Artificial Intelligence-based Patents: Perspectives for Practitioners and Patent Owners

Article co-written by Yuri Levin-Schwartz, Ph.D., a law clerk at MBHB.

Innovations involving artificial intelligence (AI) and machine-learning (ML) are being developed at an ever-accelerating pace. For example, as illustrated in Figure 1, the number of patent applications published by the United States Patent and Trademark Office (USPTO) including the phrase “artificial intelligence” or “machine-learning” has increased from 406 publications in 2009 to 4,091 publications in 2019 (projected from statistics available on September 11, 2019).

Figure 1[1]

As evidenced by such increasing patent activity, AI/ML-based inventions and applications are emerging in every area of technology, including robotics,[2] integrated circuit design,[3] and health care,[4] to name just a few. Such a burst of discovery may be comparable to only a few past developments, such as the integrated circuit and the combustion engine.

As ever more patent applications are being drafted in this technology space, this article provides practitioners with best practices for drafting and claiming AI/ML-based inventions, as well as guidance to patent owners for building a valuable AI/ML patent portfolio.

Practitioner Perspective: Drafting Effective AI/ML-based Patent Applications

From the outset, practitioners should try to obtain as much publicly-disclosable information as (reasonably) possible about the AI/ML-based invention. In particular, practitioners should gather information on the novel, non-obvious aspects of the given invention including, but not limited to:

  • Input data preparation – how is data gathered, pre-processed, handled, or parsed upon use by the AI/ML model? Is the data obtained from a specific type of sensor (e.g., a camera or radar device)? What does the data represent (e.g., letters, words, Boolean values)?
  • Model structure – does the model have specific non-generic features (e.g., a neural network with non-conventional number of nodes at given layers, multiple hidden layers, etc.) and how is the input data mapped to this specific structure?
  • Training phase process – how is the model trained, and in what manner (e.g., tagged input-output pairs, unsupervised learning, etc.)?
  • Execution phase process – what weights are used with respect to what variables? What advantages result from execution of the AI/ML model?
  • Output data post-processing / analysis – how are the outputs utilized and what do they represent (e.g., classification, recommendation, likelihoods, etc.)? Is the data used to control a specific device (e.g., a speech synthesizer or an autonomous vehicle)?
  • Locus of AI/ML processing – is the substantive computing performed locally (e.g., “AI on the edge”), in the cloud, and/or elsewhere?
  • AI/ML-based hardware – does the innovation utilize AI-specific integrated circuits, such as AI-optimized graphics processing units (GPUs)? How does the model structure map to such hardware?

After obtaining this information, but prior to investing a significant amount of time and resources to draft a full patent application, inventors and their organizations should consider conducting a prior art search based on a brief, written disclosure summary and one or more sample claims. Such a search may help “short circuit” over-ambitious AI/ML-based applications or overly-broad claim scope, and can almost always provide helpful context about related prior art in a given technical area. In the near term, a prior art search can help save the client money if close prior art is found. Furthermore, in the longer term, the references produced in such a search may help practitioners and companies work around certain competitor filings.

Since the patent claims define the scope of the future legal right, practitioners should take special attention to properly capture the full intent and various embodiments of the invention whenever possible. For example, when drafting AI/ML-centric patent claims, practitioners should:

  • Include a “patentable hook” in each independent claim, which hopefully relates back to the AI/ML-based nature of the invention.
  • Separate training phase processes from execution phase processes by utilizing independent claim families directed to the respective methods. Doing so may help avoid split infringement issues down the road.
  • Try to incorporate as much physical structure (e.g., controller, computation unit, circuits, etc.) as possible into the claims to obviate issues with 35 U.S.C. § 101 (patentable subject matter). This structure can include both special hardware used for the training and execution phases, as well as physical devices that provide input data or receive output data.
  • Utilize patent analytics by “testing” sample claim language to predict art unit assignments and iteratively adjust claim terms to actively avoid business method type art units (e.g., art unit 3600) or other low-allowance-rate art units.

While drafting the patent application, practitioners should strive to describe a primary embodiment in depth while exploring several alternative embodiments to help broaden the overall disclosure. Breaking the patent specification into multiple sections that correspond to the major elements of the AI/ML-based invention (e.g., input data preparation, model structure, training phase, etc.) can compartmentalize the disclosure and may help ensure that each novel, non-obvious detail is described thoroughly. Patent drawings should expressly illustrate each substantive claim term in a schematic-type diagram or a method flowchart. Ideally, the drawings should be organized and illustrated so that another patent practitioner might be able to identify the patentable hook upon a brief inspection of the drawings.

Patent Owner Perspective: Developing an Intentional AI/ML-based Patent Portfolio Strategy

Companies are time and resource limited. Meanwhile, building and maintaining an AI/ML-based patent portfolio can be expensive and time-consuming. Accordingly, pursuing an intentional, introspective patent portfolio strategy may help maximize return on investment for a company by 1) focusing on the company’s core needs and strengths and 2) considering the company’s competitors and industry to help maximize its competitive advantage. From a high level, a central question to a patent owner is: “Would you care if someone copied this AI/ML-based invention?” However, the answer can involve many further details and potential considerations. When an invention satisfies many of the factors in the checklist below, it may indicate that preparing and filing a patent application may be in line with the company’s interests.


First, consider the invention and how it relates to the business and patent portfolio goals of the company.

  • Consider key industry players (competitors, partners, customers)
    • Is this AI/ML-based invention directed to technology related to the key industry players?
    • Is anyone outside of your company (and notably, any key industry players) using this invention or a related technique?
    • Does this particular AI/ML-based innovation represent a new approach within your industry?
  • Consider this AI/ML-based invention with respect to your own company
    • Is this invention directed to a fundamental/core technology and/or product of your company?
      • If so, consider filing an “omnibus” provisional application in the United States with many (e.g., 30-50) multiply-dependent claims and as much disclosure as possible to obtain an early priority date. This can provide maximum express support for future Patent Cooperation Treaty (PCT)/international filings and retain options for claim scope/direction at the US conversion stage.
    • Is the AI/ML-based invention currently in use at your company? If so, is it successful?
    • Is this AI/ML-based invention a customer driver for your company (i.e., generates more customers, makes platform more user friendly, a reason why customers would choose your product/company over a competitor, etc.)?
    • Does this AI/ML-based invention provide a competitive advantage?
  • Consider your company’s investment in this AI/ML-based invention
    • Has a substantial amount of money been invested in the research and development of this invention? Patents are a form of insurance, so more money invested leads to higher interest in patenting.
    • Have a substantial amount of employee hours been invested in the development of this invention?
  • Consider the impact of the AI/ML-based invention
    • Does the invention work (or work well) (e.g., AI-based invention reduces power usage by 10%)?
    • Is a prototype made and working?
    • Does this invention provide a disruptive solution to an existing problem in the industry?
  • Consider the “shelf life” of the AI/ML-based invention
    • Does the invention have a short product life (e.g., software revisions likely to occur within 1‐2 years)?
    • Does the invention have a long product life (e.g., disruptive, fundamental new idea)?
  • Consider publicizing the AI/ML-based invention
    • Are you okay with making this invention available for the world to inspect?
      • Consider whether to file a non-provisional application in the United States along with a non-publication request to prevent publication. A rescission of the non-publication request can then be filed at the PCT stage, if necessary.
    • Does the invention relate to or contain any sensitive/confidential information (a patent will be publicly available for anyone to read)?
      • Weigh eventual patent right against potential adverse impact to existing trade secret portfolio.
  • Consider patent ownership
    • Will the company be able to claim full or partial ownership over the patent application?
    • Was the AI/ML-based invention jointly developed in combination with a supplier, vendor, independent contractor, customer, or other third party who may have rights?
    • Does the AI/ML-based code include third-party code and/or open-source code?
      • Open-source software licenses vary widely and can, in some instances, prevent patenting of software that implements it.
  • Consider the potential market for this patent and monetization potential
    • Is there a potential for licensing revenue?
      • Consider utilizing different independent claim families to separate training phase and execution phase to keep licensing compartmentalized.
    • Is there a potential for sale of the patent (and return license)?
    • Is there an interest in enforcing the patent for monetary damages?
      • If enforcement is a primary interest, use claim charting and infringement analysis to streamline independent claim elements and method steps.
  • Consider whether you would like to be able to prevent competitors from using the invention
    • Would you like to have control over who can use the invention?
    • Is the invention easily detectable?
  • Consider copying and reverse-engineering of the invention
    • Is the invention very difficult to reverse-engineer (e.g., AI-based application running on private cloud server – perhaps maintain as trade secret)?
    • Is the invention relatively easy to reverse-engineer (e.g., AI-based application deployed on 100,000+ mobile devices – may need patent protection to stop copying)?
  • Consider international market/competitors
    • Would you like to protect the AI/ML-based invention in foreign jurisdictions?
      • Consider filing in countries with the most AI/ML development. (e.g., China, Europe, South Korea, Japan, India, Singapore, Taiwan,[5])


Second, consider the AI/ML-based invention and whether the legal requirements to obtain a patent can be satisfied.

  • Consider the prior art
    • When considering the technology that exists today, is the AI/ML-based invention clearly/substantially different than conventional methods/systems (e.g., a convolutional neural network with novel structure and/or novel input data mapping, etc.)?
  • Consider the potential claim scope in light of the prior art
    • Is the possible claim scope valuable to your company?
    • Would the possible claim scope be too narrow?
      • Too narrow and hard to enforce?
      • Too narrow and easy for a competitor to design around?
  • Consider enforcement of patent rights
    • Is it easy for you to identify when someone is using the AI/ML-based invention (i.e., identify infringers)?
    • Does the invention find use only internally or internal to a company (e.g., company is running AI-based models on a private server, for test purposes only)?
    • Does the invention find use by customers (e.g., customers are deploying trained models to provide recommendations or classifications)?
  • Consider patent eligibility
    • Is the invention “directed to” laws of nature, mathematical theories, etc.?
      • An algorithm (e.g., a neural network or support vector machine) in and of itself is not patentable subject matter.
      • However, for an algorithm utilized in a broader method or utilized in a system, the method or system generally could represent patentable subject matter.
    • Does the invention involve non-generic hardware (e.g., AI-optimized GPUs, FPGA trained with machine learning model, etc.)?
      • Claiming an AI/ML-based model that interacts with non-generic hardware is an ideal scenario.
  • Consider whether the invention is sufficiently developed
    • Is the invention past the “idea” stage (e.g., AI/ML-based code has been developed and implemented in internal tests, training phase has been completed, etc.)?
      • A working prototype is not necessary, but the patent applicant must be able to describe how to make and use the invention in detail.

AI/ML-based patent filings will continue to rise for the foreseeable future as artificial intelligence takes a larger role throughout our lives and within society more generally. However, such patent applications should be drafted with special considerations tied to the underlying AI/ML-specific subject matter and claims. Furthermore, patent owners should strive to employ an intentional strategy aligned with their company’s business interests when building their AI/ML-based patent portfolio.

[1] Yearly totals of USPTO-published patent applications with disclosures that include the phrase “artificial intelligence” or “machine-learning.” Google Patents, (accessed Sept. 11, 2019).

[2] See, e.g., Intelligent Robot Based on Artificial Intelligence, U.S. Pat. App. Pub. No. 2019/0111566.

[3] See, e.g., Machine-learning Circuit Optimization Using Quantized Prediction Functions, U.S. Pat. App. Pub. No. 2019/0228126.

[4] See, e.g., Artificial-intelligence-based Facilitation of Healthcare Delivery, U.S. Pat. App. Pub. No. 2018/0182475.

[5] Note that Taiwan is not a signatory of the Patent Cooperation Treaty and direct filings are due in that country within 12 months of the earliest priority date.