A Game-Changer for Machine Learning Patents in the UK?

In a remarkable decision, published on 21 November 2023, the High Court has overturned a decision by the UKIPO relating to machine learning. 

This decision appears to significantly broaden the type of inventions that might be patentable in the UK.  This will be welcomed by companies and inventors working on inventions in the field of machine learning.

This was an appeal from a decision by the UKIPO to refuse a patent application.  The underlying patent application involved the use of an Artificial Neural Network (ANN).  The UKIPO took the view that the invention described in the application was a computer program as such and did not involve any technical contribution.  On appeal, the judge, Sir Anthony Mann, disagreed on both points.

The invention related to a technique for providing media file recommendations to a user.  The claimed method would provide recommendations for media files that were similar to an input file, based on apparently similar tastes of other humans.  This was achieved by training an ANN using a training set that included media files that could be described objectively, based on their physical characteristics, and subjectively, in written form.  The method was able to identify similar media files by merging an assessment of these two criteria together, mimicking the way in which they might be assessed by a human.  The output was a file recommendation for a user.  This file recommendation was determined to be “better” in the sense that it is more likely to meet the needs of a human user by providing an appropriate recommendation.

There are two independently intriguing aspects to the decision.  First, whether the computer program exclusion is engaged at all by an ANN.  And second, in any event, whether the invention involves a technical contribution.

Is the Exclusion Engaged by an ANN?

Once trained, an ANN applies weights and biases in its individual nodes.  In this way, an input file can be provided, and an output generated without any human involvement, beyond the training phase.  The actual weights and biases applied at each node are determined by the training process, and are not based on any computer program that has been written by a human. 

On this basis, the judge decided that there was no computer program in a trained ANN that could engage the exclusion.  Furthermore, the judge thought that it made no difference whether the ANN was implemented in hardware or as a software emulation.

Does the Invention involve a Technical Contribution?

Case law in the UK requires that a patentable invention must pass the four-stage test established in Aerotel.  On this point, the judge decided that, even if the computer program exclusion was engaged by an ANN, the invention nevertheless made a technical contribution.

The effect of the invention in this case was the output of a specific file recommendation to a user.  The specific file recommendation provided was only possible because of the model that was applied.  In the absence of this model the file recommendation would not and could not have been provided to a user.  In the judge’s opinion this involved a technical contribution because the file was transferred to a user, which is external to the computer.  In the judge’s view, it did not matter that the transferred file might have some subjective impact on a human user, by prompting them to select a media file that was “similar” to an input file, as determined according to human tastes.

Comments and Analysis

This decision appears to dramatically overturn the established practice of the UKIPO.  It also appears to diverge from established jurisprudence at the EPO.

The first consequence is that trained machine learning models are likely to have a much better chance of success in UK patent applications.  Secondly, following the logic in the decision, it seems that many more types of invention could be considered to involve a technical contribution, since a technical contribution was acknowledged in this case simply by the transfer of a file, even though the key benefit could only be determined by a subjective human subject, based on the correctness of any recommendations.

This patent application was not filed at the EPO, which means we are unfortunately denied a direct comparison in terms of examination process.  If this application had been filed at the EPO then it could have been a very different story.  The EPO would have assessed this using the standard Comvik approach (T641/00), to establish whether there was a technical solution to a technical problem.  Following this approach, it is quite likely that the EPO would argue that the improved file recommendation was “better” only when viewed through the eyes of a subjective human.  In that sense, we anticipate that the EPO would have argued that this invention addresses a non-technical subjective problem. 

We await the UKIPO’s next step with interest.  It will be interesting to see whether they decide to update their practice guidance, or whether they will appeal the decision.  An appeal seems likely, and, if this happens, this story may have more chapters to come.

Please do get in touch if you have any questions concerning machine learning and artificial intelligence and what it might mean for your innovation. Find contact details on my website profile here or contact us at gje@gje.com.