Has AI Rendered All Future Inventions Obvious?

The World has changed; we can be certain that in an age where generative AI has become readily available to scientists, engineers and creatives, the prospects for potential leaps in innovation seem boundless. Whilst this might appear dispiriting to budding innovators worried about how a mere human can compete with an all-knowing LLM, from the perspective of patent law has the threshold for what qualifies as ‘inventive’ really changed?

The AI of skill in the art

Novelty is the first hurdle of patentability for an invention and generally requires a fairly objective assessment: either the invention is new, or it isn’t. The second hurdle set by most patent legal systems is more subjective and relies on the assumption that a patent claim demonstrates an inventive step with reference to whether a person of skill in the relevant field of art would have deemed the claimed invention obvious. The determination of novelty and obviousness is made at the appropriate priority date for the claimed invention. The question of who the skilled person is might be a little bit more interesting, though.

 

The person of skill in the art is not a genius but nor are they an unimaginative loafer. Rather we are told by the case law that they are best seen as someone who is a competent, if slightly boring, nerd. This is a person who has a lot of knowledge but not a lot of ingenuity. They don’t act out of idle curiosity but do respond to prompts, or pointers, in the right direction. They need a push; they need motivation and they may or may not arrive at the claimed conclusion depending on what they learn from the enormous body of common general knowledge that they have ready access to at their fingertips.

 

In practical terms, as patent attorneys we spend a lot of time arguing with patent office examiners, opposition divisions, judges and other attorneys about who this fictional person of skill in the art is (or is not, depending on your perspective). It varies on a case-by-case basis as does the type of prior art information they are readily exposed to. However, if we look at these well-established definitions of the person of skill in the art, they do bear a resemblance to an AI agent. We have an ‘individual’ with access to an enormous body of knowledge but requires prompts to analyse it effectively and generate something of value. Sometimes they go off on the wrong track, though, and might reach an incorrect conclusion based on poor prompting or erroneous underlying data. We won’t mention hallucinations – the person of skill in the art doesn’t yet do that.

 

Maybe technology has simply caught up with decades of established case law, and instead of a purely legal construct we now actually have an actual embodiment of the notional person of skill in the art in the form of an agentic AI.

 

Modelling the data

Does this mean that a patent will be harder to obtain now? Will the patent offices defer to a familiar refrain of “that’s obvious, you could just ask Claude!”. In some low-tech fields of R&D, maybe. More often than not, I suspect that it could actually make life more straightforward for those inventors who are working in complex technical fields.

One example of a transformative use of AI has been the development of Alphafold2 which is an algorithm that can predict protein structure based on information of the underlying amino acid sequence. The algorithm has been validated by multiple lines of evidence from structural biology experiments after Google DeepMind entered structure predictions from AlphaFold2 into the critical assessment of structure prediction (CASP) competition in 2020. The software reportedly outperformed all the other entrants by a wide margin and heralded a new era of AI powered structural bioinformatics, drug discovery and data driven diagnostics.

There is no question that structural biology was advanced by Alphafold2 and other similar developments. However, as with all algorithmic approaches that are applied to highly complex systems, there are limitations, particularly where there are deviations in underlying training data that reduce the confidence in a given prediction. The European Molecular Biology Laboratory (EMBL) mentions on their online training that “AlphaFold2 gets the fold and majority of the side chain positions right, but if you are interested in the features that require atomic precision, like understanding the fine details of catalysis, AlphaFold2 may not be able to match such requirements”. Even AI has its limitations.

 

Inventions can be found in complexity

The complexity of large multi-factorial systems, such as cell biology or weather prediction, where higher order interactions may generate unpredictability in outcome reinforces that predictive models may be useful but do not eliminate the need for real world experimental data. An invention will often lie in the areas where the projections 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 agentic AI trained to avoid radical deviation from normal design space. The skill of the inventor and their patent attorney is now in leveraging this divergence between an AI prediction and the actual real-world outcome in order to demonstrate an inventive step in their patent claims.

In conclusion, I don’t think the era of AI powered innovation has rendered future inventions obvious, but it has emphasised an even greater need to step out of the virtual world of predictive modelling and to invest in including real-world data in future patents.

 

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