Can you get a patent for AI? In particular can you get one for an invention concerning the use of an Artificial Neural Network (ANN)? To date, most practitioners would have had serious doubts due to patent eligibility/excluded subject matter provisions. This has prompted a renewed interest in the laws governing trade secrets and copyright as a means of protecting and monetising the significant investment in all things AI-related.
However, the UK High Court has recently issued its first judgment on this topic, Emotional Perception AI v Comptroller-General of Patents  EWHC 2948 (Ch) and has said 'yes'.
This case was an appeal from a decision of the UK Intellectual Property Office (IPO) which refused the patent application on the grounds it related to excluded subject matter for being no more than a "computer program as such".
As discussed in more detail below, the reasoning applied in the decision is a departure from the way similar inventions are assessed before the European Patent Office (EPO). If upheld subject to any further appeal, and implemented accordingly by the IPO, this would potentially pave the way for the UK to become a more patent-friendly jurisdiction in the field of AI innovations, both compared to what it was prior to this judgment and as compared to the EPO.
However, although this is a patentee-friendly decision, the courts have issued several similarly helpful decisions in the wider field of computer-implemented inventions over the years that have not always translated into smooth prosecution before the IPO. It therefore remains to be seen how the IPO will adapt its examination procedure in practice but this judgment is a positive step.
What was the invention?
The invention as claimed related to the provision of recommended media files to a user. The claims were not limited to a particular type of media, but for the purposes of illustration here will be taken to be music, which was the context in which arguments were presented by the parties and the invention described in the judgment.
During a training phase, two ANNs were trained on pairs of music files. The first ANN receives as inputs a pair of music files, each accompanied by a natural language (or "semantic") description of how the music in that file would be perceived by a human (in a simple example "happy", or "sad"). From this, the ANN produces, for each track in the pair, co-ordinates in a 'semantic space' that characterises the track. The distance between the pair of co-ordinates represents the semantic similarity of the tracks - the shorter the distance, the more semantically similar the tracks. A second ANN analyses the same pair of tracks in relation to properties of the music - such as tone, speed and loudness - and produces a pair of coordinates in 'property space'. The second ANN is then trained to make the distance between the co-ordinates in property space converge or diverge based on the distance between them in semantic space. The process is then repeated over many pairs of tracks.
Once trained, the second ANN can provide a single vector in property space for a given input track of music that can be used to identify other tracks in close proximity in property space that, by virtue of being in proximity in property space, will have a similar semantic style. In effect, the ANN can identify semantically similar tracks from the properties of the input track. A recommended track is then communicated to a remote user.
How is the patent eligibility of AI inventions assessed?
In essence, patent eligibility in the UK is assessed by identifying the 'contribution' the claimed invention makes over the known art and determining whether that contribution resides solely in excluded subject matter (e.g. a computer program or mathematical method ) or not. Those inventions that are deemed to make a 'technical contribution' fall outside the exclusions and are eligible. There are a number of signposts that can be used as guidance when making this assessment, the key one in this case was whether the claimed technical effect has a technical effect on a process carried on outside the computer.
The procedure before the EPO is different but, up until now at least, produces similar outcomes. Before the EPO, an invention will fall outside the excluded subject matter provisions if it involves any 'technical means'. Use of any generic computer is sufficient to overcome this hurdle. The tougher assessment comes under inventive step, where only those parts of the claim said to produce a technical effect serving a technical purpose are considered in the assessment. It may be understood that eventually both the IPO and the EPO are asking themselves the same question about technical contribution, it is simply that they ask it at different points in the examination of the application.
What does this decision say?
Where's the computer program?
The judge found that the claims did not invoke the statutory exclusions at all because the ANN was not a computer program, regardless of whether the ANN was implemented in hardware or emulated in software running on a computer. The basis for this finding was that the ANN does not execute a series of instructions prepared by a human; instead, it operates according to an approach it has learned itself. When emulated in software, the judge found that the ANN operates at a different level from the underlying software on the computer. Though the emulation of the architecture of the ANN (in terms of its nodes and layers) was a result of programming, the actual operation of those nodes and layers is created by the ANN itself and not provided by a human.
This finding, if upheld on any appeal, is particularly powerful due to its general applicability. It would seem to apply to any trained ANN and, by not invoking the exclusions at all, sidesteps any difficult questions that can arise on technical contribution depending on the use of the ANN.
What if there is a computer program?
Helpfully, the judge went on to consider the issue of technical contribution in the event he was wrong about the claims not being a claim to a computer program at all.
It was agreed between the parties that, since the claim included the step of providing the chosen file to an external user, there was an effect taking place which was external to the computer. What was in dispute was whether this effect was technical or not. It is important to bear in mind at this point that the file shared with the external user encompassed music files provided on the basis they were determined to be a good semantic match to a chosen input file.
Crucially, the judge found that the chosen file was identified through the application of technical criteria that the ANN has worked out for itself, and this was sufficient to conclude there was a technical effect - the fact that provision of the chosen file might have a subjective (i.e. non-technical) effect on the user is irrelevant. In other words, it did not matter that the file itself contained 'non-technical' information (music), nor that the file was chosen for the non-technical purpose of being semantically similar to an input file.
This reasoning looks to diverge from the general body of case law developed at the EPO, where the assessment of whether there is a technical effect would look at the purpose of the claim features. Though only hypothesis, the fact the ANN operated to provide to a user files semantically similar to a chosen input file would likely be deemed a non-technical purpose under the EPO's framework.
Implications and takeaways
The decision is welcome news to those looking to obtain patent protection in the UK for AI inventions involving ANNs. However, two points of caution are worth bearing in mind.
The first is that the analysis of technical contribution was undertaken in the context of whether there was a technical effect outside the computer (one of the signposts used to indicate a claim is outside the statutory exclusions). This was possible because the claim recited the step of actually providing the identified file to the external user. Had the claim stopped at merely identifying a file using the ANN, the outcome would likely be different. This reinforces the value there can be in including within claims a step of providing an output, if appropriate within the context of the invention. Similar reasoning was employed by the courts in finding claims did not relate to excluded subject matter in Protecting Kids the World Over (PKTWO) Ltd,  EWHC 2720 (Pat).
The second is that helpful decisions from the courts (from the perspective of the patentee) already exist in relation to computer-implemented inventions more generally, yet applicants could still face strong resistance from the IPO during prosecution. Practical experience has indicated that it is not unusual for the IPO to apply the criteria for patent eligibility quite strictly, and it remains to be seen how this latest decision is applied in practice.
Overall, a positive outcome for those looking to protect their investment in AI in the UK.