AI will provide the basis for future U.S. economic growth and prosperity, and is something that the USPTO will continue to facilitate with our corps of patent examiners and other professionals who specialize in the nuances of this broad-based and far-reaching technology.”
The USPTO recognizes the importance of artificial intelligence and machine learning, and is expanding the scope of their policies to include such technology. As a whole, artificial intelligence includes several subfields of innovation:
Planning/control: Developing processes used to identify, create, and execute activities to reach specific goals
Knowledge processing: Deriving and representing facts about the world using automated systems.
Vision: Extracting and understanding information using photos or videos.
Speech: Understanding a sequence of words given an acoustic signal.
AI hardware: Fabricating physical computer components to meet AI requirements.
Evolutionary computation: Making sense of aspects of nature and evolution through computational routines.
Natural language processing: Utilizing data encoded in written language.
Machine learning: Learning from data using a broad range of computational models.
Under this broad definition of AI component technologies used in the patent landscape, several different innovations can be patentable as long as they meet the novelty and utility requirement. Novelty and utility refers to the usefulness of an invention within any given industry in combination with the newness of the invention.
Utility is the easier prong to meet. As long as an invention has application in a given industry, and benefits users somehow, utility is present. The invention must be somehow desired and useful to the public. Novelty, however, is the most stringent requirement. For novelty to be met, the invention must be new; it must not exist already, in any capacity.
For example, a developer may create a software that uses machine learning to cater recommendations to search-engine users. If the same software already exists in another country, however, then the development would not be considered novel. If an invention is already known to the public before the date of the AI patent application, then it would not be considered novel.
This is only one half of the issue of what can and cannot be patented. To fully understand this dichotomy, one must equally consider what cannot be patented, and what types of AI patent applications may not be accepted by the USPTO.
What Cannot Be Patented?
The USPTO also discusses what is not patentable under federal law. As discussed above, inventions that do not possess inherent utility and novelty cannot be patented. An invention must further a goal and it must be entirely new. In addition, there are other legal nuances that impact the patentability of an invention or developed AI technology.
The USPTO outlines that an invention cannot be patented if:
The invention was described in a printed publication or was otherwise available to the public before the application was filed with the USPTO.
The invention was described in another patent issued to a different inventor.
The invention is not sufficiently different from an original technology, such that it would be considered non-obvious to an individual of ordinary skill.
The phrase “otherwise available to the public” is quite important, as it encompasses a wide range of activities. For instance, an invention is considered otherwise available to the public if it was discussed in a scientific presentation, a lecture or speech, a video published on social media, or even in statements made on radio and television shows. If the invention was made available within online material, this could greatly impact the patentability of an invention.
The second aspect is clearer, meaning that an invention cannot already exist within another patent application or granted patent. It should be noted that this caveat comes into play on the actual filing date of a US non-provisional patent application.
Lastly, and importantly, an invention must be non-obvious to an ordinary individual with basic knowledge of the technology. This connects closely with novelty, suggesting that an invention must be sufficiently new. Surface-level and superficial changes are not enough to create a patentable technology. New contributions to an already-existing technology must be non-obvious and non-predictable.
Patent Inventors Must Be Human: What This Means For AI And Machine Learning Patent Applications
In April 2020, the USPTO issued an important decision that affects AI patent law moving forward. Artificial intelligence and machine learning softwares are reinventing the way that the world communicates and spreads information. To this point, a major facet of such technology is the analysis of data and subsequent generation of solutions.
As such, it is natural to assume that artificial intelligence itself may produce novel inventions that are useful to our society. This is precisely what happened in the 2020 landmark USPTO decision. In this case, an AI machine named “DABUS” generated a fractal light signal using pulses to replicate human brain waves. DABUS was named the sole inventor and Dr. Stephan Thaler, the inventor of DABUS, was the patent’s assignee.
The USPTO held that the patent was unenforceable. They stated that under 35 U.S.C 101, an inventor must be a natural person. This is largely due to the use of the word “whoever,” which implies a human individual. Therefore, qualifying the AI machine DABUS as an inventor contradicts the statute and the resulting patent application is considered invalid. In similar situations, a human inventor must be credited, even if a machine generated the novel approach or application.
With this in mind, AI patent applications cannot name AI technology as the inventor of a machine learning software or application. As the state of AI technology and machine learning continues to progress, this USPTO decision will become even more important and may open up further discussion regarding what can and cannot be patented.
In Sum: AI Patent Applications
As mentioned above, an idea alone is not patentable. There must be a deliverable, a finished product. This becomes relevant when discussing what can and cannot be patented regarding AI algorithms and softwares. Generally, an AI invention would be considered abstract (and therefore unpatentable) if it is considered:
A mental process
A method of organizing human activity
A mathematical concept or formula
According to the World Intellectual Property Organization, the number of applications for AI technology has sharply increased. This trend is expected to continue as researchers and software developers continue to innovate AI technology.
In Sum: Machine Learning Software Patent Applications
Machine learning software patent applications have increased steadily at an average rate of 28% each year. As a subfield of AI, machine learning falls into similar patterns regarding US patent applications. There is some question surrounding whether or not it is possible to patent machine learning algorithms and software.
Machine learning software is undeniably patentable, as it is considered a finished product. Machine learning algorithms, however, are a bit more difficult to define. Federal patent law states that mathematical formulas may not be patented. As such, a machine learning algorithm alone may be considered too abstract to patent.
In these cases, it is possible to patent a series of steps found within your developed algorithm. If you can break your machine learning into its vital components–such as mathematical procedures and underlying software processes–it is possible to be successfully granted a patent.