We could be approaching a point where there are many more inventions that do not meet patent law’s inventive step requirement

By Sharaz Gill

Anyone who has been following my series of articles on the impact of AI on patent law and licensing will know that much of my navel-gazing is prompted by reading about new applications or developments in the space. My latest offering is no different. I recently came across an article in the Singularity Hub about the Jacobian Conjecture, a mathematical problem that has stubbornly resisted solution for almost 90 years. The mathematics is beyond me, but it was not the conjecture itself that caught my eye; it was the role that AI had apparently played in finding the long-sought-after counterexample.

Mathematician Levent Alpöge had been working with Claude Fable 5 to explore the problem. What I found interesting was not that the AI had disproved the conjecture or whether Alpöge or the machine deserved the credit. It was rather the interaction between them. The counterexample identified by Alpöge was apparently remarkably simple; the real difficulty lay in finding it among an enormous number of possible polynomial mappings. AI had helped a highly skilled mathematician to navigate that search space and find something that had eluded others for decades. It had, in a meaningful sense, extended the reach of an already exceptional scholar.

There is a potential terminology trap here, particularly for a patent lawyer. Problem-solving ability, innovation and inventiveness are not the same thing. Alpöge’s mathematical discovery demonstrates remarkable problem-solving ability, but it is not an invention in the patent law sense. Equally, something can be innovative without necessarily involving an inventive step. I will inevitably use some of these terms rather loosely in what follows, but the distinctions become important when we ask what AI is actually doing to inventiveness as patent law understands it.

In any event, that brought me back to an argument I made in my last article for Sisvel Insights, in which I considered what happens to inventive step when AI becomes an ordinary tool of scientific and technical research. Patent law already assumes that the skilled person has access to the ordinary tools of the relevant field. If AI is one of those tools, I suggested that it becomes difficult to place it in the hands of the real inventor while denying it to the hypothetical skilled person.

The Jacobian Conjecture story made me wonder whether there is a further consequence: if AI amplifies the problem-solving capabilities of individual researchers, should we expect it to lead to more inventions? Scientists and engineers would be able to explore more possible solutions, test more hypotheses and make connections that might otherwise have remained undiscovered. AI does not need to become an autonomous inventor for the rate of invention to increase dramatically; it merely needs to make human inventors better at finding new solutions.

Now, I am fond of a good paradox and there is certainly one here. If the same AI capabilities become part of the ordinary toolkit of the skilled person, the threshold for inventive step must rise. As a result, a solution that once required considerable ingenuity may become relatively straightforward to find with AI assistance.

We could therefore be heading towards a world in which there are many more inventions, but fewer of them are inventive in the patent law sense.

AI: be all you can be … and a bit more

The most immediate effect of AI on invention is likely to be an increase in the number of inventions. This is partly because AI accelerates existing research processes. However, speed alone does not capture what is changing. More importantly, AI allows researchers to explore a much wider range of possible solutions than would previously have been practical.

A conventional researcher faced with a technical problem has finite time and resources. There may be hundreds or thousands of possible approaches, but only a small number can realistically be investigated. Decisions must thus be made about where to look and which avenues are worth pursuing. Some possibilities will never be considered at all.

AI changes those constraints. It can generate and compare large numbers of candidate solutions, search across different bodies of knowledge, suggest connections between apparently unrelated fields and repeatedly refine its proposals in response to new information. Approaches that might previously have been dismissed as too remote, or that simply may never have occurred to the researcher, can now be investigated fully. The result is potentially more than enhanced research productivity: AI expands the effective intellectual search space available to the researcher.

This does not mean that the human contribution disappears. Quite the opposite may be true. The utility of an AI system depends heavily on the person using it. The researcher must still identify the problem, identify relevant information and constraints, assess the proposals that are produced, recognise promising results and decide which paths merit further investigation. This requires key expertise. As the capabilities of the AI increase, the value of these human contributions will likely increase with them.

If this effect can be reproduced across science and engineering, AI should boost the rate of invention, perhaps dramatically. This creates the other side of the paradox: if AI can amplify the capabilities of the actual researcher in this way, how much of that same amplification should patent law attribute to the person skilled in the art?

Moving the inventive goalposts

There is an obvious difficulty with this amplification of human inventiveness. Patent law does not measure inventive step against what the inventor could have achieved without assistance. It asks what would have been obvious to the person skilled in the art at the priority date.

As I argued in my previous article, the skilled person cannot sensibly be isolated from the ordinary tools used in the relevant field. If AI becomes part of the standard research environment, it must eventually form part of the capabilities attributed to the skilled person as well, in the same way as scientific databases, simulation software and computational modelling already do.

Moreover, inventive step is not a fixed threshold. What requires ingenuity depends in part upon what an ordinarily skilled person can do with the tools that are generally available at the relevant time. Imagine, for example, a technical problem for which the prior art contains all the information necessary to reach a particular solution, but the connections between the relevant pieces of information are difficult to see. A skilled person working conventionally might never make those connections. If an AI system routinely available in that field can identify them, propose the resulting solution and explain why it is worth pursuing, it becomes much harder to say that reaching the solution involved an inventive step.

This produces a ratchet effect: as AI systems become more capable, problems that once required exceptional ingenuity may fall within the capabilities of the ordinarily skilled person. The better AI becomes at exploring technical possibilities, the higher the threshold of inventive step may climb.

This effect will not be uniform. Different technologies will adopt AI at different rates and use it for different purposes. The assessment will remain tied to the priority date and to the capabilities and practices of the skilled person in the relevant technical field. An AI functionality that is routine in computational chemistry may, at the same date, be unusual in telecommunications.

If AI makes actual researchers better able to find technical solutions, and those capabilities become ordinary within their field, patent law must eventually reflect this in its assessment of obviousness. AI may therefore cause the number of inventions to increase while reducing the proportion that involve an inventive step. Thus, the technology that facilitates invention also raises the standard by which inventions are judged.

This prompts a further question: if increasingly capable AI systems are doing more of the work involved in identifying the technical solution, does human inventive skill disappear from the process or does it simply move somewhere else?

So, where’s the clever bit?

As AI becomes more capable, human ingenuity does not necessarily disappear from the inventive process. Perhaps some of it will move to an earlier stage.

Take two researchers using the same AI system on the same technical problem. One asks it to propose solutions and receives nothing remarkable. The other identifies an overlooked constraint, asks the AI to approach the problem on that basis, follows a connection suggested by its response, changes the parameters and eventually obtains the solution that becomes the invention.

The AI may have generated the final solution, but it would be difficult to say that the two researchers had made the same contributions. The second researcher brought technical insight to the process. Without that insight, the AI might never have been directed towards the invention.

This goes beyond simple ‘prompt engineering’. The important contribution is not simply knowing how to phrase a request, but knowing what to ask the AI to investigate, what technical context and constraints to provide, and which of its responses are worth pursuing. That requires technical understanding as well as skill in using the AI. This suggests that some of the ingenuity involved in invention could move upstream. The researcher may contribute less by personally conceiving the final solution and more by directing the process through which it is found.

Patent law cannot simply treat a clever prompt as conferring an inventive step because inventive step attaches to the claimed invention and is assessed objectively. Ingenuity in reaching an otherwise obvious solution does not make it less obvious.

There may nevertheless be cases where the human contribution also tells us something about obviousness. If the AI found the solution only because the researcher identified an overlooked constraint, reformulated the problem or made a connection that the skilled person would not have made, this may help to explain why the claimed solution was not obvious. Prompts and responses may also provide unusually good evidence of this process, showing how the researcher framed the problem, changed constraints, rejected false leads and steered the AI towards a result.

The significance of prompting is therefore likely to change over time. What initially requires unusual skill may become ordinary competence as researchers become better at directing AI. Increasingly, AI systems will also perform some of these tasks themselves.

Prompting can provide a possible location for human ingenuity only if it requires something beyond ordinary skill. Its significance will depend on what the skilled person could ordinarily do with the AI tools available at the priority date. This makes inventive step more fact dependent. It may be necessary to understand which AI systems were available at the priority date, how they were used in the relevant field and what they could do.

Navigating the Library of Babel

AI may have a particular impact where inventions involve choosing between many possible solutions.

In my previous article, I considered the EPO’s ‘one-way street’ and the ‘obvious to try’ analysis applied by the English courts. Where the prior art points towards a particular course and the skilled person would pursue it with a reasonable expectation of success, inventive step becomes difficult to establish.

The position differs where there are many possible directions and no obvious basis to choose between them. A pharmaceutical researcher may face thousands of compounds; an engineer, many combinations of parameters or configurations. The claimed solution being somewhere among them does not make it obvious to select.

My attempts to re-read Borges’s “The Library of Babel” (this time in Spanish) while writing this article yielded an unexpectedly good analogy. Borges imagines a library containing every book that could possibly be written. Somewhere within it lies every truth, surrounded by an effectively limitless quantity of nonsense. Having the answer somewhere in the library is of little use if there is no practical way of finding it. There is something similar about many technical search problems. The potential solution may already be there. What matters is whether the skilled person can find it.

AI could change the practical significance of this problem. Suppose that there are 100,000 possible approaches to solving a technical problem. It may be unrealistic for a human research team to assess more than a small proportion of them. An AI system, however, may be able to analyse the possibilities, eliminate those that appear unpromising and identify a much smaller number for further investigation.

The number of possible solutions has not changed. What has changed is the difficulty of finding the promising ones. This is what makes the Jacobian Conjecture example so interesting. The counterexample was not apparently difficult to understand once it had been found. The challenge was finding it within an enormous field of possibilities. AI appears to have changed the search rather than the solution.

This could considerably expand the range of research that is regarded as routine. A search that once would have been impractical because of the number of possible alternatives may become a relatively straightforward exercise. The skilled person need not know in advance which of the 100,000 possibilities will work if an ordinarily available AI system can reduce them to a shortlist that can readily be tested.

There are limits to this argument. The fact that an AI system can be made to identify the claimed solution does not establish that the solution was obvious at the priority date. Much will depend on the capabilities of the AI systems then ordinarily available in the relevant field. If finding the solution involved an unusual model, an unconventional way of using it or considerable technical insight in directing the search, the route may still be far from routine.

Nor does identifying a candidate necessarily provide a reasonable expectation of success. An AI may be very good at generating plausible possibilities without being equally good at predicting which of them will work. Experimental uncertainty does not disappear merely because the initial search becomes easier.

Even with these qualifications, AI could alter the practical meaning of ‘obvious to try’. In some fields, the issue has never been a shortage of possible solutions; it has been deciding which of them are worth trying. As AI becomes better at making that choice, technical territory that once required inventive exploration may increasingly become searchable.

Levelling up the problem

So far, we have largely assumed that increasingly capable AI will continue to be applied to the same kinds of technical problems. In practice, however, that seems unlikely. If AI makes it easier to solve today’s challenging problems, it follows that researchers will use it to tackle harder ones, including areas previously considered too complex, uncertain or expensive to investigate. The threshold for inventive step may rise, but the difficulty of the problems being addressed may rise with it.

There will therefore still be a frontier beyond what the skilled person, equipped with the ordinary AI tools of the time, can readily achieve. At that frontier, AI may be asked to tackle problems for which there is no known route to a solution at present. It may have to generate new hypotheses, make unexpected technical connections or develop approaches that were not apparent from the existing state of knowledge.

Suppose that a researcher poses such a problem in a conventional way and the AI produces a solution that neither the researcher nor an ordinarily skilled person using the same AI would have been expected to find. There is no obvious reason why such a solution should fail the test for inventive step. The fact that the AI has done something beyond what the skilled person, equipped with ordinary AI capabilities, could have been expected to achieve may support the conclusion that the invention was not obvious.

The issue of inventorship then becomes trickier, however. If the researcher did no more than pose the problem, recognise the value of the answer and verify that it worked, it may be difficult to identify a human contribution to the conception of the invention. This process may accelerate further if AI systems increasingly participate in improving the tools themselves, as current work on recursive self-improvement would suggest.

Patent law may therefore encounter an uncomfortable category of invention: a technical solution that is beyond the ordinary capabilities of the AI-augmented skilled person, with no human who can readily be said to have conceived it. This takes us back to AI inventorship by a rather more exotic route. Patent law would then face the curious prospect of recognising an inventive step without being able to identify a human who took it.

O brave new world, that has such patents in it

AI thus presents patent law with yet another curious paradox: it may make us considerably better at inventing while raising the standard that determines which of those inventions deserve patents.

The frontier of inventive activity will not disappear. As today’s difficult problems become easier, researchers will likely move on to harder ones. The boundary between ordinary and inventive activity will have to move with them.

There is also a wider consequence. AI is making it easier not only to invent, but also to turn inventions into patent applications. If applications increase dramatically while the proportion that satisfy the inventive step requirement falls, patent offices could face a very different problem. I will return to that little conundrum in my next article.

Sharaz Gill is Head of Portfolio Management at Sisvel

The opinions expressed within this article are the author’s and do not necessarily reflect the views of Sisvel. The content is for informational purposes and should not be taken as legal advice.