AI report

2023

Cutting edge insights from the leading intellectual property AI experts

Introduction

For those of us that have been involved with Artificial Intelligence (“AI”) for some time, it has been fascinating ​how quickly it entered public consciousness and debate after the release of Large Language Models such as ​ChatGPT. For those that have been claiming that AI already had a “hype” issue, 2023 must have come as ​something of a shock. Of course, it is too early to see whether the increase in attention that AI has received ​of late will translate into more patent filings at the European Patent Office (“EPO”), but it will certainly be ​interesting to monitor this in future reports.

The strong growth in AI patent filings in Europe has continued, with the number of AI publications at the EPO ​increasing by around 17% between 2021 and 2022. Whilst US applicants are still filing more AI applications ​at the EPO than applicants from any other country, the dominance of the US has decreased over the last four ​years, as other countries, such as China, have begun to increase their AI filing numbers. Whilst the strong ​growth in AI filings from Korea that we noted in our previous report has tailed off slightly, Korea still remains ​the largest filer of AI applications at the EPO per capita.

In our last report we predicted that AI techniques had matured to the point that we would start to see more ​applications in the physical economy and from our latest analysis this appears to be borne out. Life and ​medical sciences, telecommunications and physical sciences remain the strongest industry sectors for AI ​patent filings at the EPO. The percentage of AI applications relating to computer vision has shown significant ​growth over the last five years, demonstrating the significant investment that is being applied in this area. On ​the flip side, the relative number of speech processing applications has continued to decline.

To us, one of the most exciting areas of innovation is in Medical Technology (“MedTech”) and in this report ​we have a whole section devoted to trends in AI patents applications in this space. Year-on-year growth in ​the number of AI MedTech publications surpassed that for AI overall each year from 2014 to 2021. This ​shows MedTech to be a key area of development in the AI space. Of the MedTech AI applications published ​at the EPO in 2022, nearly 50% related to medical imaging or diagnostics. This demonstrates that MedTech ​AI innovation appears currently to be focussed less on treatment, and more on prediction, prevention, and ​early diagnosis and intervention.

Overall, our analysis shows strong growth in both filing numbers and allowance rates for AI applications at the ​EPO. This reflects not only significant investment in AI from industry, but also an increasing ​acknowledgement of the practical applications of AI by the EPO.

Methodology

As in previous years, the data analysed in this report starts with the International Patent Classification (“IPC”) code and keyword definitions used for patent data in the “WIPO Technology Trends 2019: Artificial Intelligence” report (as defined in the “Data collection method and clustering scheme: Background paper” for the same report). Cases matching the definitions used for the World Intellectual Property Organization (“WIPO”) report were identified using the Derwent Innovation database, and data from Derwent Innovation was combined with data from European Patent Bulletin. The WIPO definitions were refined based upon manual analysis of the data. We then wrote custom formulae using the raw data to generate our own fields for the analysis.

In a change to this year’s report, we have based some of our analyses on publication date, rather than filing date, to reduce inconsistencies from year-to-year. While performing this year’s analysis, we identified that divisional applications were creating inter-year inconsistencies, due to being accorded the same filing date as their parent applications. It should be noted that further inter-year inconsistencies arise from IPC codes being adjusted on grant or otherwise, which can change the applications that are analysed. We do not try to account for IPC code changes, but do show the filing trends over time separately in each of our reports, so that each report can stand alone.

Can AI be an Inventor? DABUS update

The question as to whether an AI algorithm that develops a patentable invention can be deemed an “inventor” continues to attract interest - even more so this year, given the rapid development and adoption of generative AI methodologies. This issue has been considered in depth in various jurisdictions in relation to a set of patent applications which list an AI machine – referred to as “DABUS” – as an inventor. These cases were discussed in detail in our previous AI Report. Here, we give a brief update on some of the more recent developments.

As reported previously, the UK Court of Appeal decision[1] in September 2021 determined that an AI-based machine is not a person and cannot be an inventor for the purposes of the relevant UK Patents Act. Leave to appeal this decision to the Supreme Court was granted in August 2022. It is interesting that this further appeal was allowed – this indicates perhaps the important implications, given the increasing use of AI systems as tools in such fields as drug discovery. The case was heard in March 2023, and a decision is now awaited.

In the US, the Court of Appeals for the Federal Circuit (“CAFC”) also came to the view that an AI software system cannot be an "inventor" in a decision issued in August 2022, with the US Supreme Court recently denying an appeal in April 2023.

In July 2022, the EPO Legal Board of Appeal issued its written decision[2] in the case, confirming that a machine is not an inventor within the meaning of the EPC. A request that two questions be referred to the EPO Enlarged Board of Appeal (the highest legal judicial authority of the EPO) was also considered not to be warranted by the Legal Board of Appeal.

Given these further decisions, we can conclude that, for now, AI algorithms should not be listed as an inventor on a patent application. Interestingly, the EPO Legal Board of Appeal did note that they were “not aware of any case law which would prevent the user or the owner of a device involved in an inventive activity to designate himself as inventor under European patent law”. In doing so, the Board appeared to acknowledge that patent protection might be obtainable by specifying the user or owner of the AI as the inventor, rather than the AI itself.

It should be noted that these decisions are based on the law as it currently stands. If changes to the law are required, this may be a question for legislative bodies, rather than the courts. We may well see changes to the law in this area in future. For instance, the USPTO recently requested comments regarding AI and Inventorship, indicating that the law in this area continues to be considered. It may be that in time, patent laws are reviewed, as the importance of AI generated innovation increases.

1. Thaler v Comptroller General [2021] EWCA Civ 1374

2. J 8/20

1: General

Publication trend

1.1

The growth in AI patent filings in Europe continues, with Figure 1.1.1 below showing the yearly publications for AI patent applications at the EPO continuing to increase into 2022. This growth appears to be easing, with the yearly number of AI publications increasing by around 17% between 2021 and 2022 compared to 19% in 2021.

This rate of growth for AI patent filings continues to significantly outstrip the general filing trend for European applications regardless of technology, which in fact saw a decrease of 0.6% in the number of patent applications filed in 2020 compared to 2019 [3] - likely due to the COVID-19 pandemic. While this trend has since reversed (total patent filings at the EPO increased by 4.5% in 2021 and by 2.5% in 2022), the number of AI publications shows an even stronger growth rate.

3. Source: EPO Patent Index 20222. J 8/20

1.2

Pending applications

Over the last few years, the EPO has attempted to increase the speed of examination and therefore reduce the average age of applications at closure[4]. Figure 1.2.1 below shows the progress the EPO has made in this respect - the average age of AI applications decreased to just over 4 years in 2021. However, the rate of decrease appears to be slowing, likely reflecting the EPO now approaching a stable rate of examination. Similarly, the rate of increase in the number of closures also appears to be levelling off.

Figure 1.2.2 below shows the number of pending AI applications for any given year, calculated as the difference between the total number of applications filed (since the start of 2000) and the number of applications closed each year. The number of pending AI applications continues to increase into 2020 (note that, due to the 18-month delay in publishing newly filed applications, we are only able to assess the number of pending applications up until 2020) - the examination load on the EPO in this area continues to increase year on year. Whilst the EPO has made headway in reducing the average age of applications at closure, it will be interesting to see if the significant growth in AI filings begins to cause this to slip.

4. Closure is defined as any act that causes an application to no longer be pending, such as grant, withdrawal or refusal

1.3

Filings by Country

In this section, we look at the publication trends of AI applications at the EPO for applicants based in different countries.

Figure 1.3.1 shows the relative proportion of applications published by applicant country. With the exception of Republic of Korea (“KR”), each country we analysed has seen a continued increase in the number of European patent applications in the AI space between 2021 and 2022. Since 2015, the US has been, by a large margin, the country with the most applicants filing AI applications at the EPO. Having said this, there has been a slight decline in the proportion of AI applications that originate from the US over recent years, as other countries (and in particular China) have begun to increase their filing numbers.

If we consider the member states of the European Patent Convention (“EPC”) as a block (“EP”) [5], we can see that it has been a close race between US and EP applicants since 2015. However, even taking EP member states together, US applicants are filing more AI applications at the EPO in most years.

Between 2015 and 2021, growth in the number of Chinese (“CN”) applicants has outstripped that of other countries. Although growth in the number of Chinese applicants slowed a little between 2021 and 2022, growth was also slower in Japan (“JP”), Republic of Korea (“KR”) and the rest of the world (“RoW”). There has been a slight decline in the dominance of EP and US applicants since 2015; although there are some signs to suggest that this trend has reversed since 2020.

Once again, the performance of EP applicants looks relatively weak, particularly considering that the EPO is arguably the “home” patent office of these applicants. This is particularly the case since 2015 when, per capita, EP, JP and KR applicants were filing similar numbers of applications.

The recent advances in AI show no signs of slowing down and so we anticipate continued growth in applicants from all countries. However, there are a number of global trends which have the potential to alter the relative filings between countries over the coming years. Different rates of population growth (and decline), on- and near-shoring, government AI policy and trade disputes all seem likely to affect the numbers in this section in ways that likely won’t be clear for a number of years.

5. We define European applicants as applicants from member states of the European Patent Convention (“EPC”). This is a larger group than just EU member states and for example includes the United Kingdom

1.4

Industry data

In this section, we look at how the AI patent applications break down by technology sector. Whilst some inventions are directed solely to improvements in AI methods themselves, most patent applications are directed towards the application of AI in a specific field.

Figure 1.4.1 below shows the breakdown of applications published over the last 22 years based on technology sector [6]. The largest technology sectors were Life & Medical Science, Telecommunications and Physical Sciences.

As shown by Figure 1.4.2 below, most sectors continued to see increasing numbers of AI applications since our last AI report. Interestingly, the decrease in the number of publications in 2021 in the Transportation sector reported in our previous report has not continued, with this sector seeing an increase in publication numbers from 2021 to 2022. It therefore seems that the decrease in 2021 may have been an anomaly, with Transportation remaining a key sector in which applicants are seeking to protect their AI innovations. The Life and Medical Sciences sector continues to see a strong growth rate, and it seems likely that this sector will remain the largest sector for AI patent filings at the EPO for some years to come.

The allowance rate between different technology sectors varied widely, from 23% for Business to 62% for Transportation. As in previous years, we see that the categories with higher allowance rates tend to fall into areas that the EPO considers “technical”, whereas the categories with lower allowance rates tend to be considered “non-technical”, such as Banking & Finance, or Business. The following categories each had an above average allowance rate (shown in yellow in Figure 1.4.3 below): transportation, agriculture, energy management, telecommunications, security, life and medical science, physical sciences, arts and humanities, cartography, military and law and social science). The categories with a below average allowance rate (shown in blue) were: entertainment, personal computing, education, networks, document management, publishing, industry and manufacturing, computing in government, banking and finance, and business.

It is noted that “Networks” had a relatively low acceptance rate. This sector includes applications relating to social networks, which are likely to be considered “non-technical” by the EPO. “Arts & Humanities” on the other hand includes applications related to music, and this probably contributes to the relatively high acceptance rate for this category – digital audio enhancement is considered a technical application by the EPO. The Industry and Manufacturing sector also had a relatively low acceptance rate, which may be due to a large number of AI applications relating to planning and scheduling in this field – such aspects are more likely to be deemed “non-technical” by the EPO. However, this sector appears to be seeing an increasing allowance rate, which may reflect a change in the type of technology being protected in this sector, moving towards areas that the EPO considers more technical. It is also notable that even in sectors such as Banking and Finance, around 20% of applications are allowed – demonstrating that it is possible to obtain grant of a patent in these areas in Europe.

Figure 1.4.4 shows changes in the industries which applicants of different countries have focussed on since 2010.

We see that 2018-2020 marked something of a high watermark for applications in the Transportation industry, with relative declines in most countries since then (although as indicated above, growth has resumed in 2021-2022). However, applying AI to transportation continues to be a key focus of the Japanese automotive industry and still accounts for the largest proportion of applications from Japanese applicants. Despite it being widely reported that progress is harder than expected in this area, this is a signal that Japanese industry is still devoting significant resources to advancing the use of AI in Transport.

As reported above, Life and Medical science is the area that has the most AI filings overall, but there are clearly some countries that are bucking this trend. In 2010, applicants from the Republic of Korea filed proportionally fewer Life and Medical science AI applications at the EPO than any other country. However, in 2021 the proportion of applications made by KR applicants saw an increase for the first time since 2016.

Trends for EP and US applicants are subtler than those of other countries, likely an artefact of the much larger absolute numbers of applications. Still, a greater proportion of applications from EP applicants are directed to life sciences, while relative proportions in the US are broadly flat over the period in question.

6. Based on classifications used in the WIPO Technology Trends report on Artificial Intelligence, with some refinement of definitions based upon manual analysis of the data set

Updates to the EPO Guidelines for Examination in 2023

The EPO Guidelines for Examination are updated on an annual basis. As covered in our previous report on AI filing trends, Section G-VII, 5.4.2 of the Guidelines was updated in 2022 with the inclusion of a new AI-based example (Example 5), demonstrating how inventive step is assessed by the EPO. Modifications have now been made to this new Example 5 in the 2023 update. This update is important since, by drawing analogies with the examples given in the EPO Guidelines, it can be possible to convince EPO Examiners of the technical character of an invention, and therefore to obtain grant of the patent application.

Example 5 relates to a method of thermal spray coating, where process parameters are adjusted automatically by a neuro-fuzzy controller (combining a neural-network and fuzzy logic rules). In the absence of prior art, the claim is considered allowable by the EPO in the example. However, Example 5 also introduces a prior art document, D1. D1 teaches process parameters adjusted automatically due to neural network analysis. Thus, the difference between the method of Example 5 and D1 concerns the use of a neuro-fuzzy controller combining a neural network and fuzzy logic rules.

This feature, of combining results of a neural network analysis and fuzzy logic rules, defines a mathematical method when taken on its own. However, together with the feature of adjusting the process parameters, it contributes to the control of the coating process. Control of a specific technical process is considered a technical application at the EPO. Thus, even though the difference over D1 is the use of a mathematical method, the technical effect required by the EPO is considered to be present since the mathematical method serves to control a technical process.

However, the Example 5 also considered a second prior art document D2. This disclosed a combination of a neural network and fuzzy logic rules providing a neuro-fuzzy controller in the technical field of control engineering.Based on this, claim 1 of Example 5 was considered to be obvious by the EPO. However, the EPO made some further comments in the Example 5 suggesting that, had the description and the claims provided further details, a different decision might have been reached, even in light of D2. In particular, the comments set out that “claim 1 does not contain any information about the coating properties to be achieved” and “No features of the neuro-fuzzy controller are linked to any technical properties of the spray coating”.

Following the 2023 update, the Example 5 now further clarifies these comments, stating that the “availability of the general teaching of using neuro-fuzzy controllers in the field of control engineering resulted in the objection that the controller of claim 1 was an obvious alternative. This particular objection could have been avoided if the claim had recited further features of the fuzzy control method linked to some technical properties of the spray coating process” and that “if the desirable coating properties resulted from specific input and output variables of the neuro-fuzzy controller, how the controller is trained or how the output is used in the regulation of the process parameters, these features would have had to be recited in the claim. The description and figures as filed could have provided evidence that the desirable coating properties are indeed achieved”.

These modifications to the Guidelines emphasize the importance at the EPO of providing a fully detailed description of AI inventions in the patent application, including details of the input and output to the model, the training process, and the use of the model output. Such details can be critical for obtaining grant of the application at the EPO.

It is noted that the sections of the Guidelines setting out the general EPO approach to Mathematical Methods (G-II, 3.3) and to Artificial intelligence and machine learning (G-II, 3.3.1) have not been modified this year, reflecting the more settled approach to AI inventions that the EPO now appears to have fully adopted.

1.5

Technology data

In this section we analyse the types of AI technologies featured in AI patent applications. Given that the EPO treats different AI technologies differently when considering patentability, it is interesting to see whether this treatment affects filing trends. While other factors, such as technological progress and market forces also have an impact, it seems clear that the EPO’s treatment of certain technologies does have an effect on filing strategies.

For example, the EPO’s position on the patentability of image or speech processing is much more favourable than that on text processing. This reflects EPO decisions that deem text and natural language to relate to “cognitive content”, in contrast to images or audio data, which are considered to reflect data dictated by the laws of physics.

Accordingly, we would expect to see lower allowance rates on average for applications classified as natural language processing than for applications classified as computer vision or speech processing. As shown in the next section, the data supports this expectation - the allowance rate for applications filed between the years 2016 and 2021 and classified as computer vision is 54%, whereas for natural language processing, the allowance rate is 38%.

It may be reasonable to expect, therefore, that filing trends would also be lower for applications directed to AI technology which exhibit a lower-than-average allowance rate. However, the data indicates that this is not necessarily the case. For example, Figure 1.5.1 shows that in 2022, a larger proportion of cases were classified as natural language processing (953) than were classified as speech processing (511), robotics (745), or control methods (601).

In our 2022 report, we suggested that one reason for the decrease in the proportion of speech processing applications could be that current speech processing methods sufficiently serve their purpose in industry. In contrast, developments in the field of natural language processing have been at the forefront of the AI industry in recent years, evidenced by the recent successes of Large Language Models, which likely explains the high proportion of applications directed to this technology, despite the EPO’s position on patentability and its relatively low allowance rate.

Previously, we also predicted that as computer vision techniques mature, we may expect to see a relative decrease in the number of computer vision applications. The data for 2022 indicates that no such shift is evident as of yet. In fact, the proportion of computer vision cases published between the years 2021 and 2022 has increased from 35% to 36%. Interestingly, the recent developments in language processing techniques have resulted in multi-modal models that are having an equally disruptive role in computer vision and image processing. Given this, it may be many years before such a shift occurs.

We noted in our last report that we expected to see increased use of AI in the physical economy being reflected in patent filing numbers. The proportion of robotics applications has been increasing since around 2017. Interestingly, however, after growing as a proportion since 2015, control method applications have seen the largest proportional decrease between the years 2021 and 2022, dropping from 11% to 8%. However, this may be a result of a transient shift in focus in the AI industry to natural language processing and computer vision. We may therefore expect to see a resumption of relative growth in control method applications in future. Notably, applications classified as control methods have a significantly higher allowance rate (72% for cases filed between 2016 and 2021) given their direct link to physical reality.

1.6

Allowance rates

The allowance rate for AI applications has generally increased since 2015, although with a small decrease in allowance rate between 2019 and 2021. For applications closed in 2021 and 2022 [7], the allowance rate for AI applications has once again assumed an upward trajectory, reaching a high of 54% in 2022. Interestingly, the pause in the improvements in allowance rate for AI applications partially overlaps a pause in the improvement of the overall allowance rate. More generally, however, the allowance rate for AI applications is consistently below the average for all applications at the EPO, which was 76% in 2022 [8] for example. The lower allowance rate for AI related applications likely reflects the additional difficulties associated with protecting computer-implemented inventions at the EPO.

While the allowance rate of AI applications seems to have been increasing when we look at close year, we see that allowance rates for applications by filing year appears to be decreasing since 2015.

This demonstrates an interesting change from the long-term average allowance rate by age shown in Figure 1.6.3 below.

Historically, applications that are closed relatively quickly have a higher allowance rate, which gradually declines as the pendency time increases. This likely reflects a tendency for applications with significant validity issues to undergo prolonged examination, and potentially appeals, before closure. Having said this, the allowance rate for very young applications (e.g. applications younger than 3 years old) is significantly below the average. This is likely due to an increased number of these applications being abandoned by their applicants (e.g. in response to significant objections raised by the EPO in the search report).

In Figure 1.6.4 above, we examine allowance rates across a number of AI technology designations. As the allowance rate tends to vary between each year, Figure 1.6.4 plots trends of allowance rate for each technology. Linear trends are used for clarity. Perhaps most interesting is that four of the five categories have allowance rates comfortably over 50% in 2022. Computer vision has seen a particularly strong increase in allowance rate since 2015.

Natural language processing (“NLP”) has long been considered by the EPO to be a “non-technical” application, although even here we can see a gently increasing allowance rate, hitting nearly 40% by 2022. This should be encouraging to the companies that are innovating in the area of Large Language Models.

7. Close year is defined as the year in which a case proceeds to grant, is withdrawn, abandoned, or refused

8. Source: https://www.mondaq.com/uk/patent/1167944/the-european-patent-office-the-story-in-numbers-part-2

Generative AI continues to generate controversy

Generative AI, particularly as an aid in the creation of scholarship and artistic works, has given rise to some ​newsworthy ethical and legal challenges. These have fallen into broadly three categories, namely:

1. The concern about the quality of the output.
2. Many generative AI systems have been trained on material created by or relating to third parties. This ​gives rise to data privacy and copyright infringement concerns.
3. Given that no human author has had a hand in its creation, can the creative output of generative AI merit ​copyright protection?

There have been attempts to address all these issues with high profile commentary in the media, proposed ​legislation and, particularly in the case of category 2 above, litigation.

Quality concerns

Numerous regulators and commentators have expressed concern that biases inherent in the training data ​could lead to the output being biased. The training data used to train LLMs can also be conflicting or ​incomplete, meaning that LLMs are also susceptible to “hallucinations”, whereby the text generated by the ​LLM is factually incorrect. In an extreme example, in July 2022, two New York lawyers were fined after ​submitting a legal brief with fictitious case law generated by ChatGPT. The judge concluded that parts of their ​brief were "gibberish" and "nonsensical".

Users therefore need to approach how and when they use these programs with a huge amount of caution; ​they are not simply a clever way of getting a robot to do your homework. Licensors will also need to consider ​what warranties they can give regarding efficacy and functionality, and how they should be limiting their ​liability in terms of use. Looking at the contractual relationship from the other end, users will need to ​scrutinise those terms of use carefully to see how much comeback they might have if things go drastically ​wrong.

Copyright infringement

Generative AI programs have been trawling through, drawing on and reproducing a huge amount of third ​party proprietary material and the scale and sophistication of this activity makes keeping track of what has ​been going on a real challenge. Despite this, numerous cases of alleged unauthorised use have cropped up ​and have led to some high profile claims. For example, Getty Images has initiated proceedings against ​Stability AI and textbook publisher Pearson has also sent a cease and desist letter to an unnamed AI ​company for using its content to train language models. In addition to the Getty claim, Stability along with ​Midjourney and DeviantArt is on the receiving end of another class action suit and Microsoft, Github and ​OpenAI are being sued for breaching open source licence terms. More recently, comedian Sara Silverman ​and bestselling authors Christopher Golden and Richard Kadrey launched a class action lawsuit against ​OpenAI and Meta for unauthorised use of their books. It is too early to tell how these cases are going to pan ​out but judgments going the way of the claimants will spell trouble for companies operating in this field.

The way forward may be a Spotify-style licensing model. In this regard OpenAI, and The Associated Press ​(“AP”) have made a deal for the former to license AP’s archive of news stories. Accordingly, blanket licensing ​arrangements may help to resolve this issue.

Further down the supply chain, Shutterstock is offering customers an indemnity for liability arising from use ​made of AI images in its library. Having said this, the indemnity being offered is capped. As the actual liability ​arising from a copyright infringement claim could be higher than this cap, buyers should still beware of the ​risks associated with the use of these generative models.

Regulators have also chimed in. The draft EU AI Act anticipates all use of third party material by generative AI ​suppliers having to be disclosed, including sources. This could pose a significant burden to AI users in the ​EU, similar to the issues posed by GDPR. Chinese interim measures seem to be adopting a similar approach, ​requiring reliance on legitimate sources, respect for third party intellectual property rights and the legal ​processing of personal information with the requisite consent or legal basis under Chinese law. On the other ​hand, Japanese counterparts are leaning towards a softer approach, treating such use of third party material ​as inspiration rather than infringement and expressing concern that an overly strict approach will just stifle ​innovation.

Meanwhile, the University of Chicago has offered a practical solution – GLAZE, a protective cloak which ​prevents AI programs from scooping up proprietary images without permission.

As was touched upon above, a related concern is privacy – personal data can be trawled and used for training purposes without the owner’s permission just as easily as copyright material. Millions of people have probably had their photographs used in this way to train AI image generation programs without their knowledge or permission and the dangers of deepfake pictures and film footage are already well known

Copyright ownership

Most jurisdictions, albeit for slightly different reasons, seem to be taking the stance that human creative input is required for copyright or patentability to arise, meaning AI-generated works will be unprotected. In the US this has resulted in partial copyright registrations, with human-generated elements to a creative work being granted protection and AI-generated ones being denied it.

Inevitably, this position has also resulted in a flurry of accusations to the effect that human authors are taking credit for AI-generated works so they can take advantage of copyright protection. As time goes on it may become more difficult to detect which works have an AI author and which a human one.

Conclusion

As with the challenges arising from illegal downloads, gene patenting and data privacy, the generative AI controversy is another case of law and regulation rapidly being outstripped by technological progress. A way forward in the form of regulation and licensing will clearly emerge when the dust has settled but it is unlikely to be perfect and it is already clear it will vary from region to region. As is often the case, the EU regulatory framework looks like it may be the most stringent. The ideal scenario of one set of rules applying worldwide seems sadly out of reach.

2: MedTech

2.1

Publication trend

Figure 2.1.2 indicates that year-on-year growth in the number of AI MedTech publications surpassed that for AI overall each year from 2014 to 2021.

Figure 2.1.2 - Growth rate per year for MedTech AI publications compared to overall AI publications at the EPO

2.2

Sector breakdown

As illustrated in Figure 2.2.1, the largest subcategories according to our classification system are medical imaging and diagnostics. There is a chance that many of the publications classified as medical imaging are also classified under diagnostics, given the potential for AI-based medical imaging to support clinicians in making diagnoses. Accordingly, there may be some overlap between these two subcategories (and other subcategories).

Interestingly, the number of filings in the field of AI-based drug discovery is much lower, with just 12 applications published in 2022 (which is too few to be visible in Figure 2.2.1). This may be particularly unexpected given the media attention that this area has generated in recent years. There could be a number of reasons for this. Perhaps this technology has simply yet to reach maturity, and it is only a matter of time before the number of applications begins to climb. Alternatively, it could be that inventors are choosing not to file patent applications based on concerns as to the patentability of such inventions in Europe (e.g. with regards to sufficiency, obviousness, or demonstrating a “technical effect”).There is some evidence in our data that such concerns would be unjustified, as the allowance rate of applications in this field appears to be fairly strong (see figure 2.3.1). Maybe the true explanation relates to long-term strategy, with inventors instead choosing to use patents to protect the fruits of their drug discovery platforms, whilst keeping the algorithms confidential.

We can also analyse the list of MedTech applications according to the type of technology that underpins them. As shown above, and as we saw for AI more generally, computer vision is again the largest contributor. For MedTech, this is followed by predictive analytics, which is significantly more prevalent in MedTech filings than for AI overall. In combination with the relatively low proportion of Drug Discovery AI applications, perhaps this shows that MedTech AI innovation is currently focussed less on treatment, and more on prediction, prevention, and early diagnosis and intervention.

When we look at the trends in publications across our subcategories of MedTech, we see that, despite the overall slow-down in growth seen in 2022, the number of filings still continued to increase across the board.

The one exception is the field of drug discovery, where the number of published applications actually fell by almost 30%, reverting to the level previously seen in 2018.

At the other end of the spectrum we find bioinformatics, which is the only subcategory where growth continued to accelerate last year (publications increased by 54% in 2022, up from 41% in 2021). It is possible that this is simply an artefact of the WIPO classification system catching up to this relatively new and exciting area. On the other hand, this might reflect the increasing utility of bioinformatics across all of the other technical areas. Indeed, while many applications relate to more than one subcategory, Bioinformatics applications are associated with other subcategories more often than any other MedTech subcategory. The same analysis also revealed that applications relating to medical imaging are, on average, associated with the least number of other subcategories, which makes the high number of filings relating to medical imaging all the more striking.

Figure 2.2.5 shows changes in the types of MedTech applications filed and published by applicants from different countries between 2015 and 2022.

The relative proportions of different types of MedTech AI applications filed by US and EP applicants, particularly from 2020 onwards, are very much alike. Notably, the proportion of applications filed by US applicants in the field of Bioinformatics is consistently the highest of all countries included in this analysis, with exception to those applications published with CN applicants between 2016 and 2017. For US and KR applicants, the proportion of MedTech AI applications directed towards diagnostics has seen a steady decline, as other areas of innovation (such as medical robotics, medical prosthetics, and medical implants) have grown.

Medical Imaging appears to be a key area for AI MedTech development in Asia and the Far East. Between the years 2018 and 2022, medical imaging applications filed by CN applicants have seen a large proportional increase to make up just over 59% of all MedTech cases emanating from this country. Likewise, applications filed by JP applicants also exhibit a higher than average proportion of medical imaging cases for this period.

9. The selected MedTech subcategories were “Bioinformatics”, “Diagnostics”, “Medical Imaging”, “Medical Robotics”, “Medical Prosthetics”, “Patient Monitoring”, “Drug Discovery”, “Consumer Healthcare”, “Medical Implants”, and “Medical Software”. Each record was classified into one or more of these MedTech subcategories according to their associated IPC codes. A case was classified as MedTech when classified into at least one MedTech subcategory.

Allowance rates

2.3

As shown below, the allowance rate for MedTech applications has remained relatively flat since 2019, with ​the trend lines for most subcategories gathered around the average allowance rate for AI applications shown ​in Figure 1.6.1. While the absolute number of applications is relatively small, the allowance rate for Drug ​Discovery applications shows a clear upward trend, with an allowance rate of 83% in 2022. Healthcare has a ​similarly low number of absolute applications but shows a significantly lower than average allowance rate.

Opposition

2.4

Opposition is a post-grant procedure that allows a third party to challenge the validity of a European patent within the first 9 months after grant.

Just under 2% of AI-related patents granted since the year 2000 have been opposed, and the figure for MedTech is almost identical.Interestingly, AI patents relating to medical imaging are less than half as likely to be opposed, with just 10 oppositions lodged against almost 1,000 applications granted. By comparison, two oppositions have been filed against just twenty-four drug discovery patents granted over the same period. One explanation for this could be that the presence of a “technical effect” is often more clear-cut in the case of medical imaging, making this line of attack more difficult for would-be opponents.In fact, the EPO Guidelines for Examination specifically refer to the classification of digital images based on low-level features as an example of a “technical” application of AI. Another explanation may simply be that drug discovery patents are more valuable, and therefore may be more at risk to opposition than patents in other fields.

2.5

Embedding analysis

For this year's report, we have used a trained machine learning model (PatentSBERTa) to generate patent claim embeddings for claim 1 of each record in our dataset. In industry, such techniques are used to analyse the semantic similarity of text – in other words, how similar is one piece of text with respect to another piece of text.

As a high level overview, PatentSBERTa is a modified version of sBERT fine-tuned on patent data for various tasks including predicting IPC and CPC codes in addition to predicting the next word in sequences of claim text. The resulting model should, in theory, be able to better represent patent claims in high-dimensional embedding space.

We clustered the claims according to the IPC based MedTech classification system described in this report. The position of each data point in Figure 2.5.1 represents the semantic content of a single patent. Data points that are close together indicate patent claims that are semantically similar. The colour of each data point represents the MedTech subcategory. By examining the resulting clusters, we were able to investigate whether PatentSBERTa accurately represents claims from different MedTech subcategories.

In general, PatentSBERTa appears to embed the semantic meaning of patent claims relatively well. We see distinct clusters for Diagnostics and Medical Imaging. Bioinformatics also shows a primary cluster but a significant number of distributed data points. As we noted above, Bioinformatics is the subcategory most often combined with other subcategories and this appears to be reflected in our analysis. Overall, this portion of the report shows that it is possible to leverage machine learning models for the comparative analysis of patent claims.

3: Looking ahead

Whilst Europe and the US continue to dominate the AI patent filings at the EPO, we have seen increasing numbers of filings coming out of the Asia and Far East, and in particular China. Given that AI is a key focus of China’s current five-year plan, we expect this trend to continue into the future.

Furthermore, it will be interesting to see what effect AI regulation has on patent filing trends. The currently proposed EU AI Act could pose a significant burden to AI development in the EU, similar to the issues posed by GDPR. This could lead to a slowdown of EU originating AI patents relative to other countries that are applying a more soft-touch approach to regulation.

Given the 18 month delay between patent filing and publication, the recent surge in generative AI that has occurred over the last year is too fresh to show up in the current patent data. Nevertheless, it will be interesting to see whether this has an effect in the years to come.

Whilst LLMs have been a huge success in recent years, this is unlikely to lead to a surge of granted patents at the EPO given the EPO’s current negative outlook on NLP inventions. Having said this, the EPO’s view that NLP inventions are non-technical could be open to appeal. If successfully challenged, this could open up a much wider scope for protecting innovations resulting from LLMs and other NLP methods.

Mike Williams

Patent Attorney

Mike is an expert in patent matters relating to digital technologies, and in particular artificial intelligence (AI). Mike is the lead partner on our AI reports, in which we studied the impact of the spectacular growth in AI on patent filings at the European Patent Office (EPO) over the last two decades.

Mike’s background in Computer Science, in which he holds a master’s degree from University of Manchester, gives him the ideal technical foundation to understand his client’s inventions in all matters of digital technologies. In addition to artificial intelligence, Mike has extensive experience of patent matters relating to all aspects of computer science, including signal processing, image analysis, communications protocols, computer graphics. Mike also is also an expert in the area of lithographic systems, analogue and digital electronics.

Mike's uses his technical expertise and commercial focus to advise a broad range of clients from large multinational companies to universities and SMEs. Mike’s international practice means that he is fluent in drafting and prosecuting patent applications throughout the world, and advising clients to navigate contentious matters.

Matthew Jefferies

Patent Attorney

Matthew advises a wide range of clients in high tech fields, with particular expertise in computing and medical devices. Since joining Marks & Clerk in 2011, he has been active in protecting innovations within the fields of telecommunications, machine learning, medical devices and non-volatile semiconductor memory.

Matthew has experience in neural networking across a variety of programming languages and represents a number of clients in the AI and machine learning space.

Matthew has drafted and prosecuted a wide variety of UK, European and International patents across a diverse range of technologies. He has also assisted clients with appeals and oppositions before the European Patent Office. His telecommunications experience, particularly in wireless networks and video codecs, has led to him advising on the applicability of technical standards and the validity of patents within litigation proceedings and European oppositions.

Lara Sibley

Patent Attorney

A large part of Lara’s practice relates to protecting inventions in the Artificial Intelligence field, and she has particular experience with applications in areas such as natural language processing, medical imaging and home automation.

She advises clients across the world on patenting AI innovations and has given numerous presentations in Japan on the EPO approach to AI. Alongside her work in the AI space, Lara also works across a range of technologies in the software and electronics sectors, including in particular cryptography, Agritech, video coding (including standards essentiality evaluations), medical devices, wireless networking, optical devices and semiconductor devices.

Lara has successfully represented clients at oral proceedings at the EPO, and has experience of EPO opposition and appeal proceedings. Lara acts on behalf of clients of all sizes, from individual inventors, through to SMEs and large multinationals. She has experience prosecuting applications in multiple jurisdictions including Europe, the US, China and Japan.

Simon Portman

Of Counsel

Simon specialises as a commercial contract lawyer for technology companies. He works for clients in the electronics, bioscience, defence, software, nanotech and creative industries, advising companies ranging from small start-ups to big multinationals as well as individuals, public bodies and charities. He heads the firm’s Extended Reality team and in that capacity supplies specialist advice and training to members of XR organisations such as Immerse UK and XR Nation.

He advises on a wide range of contracts, including licenses, R&D collaborations, manufacturing agreements and procurement documentation. On the regulatory front, he has advised on compliance with clinical trials legislation and novel food applications, as well as freedom of information and data protection issues.

Simon has co-written two business textbooks, Commercial Issues for Life Science Companies and Intellectual Property: the Lifeblood of your Company and has contributed articles to numerous publications, including the Financial Times, Patent World and Managing Intellectual Property. He regularly gives seminars and workshops at pharmaceutical conferences in the UK and internationally. He was also part of a UK IPO steering group putting together policies and contracts suitable for R&D collaborations in China.

Samuel Lowden

Patent Attorney

Samuel graduated from Durham University in 2016 with a BSc Natural Sciences in Biology and Physics. He then studied Intellectual Property Law and Intellectual Property Management at Queen Mary, University of London.

Shortly after joining Marks & Clerk in 2019, Samuel was also awarded a Research Master’s in Molecular Biophysics at King’s College London, using Macromolecular X-Ray Crystallography and Single-Particle Cryo-Electron Microscopy.

Jeremy Russ

Trainee Patent Attorney

Jeremy joined Marks & Clerk in October 2021 as a Trainee Patent Attorney in the Manchester office. He graduated from the University of Liverpool in 2020 with a Bachelor’s degree in Computer Science.

Before joining the firm, Jeremy worked on many applied AI projects involving, for instance, sentiment analysis, signal classification, image generation, speech synthesis, semantic similarity analysis, algorithmic trading with reinforcement learning, and patent analysis.

A selection of our AI experts

Tim Hargreaves

Partner / Chartered (UK) and

European Patent Attorney

Partner / Chartered (UK), European and Registered (SG) Patent Attorney

Rhian Granleese

Partner / Chartered (UK) and

European Patent Attorney

Partner / Chartered (UK) and

European Patent Attorney

Senior Associate / Chartered

(UK) and European Patent Attorney