Artificial intelligence has already revolutionised the life sciences and healthcare sectors, and its potential applications seem to continue to grow exponentially. Across industries, recent surveys have found that around 51% of people are using Large Language Models (LLMs), like ChatGPT, for informal learning and between 24 - 33% use LLMs for their work (double the number who did so only two years ago). 

As use of LLMs becomes more and more prevalent, patents with priority dates in the year 2024 or later may very well be addressed to “persons skilled in the relevant art” (or PSA, a hypothetical non-inventive skilled worker in the field of the invention of a patent) who would frequently rely on LLMs such as ChatGPT as a reference or to assist with their day to day work. We recently considered whether an LLM might itself be considered to be a person skilled in the art, in our article published in The Prescription. However, a related question is whether LLMs could form part of the “common general knowledge” (or “CGK”) in patent proceedings, which would have significant impacts on the consideration of validity both during patent prosecution and in patent litigation.

Could ChatGPT become CGK?

“Common general knowledge” is the background knowledge and experience that all those skilled but non-inventive workers in the relevant field of the invention at the priority date may be expected to have, which may be formed through their reading, observation, training and experience. What is considered to be “CGK” is often determined based on the evidence given by parties’ independent expert witnesses. 

In Australia, the CGK is often found to include the information in standard textbooks, reference works, technical dictionaries, widely read journal articles, or other materials relevant to the field that a person skilled in the art knows exist and would refer to as a matter of course. One prime example is “Remington: The Science and Practice of Pharmacy” (published by the Philadelphia College of Pharmacy), which is frequently referred to in pharmaceutical patent proceedings as part of the common general knowledge known to formulation experts.

The advent of AI thus poses an interesting question: how would a court treat evidence from an independent expert in the field of the invention to the effect that they would use ChatGPT to interpret a claim or the complete specification of a patent, or even to solve a technical problem as part of their inventive step evidence?

The United States Patent and Trademark Office Request For Comments (30 April 2024) highlights that the proliferation of AI tools, like ChatGPT, raises questions about whether AI-generated outputs could influence what is considered prior art or the knowledge base of a person having ordinary skill in the art. The USPTO asks whether widespread access to AI might shift the level of skill or common general knowledge, and whether disclosures created by AI should be treated differently from human-authored ones.

Transparency, reliability and explainability

There are several impediments to acceptance of LLMs as part of the common general knowledge, including the difficulties of explainability (why the AI model makes a prediction) and interpretability (how the AI model makes a prediction) of AI outputs. In part, this is tied to the fact that the training data used to train an LLM is unknown to the user (i.e. the expert witness), and thus is also unknown to not only the parties (who are unable to challenge the AI output on that basis) but also to the court tasked with determining whether the output of an LLM like ChatGPT could be part of the common general knowledge. 

OpenAI states that the models that power ChatGPT are developed using 3 primary sources of information: 

• information that is publicly available on the internet;
• information accessed under partnerships with third parties; and
• information that their users, human trainers, and researchers provide or generate.

However, while that information likely includes information in textbooks, dictionaries and other publications relevant to the field of a given invention and which itself might form part of the common general knowledge, the information used to develop ChatGPT also includes a vast array of material that would not otherwise form part of the common general knowledge (e.g. social media posts, blogs, or news articles that may be relevant to other fields or that may be false or disavowed by the scientific community). For now, these characteristics of LLMs are likely to preclude a finding that ChatGPT is part of the common general knowledge, even if the evidence shows that a person skilled in the art would rely on ChatGPT despite these characteristics. 

In order for information to be common general knowledge, it must be generally accepted without question by the bulk of those in the art.  There also must be evidence that the information has been assimilated into the consciousness of the skilled worker. Given the current experience of ChatGPT, including well documented hallucinations, even if experts in a particular field regularly rely on ChatGPT as a matter of course, it seems highly unlikely an Australian court would accept that those experts would accept the output of an LLM “without question”. 

Even if a court were to find that a person skilled in the art would use ChatGPT as a starting point for ideas or to direct them to further research materials, it is likely the court would find that the PSA would approach any AI generated response with scepticism, and would do their own research to verify the response using established and accepted technical literature and resources. This is particularly the case for experts working in the life sciences and healthcare sectors given the stringent regulatory environments in which they operate. 

In the near term, it seems most likely that ChatGPT could take a place in patent proceedings equivalent to a search engine, like Google or PubMed; as a tool for locating research materials, rather than a source of technical information itself. The same is likely true for other AI models, such as the recently introduced Perplexity Patents, which claims to be the “world’s first AI patent research agent” and draws on sources of online information, including patent literature, academic papers and public software repositories, to answer technical questions posed by the user. 

We can also expect attempts to use the output of LLMs such as ChatGPT as secondary evidence of obviousness, in circumstances where the output corroborates the primary evidence of an expert as to the steps the PSA would take as a matter of course in addressing a problem said to have been solved by the relevant invention. 

Where to next?

While integration of AI tools like ChatGPT into the workflows of experts in the life sciences and healthcare sectors is inevitable (and already occurring), the current lack of transparency, reliability and explainability render AI outputs of LLMs unlikely to satisfy the test for common general knowledge under Australian patent law. Courts are more likely to remain cautious, treating AI-generated information as a research tool or database rather than as information a person skilled in the art would know or would refer to “as a matter of course”. To the extent expert witness evidence refers to using an LLM in their approach to a task, the output should be corroborated by sources of established common general knowledge in the relevant field. However, as AI capabilities continue to evolve, and use of LLMs becomes routine, it seems likely that LLMs will one day play a central role in expert evidence in patent cases.