As artificial intelligence (AI) technologies have advanced over time, strategies for patenting them have evolved in kind. Most recently, the promise of generative AI to help leverage the knowledge contained in unstructured data has captured global attention, spurring novel advances from traditional high-tech companies and startups.
This article outlines a framework for creating valuable patents for protecting AI technologies, as part of a series that covers topics including why to invest in patents for AI, how to overcome the biggest challenges in patenting AI, and business factors to consider in building your patent portfolio. The framework highlights several complementary approaches for claiming the inventions underlying any particular AI technology, and answers the following questions that characterize useful patent claims:
- does the claim prevent competitors from performing AI functionality that provides a competitive advantage in your market
- is the claim resistant to invalidity challenges
- can infringement of the claim be detected
At a high level, most (if not all) AI technologies can be broken down into the inputs provided to an AI algorithm, such as a machine learning algorithm, how the algorithm processes the inputs (which may often be seen as a black box), and the outputs of the algorithm. For example, a system for designing new drugs could include an algorithm trained on examples of drugs that meet existing criteria, so that at runtime the algorithm can generate candidate drugs expected to meet specific criteria of interest. In another example, a customer service chatbot could fine tune a GPT-type large language model based on positive examples of effective conversations (as well as negative examples of ineffective conversations.)
As shown below, patent claims can generally be prepared to cover at least three categories of inventions for AI technologies that map to the input-processing-output breakdown.
| Input | Processing | Output |
| Training Data Generation: Where a machine learning model is used, what steps are taken to make the training data more useful -- which results in a more useful model -- such as filtering the data, reserving data for validation, automating the generation of training data or incorporating user feedback? | Model Training: Where a machine learning model is used, how is the model being trained and what is the value-add generated by using a trained model as opposed to other approaches, such as rule-based or engineered approaches? | Insights and Actions: What is the outcome of using the AI solution? What patterns or new knowledge are identified? How is a physical machine, such as a medical device, industrial automation robot, or autonomous vehicle controlled using information generated by the AI solution? |
The framework outlined below applies this breakdown in a manner
suited for machine learning technologies that
train models to generate useful insights and
actions. It can also be readily adjusted to
address AI solutions that do not necessarily use machine learning
models, such as heuristics/rule based engines:
Machine Learning Patent Protection Framework
| Category | Why | How | Detectability |
| Insights/ Actions | This is the real world result and value add |
How does the action improve the underlying technology? Are you faster/more accurate/able to respond to situations before they occur? |
Focus on inputs and outputs while varying the granularity of claims from the processing components up to the system in which the AI processing is implemented Ideally can treat the algorithm itself as a black box |
| Model Training | This is where you enable the technical improvements that make your system have a competitive advantage |
Does training the model enable your system to perform functions it could not previously? Does training the model enable your system to be faster or more accurate than other ML approaches, human approaches or rulesbased approaches? |
Focus on input data including the source of the data and the combination of input data parameters Capture the most likely types of models that could be used using varying claim scope |
| Training Data Generation and Pre-Processing |
Remedy for "garbage in, garbage out" problem Automation of the preprocessing to make implementing AI solutions faster/more efficient/ use fewer computational resources |
Emphasize computer steps used to improve the training data Identify unique approaches for manipulating baseline data into data that may be expressed in different forms relevant to the technology area, such as filtering, changing dimensionality, automated annotation, identifying a subset of parameters that are more significant than others for the model, etc. |
Focus on input data including the source of the data and the combination of input data parameters Capture the right types and scope of data, sift through to get the optimal amount of data, and properly label the data to teach your system the correct way for insights/ actions |
Patent claims can generally be prepared to cover at least
three categories of inventions for AI technologies that map to the
inputprocessing-output breakdown.
Typically, patent strategies for AI technologies focus on one or more of these aspects, depending on where the innovation lies and/or where the most value is created for end users. Companies should keep in mind that as the AI patent landscape becomes more crowded, more nuance will be necessary to properly craft applications and claims. In the case of generative AI technologies including large language models and image generation models (e.g., diffusion models), invention harvesting and application preparation efforts should focus on extracting as much information as possible regarding how the technology provides benefits that could not be achieved simply by using the baseline, publicly available models. For example, understanding how the technology is made so that it can utilize previously challenging unstructured data, how it ensures that outputs are accurate (i.e., solve the hallucination problem), and how it creates benefits that are not possible with more traditional AI will provide significant dividends for building valuable patent portfolios.
By starting from the framework shown above and incorporating innovation-specific nuances, companies can identify multiple ways for patenting your AI technology, resulting in a stronger patent portfolio that will be easier to enforce and more difficult for competitors to design around.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
We operate a free-to-view policy, asking only that you register in order to read all of our content. Please login or register to view the rest of this article.