It’s All Been Done: Will AI-Generated Prior Art Scorch the Earth for Future Inventions?

Imagine an invention machine that is capable of analysing all published human scientific knowledge and then extrapolating forward. If this invention machine were then to publish its own predictions of, say, future materials, hypothetical drug compounds, or improved cell-based therapies, could this prevent future innovations from being patented?

 

What is new?

Novelty is the first hurdle of patentability for an invention and relies on a fairly objective assessment: either the invention has been disclosed to the public before the filing date, or it hasn’t. Novelty is assessed against anything that is published either in writing, online or even disclosed orally, such as in a public meeting or presented at a conference. Novelty is examined by looking at the claims of the patent application and comparing them against everything made available to the public before the earliest filing date (the so-called priority date). If the subject matter of the claims has been described already in the public domain, then there is no novelty and the application is refused. Patent office examiners will scour databases and the internet to look for earlier patent applications, articles or public disclosures from the inventors named on a patent application in order to check whether they revealed the invention before they filed the patent application. This is the typical first point of call for search and examination, after which the Examiner then looks to whether others have described similar inventions before the patent application was filed.

 

Let’s return to the invention machine, it seems one way to scupper any future blocking patents might be to generate hypothetical inventions extrapolated from what is already published. Such an approach could rely on the use of generative AI to scrape the internet to create a training data set of academic publications, patent applications, YouTube™ videos of presentations, TED™ talks and the like. It would be fairly straightforward to have the AI then make predictions of future inventions based on what is already known. These predictions could be self-published in a searchable online database and made available to patent offices and technology courts around the world. Could this scorched earth approach really block future competitors from securing vital intellectual property?

 

Where does an AI hide a leaf? In a forest!

The task of the invention machine is to extrapolate from what is known already. This is no small task; the output of the machine would be an exponentially increasing mass of hypothetical predictions. In order to be credible as a prior art disclosure there would need to be sufficient detail present in each discrete disclosure such that it couldn’t be easily disregarded as a generic waffle. But to generate sufficient critical mass to really clear a path for future innovation, the invention machine would need to generate enormous numbers of these predictions, each one being a slight variant of its nearest neighbour in the database. You might imagine each discrete disclosure as a pixel that cooperates with the other disclosures to provide a wider picture of the prior art disclosure coverage.

This is not a new concept in patent law. In the field of chemistry and pharmaceuticals patent owners have been describing broad classes of compounds in the form of a generic formula (a so-called Markush group) and then asserting that such a formula discloses each and every member of the class which could run to many billions of individual compounds. In the English Courts, the novelty of a compound selected from such a class was considered in Dr Reddy's Laboratories (UK) Ltd v Eli Lilly & Co Ltd [2009] EWCA Civ 1362, [2010] RPC 9. The judge in that case made the point that it makes no sense to say that a generalised prior description discloses a specific matter falling within it, rather what one must look for by way of an anticipation is an individualised description. This approach was reaffirmed in the more recent Court of Appeal decision ModernaTx, Inc. v Pfizer/BioNTech [2025] EWCA Civ 1032, where the novelty of an invention said to have been selected from a previously disclosed list was held to be a matter of fact and degree depending on the precise content of the prior art. In this latter case the Appeal Court referred to EPO case law that it was necessary to distinguish between “the purely intellectual content of an item of information” on the one hand and “a specific teaching with regard to technical action” on the other hand (T296/87 Hoechst/Enantiomers).

 

It is this concept that I think could defeat the invention machine the output of which could be viewed, in part or in totality, as simply a very long list of predicted variants. Going back to Dr Reddy's Laboratories (UK) Ltd v Eli Lilly & Co Ltd, the judge referred to an old riddle: where does a wise person hide a leaf? In a forest. The reasoning for this being that it is ridiculous to say that a particular leaf has been made available to you by telling you that it is in Sherwood Forest.

Could an AI create an individualised description?

I think the invention machine would need to produce disclosures with more than merely prophetic statements if it were to meet the threshold for an individualised description for each and every pixel of disclosure. In doing this there is likely to be a risk that some of its other descriptions might contain contradicting disclosures. Asking the invention machine to sanity check its output for internal inconsistencies is quite an ask. After all, there is no real world data to test which extrapolation is correct, more likely the machine simply publishes all of them. This is the problem of lack of supporting real world data. All of the pixels of disclosure are based on hypothetical results and are yet to be tested. It is possible that real world tests might identify different results that would affect the outcome of any assessment of novelty.

 

An invention will often lie in the areas where the predictions are wrong surprisingly, or where real-world data is unexpectedly beneficial. Unpredicted synergies or overcoming unseen blocks can result in inventions that may never have been conceived of by an agentic AI trained to avoid radical deviation from normal design space. Even if an AI powered invention machine has published a prediction that on face value anticipates a claimed invention, the patent attorney should look to see if that prediction falls within a plethora of similar predictions so that the invention actually represents a novel selection (the leaf) from a much wider list of disclosures (the forest).

 

In conclusion, I don’t think the era of AI powered invention prediction will scorch the earth for future inventions, but it does emphasise an even greater need to rely on real world data to support the novelty of an invention and differentiate it from mere AI powered speculation.

 

Note: This article was not written using an AI – it is all my own thinking.