As the use of generative AI becomes ubiquitous, is the patent system ready for an influx of AI generated inventions? And how might it handle AI inventions which are beyond our understanding?
It seems AI is everywhere. If not now, then soon. The patent world is no different. Patent offices are already using AI to improve the subject classification of applications and searches for prior art documents. Patent attorneys are looking to AI to assist with drafting specifications and responses to examiners. As for inventors, the thought of using AI both to assist with the inventive process and to reduce the cost of preparing a patent specification is a tempting prospect.
Earlier generations of AI used deep-learning to discover patterns in existing data and led to some notable (and in the case of DeepMind’s protein structure predictor Alphafold, Nobel-prize-winning) inventions. More recently, generative AI, powered by large-language models (LLMs) and embodied in chatbots such as ChatGPT, Gemini, Claude and others, has taken matters in a more creative direction.
When prompted, LLMs are characterised by fluent, persuasive output, capable of passing the Turing test, confounding users as to whether they are actually interacting with a human. But they can also be notoriously sycophantic and prone to hallucination. In short, there is a fundamental problem: AIs make things up. Convincingly. For an inventor using AI, will it be clear when invention has crossed the line into fantasy?
Almost every patent attorney will at some time in their career receive a call or an email from an inventor who believes they have made a groundbreaking invention. The world’s energy crisis is solved. Interstellar space travel is possible. Machines, once set in motion, operate forever, generating limitless energy with every turn. Often these ‘inventions’ are easily shown to be nothing of the kind; some variant of a perpetual motion machine, violating conservation laws and exhibiting a misunderstanding of basic mechanics or thermodynamics. Others misinterpret more esoteric concepts such as quantum mechanics and relativity.
Occasionally, however, the situation is not so clear cut. Inventors may insist that the accepted laws are incomplete or wrong. And in truth many modern inventions, including much of modern microelectronics, would once have been considered to verge on the magical, contravening the science of earlier times. Might an AI come up with an invention which relies on an incompleteness in an existing physical law or even postulate an as-yet unknown one? If it did, could we understand it? And how would the patent system handle a patent application for such an invention? Whilst the building blocks of LLMs are understood (at least by those who develop them), they are essentially a ‘black box’, with the precise reasoning by which they arrive at much of their output remaining mysterious. How then to differentiate hallucination from insight?
Patent systems long ago formalised their rejections of so-called perpetual motion inventions, as being neither capable of industrial application nor sufficiently disclosed so that they may be carried out. But in most cases these ‘inventions’ were relatively straightforward, in concept if not in detail. Now, as AIs become increasingly powerful, it is not beyond the realms of possibility that at some point an AI will make a conceptual leap to an invention which is at odds with present science. What then?
Patent examiners, at least in the UK, have something of an existing framework to follow. Whilst it predates our current AI era and so does not address them directly, it can be pressed into service to act as a bulwark, sufficient until something better comes along.
The seminal case, albeit an imperfect one, dates from nearly two decades ago, when a US company Blacklight Power were pursuing a pair of patent applications. Both applications claimed inventions which relied on a purported new species of hydrogen. Unlike standard hydrogen, this species required the sole electron to exist in an energy state lower than the lowest possible one as recognised in standard physics. Such “hydrinos” were part of a sweeping new theory – the “Grand Unified Theory of Classical Quantum Mechanics” – proposed by Blacklight’s founder and CEO.
The initial patent examiners were unconvinced, refusing both applications. On appeal, a senior examiner was more circumspect, admitting that his understanding of physics was a long way short of what would be necessary to assess the theory on its own merits or to evaluate the voluminous supporting evidence which had been provided by Blacklight. What was clear, however, was that the relevant scientific community had not taken to the new theory; ever since it was first proposed in the early 1990s it had been studiously ignored – which suggested the theory was probably wrong. A further appeal to the high court clarified what has become the current approach.
Patent applications should only be refused if they are “clearly contrary” to well-established laws, not merely probably wrong. Otherwise, providing there is credible evidence that, on the balance of probabilities, there is a “substantial doubt about an issue of fact” which could lead to patentability – and a “reasonable prospect” of the new theory being proved correct when investigated in detail at at a full trial, with expert witness – then the application should be allowed to proceed.
The question of whether, and how much, benefit of the doubt should be afforded the applicant is a finely balanced one. On the one hand, it would be unfair to the applicant if a patent application was refused but the theory turned out to be true; on the other, “it would be completely wrong and against public interest to bestow upon misleading applications the rights and privileges of a granted patent”. The patent system is inherently scientifically conservative, but there is an acknowledgement that there is the danger of of refusing an application which depends on a disputed theory which may subsequently turn out to be correct. The application stage is necessarily “an imperfect tribunal of fact”. Patent examination is not a peer review process. At the application stage, the applicant may need to be given the benefit of the doubt because an incorrect refusal cannot be remedied at a later stage. Only if an invention is required to “operate in a manner clearly contrary to well-established physical laws” is the patent application to be refused from the outset.
In Europe, the EPO has trodden a somewhat similar path, rejecting patent applications which are deemed incompatible with with the generally accepted laws of physics, and yet keeping the door ajar for “revolutionary” inventions which seem, at least at first, to “offend against the generally accepted laws of physics and established theories”. The focus is on practicalities rather than theoretical considerations. To obtain a patent based on such an invention requires the applicant to provide a description “detailed enough to prove to a skilled person conversant with mainstream science and technology that the invention was indeed feasible”.
For Blacklight, the benefit of doubt was not enough. The evidence was unconvincing and the scientific community uninterested. Both applications were referred back to the patent office and finally refused.
The Blacklight cases provide a salutary lesson for both inventors and patent attorneys. We may be entering a new era of AI-assisted inventions, and the patent system may be willing in principle to entertain the idea of inventions which push against or even cross the limits of existing science, but credible evidence is the key. Where examiners may not understand every nuance of an invention, they will look to experimental evidence and in particular whether the new theory has been accepted in the wider scientific community as proxies for assessing “substantial doubt” and “reasonable prospect”. For patent applicants it is important not be to swept along by AI pronouncements. However convincingly an AI may present a seemingly revolutionary invention, and even if no-one seems capable of understanding it, extraordinary claims will always require extraordinary evidence. Whether it will even be possible to collect such evidence if an AI invention is truly beyond our understanding may prove a defining challenge for the patent system.
