Image recognition is incredibly difficult

• Who will invest in designing new trademark tools and input the relevant data?
• Identifying images in trademark registers, comparing them and presenting them is something that has tested researchers
• Anything that is argumentative will remain human for some time

Is there any limit to the potential of new tools in trademark work? Technologically, it would be rash to rule anything out. However, during research for this report it was indicated that there are some areas that will prove resistant for a while at least, either because the data is too messy or because there is unlikely to be enough demand to warrant investment.

One such area is image recognition. That might seem surprising given that free software in smartphones has been able to recognise people’s faces accurately for some time, but the challenges of identifying images in trademark registers, comparing them with others and presenting the results in a comprehensible way is something that has tested researchers. It was only in 2017, that Compumark launched TM go365 Image Search – a self-service image search tool. “Image recognition is very difficult,” explains one service provider. “There are lots of problems with it and it is hard to give risk-scoring percentages. Moore’s Law doesn’t exist in image technology.” This is also one reason why effective design searching is so challenging. “Designs are the next step for us,” says the head of one specialist service provider.

Another area is assessing descriptiveness. In a world where, as one executive articulates, “language is constantly changing thanks to the infinite creativity of the human mind”, AI tools have to be pretty sophisticated to judge, for example, whether a compound word mark or a slogan is descriptive for particular goods or services in a given language. Any user of an EU trademark can give plenty of examples where the nuance of meaning has been lost in a multi-lingual system.

Proving use provides another example. It seems like a straightforward binary question: has the trademark been used or not? However, as anyone who has ever been involved in such a dispute can testify, in practice that question can break down into many other questions, each of which depends on several factors: was the use in commerce? Was it in the relevant territory? Did it take place in the relevant time period? Was it for the relevant goods or services? Was it use of the mark as registered? There is no doubt that algorithms can be designed to address such questions. The big question is who will invest in designing them and input the relevant data? “It takes a lot of funding to create these new systems,” as one interviewee contends.

Opposition and cancellation actions will also likely remain the domain of humans for some time yet. Again, this will be partly due to client choice rather than technological limitations. Conducting essentially administrative actions that involve filling in standard forms will be well within the scope of many AI systems, but in cases where there are markets for actual products and services at stake rather than just marketing plans, clients are likely to prefer using real people for a while yet. One of the reasons is simply that they have someone to blame if it goes wrong (and that someone of course has professional indemnity insurance – something that service providers will have to review if they expand services in this direction).

In short, as one expert presents, anything that is “argumentative” will remain human for some time, including infringement cases and licensing negotiations. However, this does not mean technology will not play a role. “Standard arguments can be applied automatically with software,” as one provider claims. Moreover, as in other areas of the law, service providers will increasingly be called on to provide evidence in litigation. Judges can be persuaded by evidence from Google search results or spontaneous examples of confusion on social media (such as the in the Glee case in the United Kingdom). Finding such evidence can involve many hours of legal work but it would take just a few minutes for a well-designed algorithm, particularly if it is using many years’ worth of search reports and opinions.