Patents, AI implemented inventions and the effect of the Emotional Perception appeal judgment
Having had a bit of time to take in the Emotional Perception AI Court of Appeal decision, it seems it will now be harder to potentially gain patent protection for AI implemented inventions.
After the High Court decision statutory guidance said patent examiners shouldn’t object to inventions involving artificial neural networks (ANNs) under s1(2) of the Patents Act 1977. That states that inventions consisting of (amongst other things) a program for a computer or mathematical method are excluded from patent protection where the invention relates to that thing “as such”. This broadly means look at the actual/alleged contribution of the patent/patent application and examine if it falls solely within the excluded matter. If it doesn't and it makes a technical contribution, it’s potentially patentable.
It's no surprise that the High Court decision was appealed. The prospect of easier patentability for AI implemented inventions was always going to be examined by a court at the highest level. Indeed, a further appeal to the Supreme Court is possible.
The appeal focussed on these questions:
- What is a computer program?
- Is there a computer program in an ANN?
- Does the ANN make a technical contribution?
- Does the mathematical method exclusion apply to the ANN?
Computer programs and ANNs
There’s no legal definition of a computer program, so Lord Justice Birss’ comments on this are helpful when trying to work out the remit of s1(2). He held that a computer is a machine which processes information. A computer program is a set of instructions for a computer to do something. These definitions work together meaning that a computer is a machine which does something, and that thing is to process the information in a particular way. The program is the set of instructions which cause the machine to process the information in that way and not another.
In the High Court Sir Anthony Mann analysed the hardware ANN and software emulated versions separately. He thought the emulated ANN existed at a different level to the underlying software. In his view, there was no program at that point because the ANN was not operating a set of instructions but applying its own weights and biases. The Court of Appeal disagreed. It didn’t matter how the ANN was implemented, the weights of an ANN are a computer program as they are a set of instructions for a computer to do something, so the s1(2) exclusion is engaged.
This part of the appeal judgement is surprisingly brief compared to the High Court decision. Its brevity and the lack of reply to Mann’s detailed analysis makes it feel like a policy decision. Seemingly the Court of Appeal decided the computer program exclusion should apply to ANN implemented inventions, so it went back to basics and looked at the functionality of the AI. Although the AI system could ultimately train itself, this was part of the weights and biases applied by the ANN, which was part of the creation of the computer program, so it fell squarely within the s1(2) exclusion. This will be hard to get over for any applicants wishing to gain patent protection for AI implemented inventions in the future.
Technical contribution and mathematical methods
Even though the program for a computer exclusion is engaged by an ANN implemented invention, it doesn’t mean the invention is unpatentable. It simply means ANN implemented inventions are in no better and no worse position than other computer implemented inventions. The only way to save such inventions is to apply the grounds from Aerotel and see if they make a technical contribution.
The undisputed contribution of Emotional Perception's ANN was a system for providing improved file recommendations, commonly in the form of music tracks. Subject only to the step of sending the recommended file and message to the user, the whole of the remainder of the contribution was said to consist of a program for a computer.
The killer blow for Emotional Perception was that the file represented a better recommendation in the form of a track which the user was likely to enjoy. The ANN makes this recommendation by assimilating the semantic (linguistic) characterisation of a track (rock, folk, happy etc) and allocating coordinates to it. A different ANN characterises tracks by physical properties (tone, speed etc) and produces separate coordinates. The ANN is then trained until it can take any music track, calculate its property space coordinates and find semantically similar music. What makes the file worth recommending is its semantic qualities, which are subjective and cognitive in nature, not technical.
The judge also commented that the mathematical method exclusion in s1(2) might have applied to the ANN if the Court of Appeal had said the ANN wasn’t a computer program. Lord Justice Birss thought it was hard to see why ANNs aren’t a mathematical method, so the same exclusion and the same result would apply. This is likely to lead to further objections to the patentability of ANNs.
Implications
Ultimately it seems this judgment will make getting a patent for an AI implemented invention harder. In reviewing a patent application, the IPO and/or court will apply the Court of Appeal’s reasoning and are likely to say whatever the AI does, it does it because of how it’s trained. This is part of the creation of the computer program, so must be disregarded when considering patentability. If the training of the AI has to be ignored, how can an applicant prove the AI system makes a truly technical contribution that isn’t simply part of the AI applying its weights and biases as a computer program?
Although each patent application will turn on the wording of the particular claims and how the technical contribution is set out, this will be the challenge when applying for patents covering AI implemented inventions following the Emotional Perception AI appeal judgment.