The normalization and regulation of AI – IP protection now that AI has hit the mainstream

Of course, the first rule of writing any article is try not to include references that date it – when it comes to AI and its current frenetic development all I can say is good luck! Many an article about the latest “big bang” AI application in recent months has ended up looking old and out of date in days or weeks (this being a comment that I suspect will not date this particular article).

As I write this the European Parliament has passed the EU’s new AI Act, the inevitable corrigendum has been published. The official publication of the Act itself in the EU Official Journal is expected imminently and twenty days later it will be in force. AI is now firmly on the map of regulators around the world, with a handy template to follow no less – a sure sign of a mature technology! AI, or more specifically Machine Learning (ML) which is the field in which the practical innovation is currently focussed, manages to be an exciting, game changing, new technology, capable of feats that to all intents and purposes appear almost magical, whilst at the same time being a field that is over sixty years old.

A bigger bang
The pace in recent years has undoubtedly quickened. Our Boult AI team has certainly seen this first hand. As a firm that has always had a strong practice in the mathematical and computational arts there has always been a steady number of inventions coming across our desks which had an ML element. These innovations were historically produced by specialist outfits using ML to help chip away at specific problems as part of much larger systems. When I began my practice ten years ago things were already changing and now we see a torrent of ML based innovation, often where the ML is doing the heavy lifting in providing solutions to problems previously thought insoluble. Indeed, such is the reach of ML innovation today, we routinely see it coming from almost every technical field imaginable (leading to some quite interesting interdisciplinary teams, both of inventors and attorneys) – from fields that are traditionally very pro-patent (such as the life sciences) all the way to those that do not ordinarily seek patent protection.

Such diversity in innovation in the ML space is what has led us to build our AI team at Boult in the way we have over the years. It is clear that the knowledge requirements to fully understand the innovations are no longer simply related to the mathematical and computational arts, but often key input is needed from the specific fields of application themselves, whether that be biosciences, engineering, scientific instrumentation, aviation and so on. By having our team built up from all of these subject-matter expert who are also comfortable with the specific challenges ML innovation brings, we often end up mirroring the innovation teams that devised the ideas in the first place.

It is also the case that this torrent of innovation that we see[1], whilst reflecting the important role that ML is playing in the technological revolution we now find ourselves in, also indicates the relevance of the patent system, and the importance of intellectual property protection, to the innovators in this space. As patent attorneys we are, perhaps, predisposed to recognizing the value of patent protection – though the first issue our team tends to explore with a client on a new project is what the commercial rationale is for the patent – however the volume of interest in ML patents does provide us with some welcome external validation.

Can you keep a secret?|
External validation such as that is all very well but in some ways the idea people who have seemingly miraculous innovations wish to protect them is, perhaps, not all that surprising. Patents are of course only one way of protecting intellectual property. ML inventions often require large data sets and opaque (or hidden) training processes does provide some comfort to those innovators who are worried that the next big thing might just be about to walk out of the door and down the road to a competitor. Indeed, ML innovation has seen a corresponding rise in interest in Trade Secret protection, but this is not a silver bullet and should often be viewed as art of a more nuanced approach to IP protection. It is the case that some ML innovation is not so easily obfuscated from the end user. Also, the overheads of secrecy (if done well) can be onerous, and the biggest hurdle can often be – yes, it really does require you to keep a secret.

Patents on the other hand are often well placed to provide a happy balance, where one can (and indeed one is required to) positively disclose the great new idea, and in return one is granted a fairly strong monopoly. Given the pace of change the modest time limit of twenty years attached to this monopoly seems, in a lot of cases, a fair exchange.

Coming back to the EU and its Act, however, the value proposition becomes more compelling.

It is understandable that regulators are looking at AI innovation – any tool that is powerful can also prove problematic if it misfires or is applied with malicious intent. As with all regulation the trick is to work out where the risk is, and what tools you are going to use to mitigate it. In the EU AI Act it seems that for those applications that are deemed the highest level of risk (where the potential reward is not seen as enough) the answer may be an outright ban. This is not something that having a patent will save you from[2].

For everything else, however, it appears that there is sunlight – and that often goes hand in hand with the requirements for patent protection.

Transparency – through a glass darkly
Concerns around AI often tend to come down to the fact that a non-human is making decision which may affect a human. People tend to be keen to know on what basis the decision are being made, and who to blame when it all goes wrong. It’s clear that decision making, and in particular poor and biased decisions, are areas where those that drafted the AI Act foresee problems arising.

Bias is not inherently an AI issue and human society strives (imperfectly and sporadically) to reduce bias, prejudice, and inequality. AI systems, often being based on ultimately human influenced or generated data, are not immune to inheriting bias from their human creators. Over the last few years, society has begun to grapple with exactly how much these human prejudices, often with stark consequences, can find their way into AI systems.

The insurance industry presents an interesting example. In recent times, legislation has prevented the use of gender data in making decisions on insurance premiums, owing to the obvious potential for discrimination. At face level, simply discounting the gender dataset from a wider data pool is straightforward – it is simply hidden in the wider data and researchers discount it when determining premiums. For AI, however, mining a full data set and finding hidden datapoints to form a decision is the entire reason it would be employed to perform the task in the first place. If certain data which could result in a discriminatory result, even if hidden from a human researcher, finds its way into an AI model, then it will naturally form the very same biases that the datapoints were hidden to prevent.

Some EU parliamentarians were perhaps aware of the cautionary tale of The Meta chatbot BlenderBot 3 – which was already using what Meta termed “problematic or offensive language” within only a few days of being live. There have also been concerns around the accuracy of Google Bard and its potential susceptibility concentrated efforts to game its response from elements online.

This may all seem somewhat esoteric if what you are doing is using AI to detect if your sump pump is cavitation. The EU Act is clear that it’s transparency all the way down, with even those with systems that are categorized as “low risk” being asked to sign up to a code of conduct encouraging them to apply the “high risk” transparency standards. For consumer facing enterprises there may also end up being an expectation from customers to a certain level of transparency.

There are already teams like Oxford University’s CapAI team that are trying to put together systems that effect external validation of peoples AI models and tools. Of course, anyone who has ever played Cluedo has encountered situations where just one seemingly innocuous piece of data being revealed may be all someone needs to complete the picture.

Whilst it is not yet clear how the dice will fall, as the devil surely lies in the detail of how the obligations under the Act are applied in practice over the coming months and years, innovators may find themselves having to give away more information about the operation and training of their systems than they would otherwise choose to do. Against this backdrop it is perhaps clear why patents are increasingly being seen as such a good fit for those who wish to protect their innovation in this space.

Pulling it off
Having said all that it is reasonably lucky that patent offices around the world are receptive to this new technology – as the strategies only work if you can get the protection. The EPO, which often ends up setting the mood music on issues, has been vocal in its acceptance that AI innovation is key to the “fourth industrial revolution”. Indeed, they have taken to heart what is often an inventor’s mantra –  “innovate or die” – realizing perhaps that if they are not going to be part of the solution in protecting innovation in this key area they may end up an irrelevance.

That is not to say that EPO has become a “soft touch” regarding ML innovation, even despite these positive noises. The process, and the standards applied, are still rigorous, and careful consideration always needs to be given to what the patent will need to look like to be acceptable. This is often where we find our many years of experience handling prosecuting such cases are invaluable – if you don’t know what acceptable looks like it can prove difficult to get there. Patent drafting and prosecution are the key battlegrounds here and are really where a lot of the overall strategies succeed or fail – a strategy based around unachievable aims will not survive contact with reality.

As such, the role of the patent attorney often has to be one that helps guide the overall protection strategy. Whilst we are not, alas, omniscient, to give the best advice we nevertheless need to understand the client’s position and where they are being pulled, whether from commercial realities (such where the value lies), regulatory realities (such as where the transparency may be unavoidable), or practical realities (such as what protection, or level of secrecy, can be achieved). In our team we take this role seriously, and try to make sure that an overall strategy is considered from the outset. It is perhaps notable that in having a team drawn from numerous technical disciplines this also means we end up with a synthesised knowledge of how many different industries try and approach the issues – something that has led to quite imaginative approaches!

In other words, the key, often, is knowing what is possible and being happy in using a variety of approaches to achieve the protection an innovator needs – for humans this sort of thing can be a team sport. Indeed, with all of the considerations touched on above, regulatory requirements, interplay with trade secrets, commercial realities for the business, and quirks in all of these due to the particular field of applications, relying on the accumulated training data of so many professionals is a bit of an ML-like solution.

Of course, in many ways the story of ML is that of trying (and in places succeeding) to replace teams of experts with the machine. When it comes to protecting that ML-innovation we are confident that our team of experts still seem to have the edge.

[1] Likely enabled by the continual improvements in computing power, and the ability to gather truly huge amounts of data on anything and everything!

[2] This does, however, raise some interesting issues of how offices such as the European Patent office (which ultimately also grants patents in jurisdictions not in the EU) will treat patent applications for such subject-matter – something which even if you are not operating under an EU jurisdiction, your patent attorney will need to consider for the filing strategy.