AI Patents: Could a US Supreme Court case about cholesterol drugs provide guidance for Artificial Intelligence companies?

Artificial intelligence (AI) and machine learning (ML) are increasingly central to innovative companies’ product development efforts, and a frenzy of patent activity is already underway. According to Bloomberg, the global AI market is expected to grow more than sevenfold – from about USD59.7 billion in 2021 to USD422.4 billion in 2028. 

There are around 20,000 AI-related patents issued in the US each year, up from roughly 5,000 in 2018. IP Watchdog’s patent of the year for 2022 was an AI tool for generating animated cinematography from a text prompt.

Courts and government agencies have also begun to address a number of interesting legal questions relating to AI and ML intellectual property. For example, courts in a number of countries have already concluded that AI cannot be an inventor (for a patent), nor an artist (for a copyright)¹.  

In the US, another significant question for 2023 is whether the providers of AI-powered image creation tools are subject to copyright infringement claims by the creators of material used to train the AI system².

What Amgen v. Sanofi means for AI and machine-learning

At first glance, Amgen Inc. v. Sanofi (No. 21-757), a patent case recently argued in the US Supreme Court, does not appear to address questions relating to AI and ML. Instead, it is a pharmaceutical case focused on the level of disclosure required to claim a functional class of antibodies. But the Court’s decision, including its articulation of the enablement standard, could have impacts on the patentability of inventions derived from AI and ML.

The Supreme Court agreed to hear the Amgen case to address the question of whether a patent specification must enable those skilled in an art “to cumulatively identify and make all or nearly all embodiments of the invention without substantial ‘time and effort.’” Amgen’s patent claims are drafted to cover the entire genus of antibodies that achieve a particular effect: binding to a particular “sweet spot” that inhibits the body’s absorption of cholesterol.

While Amgen’s patent discloses the structure of 26 antibodies, it does not provide an exhaustive list of the antibodies covered by the claims, which could number in the thousands or millions, each with a varying degree of efficacy. For example, Amgen’s screening identified 384 antibodies with some blocking potential, and 100 antibodies with greater than 90% inhibition.

Accordingly, an ordinarily skilled artisan would have to create and test a particular antibody in order to understand whether it achieves the desired effect – and thus falls within the scope of the claims.

Many software inventions are far more predictable than the antibodies at issue in the case. But AI and ML technologies can involve an element of unpredictability that brings them closer to the facts of Amgen.

Neural networks present similarities to antibodies at issue in Amgen

For instance, many advanced AI and ML systems use neural networks, computing systems inspired by the architecture of human brains. As in the human brain, synthetic neurons transmit weighted signals in response to input, thereby signaling neighbor neurons and propagating signals that result in an output. Neural networks typically have at least three basic layers for the synthetic neurons: an input layer for receiving stimuli, an output layer for doing work, and a hidden layer for decision-making.

Programmers input relatively little code in constructing these networks; instead, the focus is on training the neural network by processing examples with known results. Training the neural network can require substantial time and effort, and involve huge data sets.

Once trained, the overall architecture of neural networks, including the arrangement, signal parameters, and flow of information, can be highly complex. Moreover, the resulting arrangement and weighting of the neurons is largely a “black box” – unpredictable in advance, and difficult to explain exactly why it works as well as it does.

And to the extent that neural network is then applied to develop follow-on inventions, the inventor (which cannot be the AI itself) may not be able to explain the basis for the resulting invention because the derivation of the invention may be locked within the structure of the neural network itself.

Thus, at least as it relates to the enablement standard under US law, certain AI and ML technologies like neural networks may be similar to the antibodies at issue in Amgen.

Accordingly, the Court’s decision in Amgen could provide some certainty around the patentability of AI and ML technologies, and other inventions resulting from applied AI.

For example, if the Court were to require the disclosure of a specific structure to confer patentability, then arguably certain types of AI and ML patents could be limited to particular embodiments, like a specifically trained neural network.

But even if the Court does not substantially revise the legal standard for enablement, a question will remain whether the disclosure of a specifically trained neural network is sufficient to enable others to understand and practice other claimed embodiments of the invention, where the disclosed neural network is effectively a “black box”.

Courts will soon weigh in on these questions, and the Supreme Court’s decision in Amgen could have an outsized impact on the tidal wave of AI and ML patent litigation on the horizon. Billion dollar cases may soon depend on how specific such disclosures must be to enable skilled artisans to recreate inventions relating to AI and ML.

Footnotes

_{1 See, e.g., Thaler v. Vidal, 43 F.4th 1207, 1212 (Fed. Cir. 2022). 
2 See Andersen et al. v. Stability AI Ltd. et al., No. 3:23-cv-00201 (N.D. Cal. Jan. 13, 2023).}