This article marks the last in HLK’s series of articles on patenting AI. Here are our top tips to ensure that you get the most out of your AI patent application portfolio in Europe.
Is it worth filing a patent application at all?
If an AI invention is “hidden” and not easy to detect, then it is tempting to think that it isn’t commercially useful to file patents for these applications at all. However, as we’ve written before, upcoming transparency regulation of AI and standardisation efforts mean that in many industries hidden AI will be pushed into the open. In fact, due to standardisation efforts in AI, some of these could end up being amongst the most valuable patents in a portfolio.
For example, it is likely to be commercially sensible to file applications to inventions that will be used in the public sector in the UK, used in safety critical systems, or in the healthcare domain, amongst others. These sorts of fields provide the greatest opportunity for standard essential patents to be obtained. Furthermore, even if your competitors aren’t forced to be transparent by law, in these fields, they may also be effectively forced to disclose their methods for commercial reasons which will also expose the technology. Afterall, who would want an inscrutable “black-box” as a doctor, or a pilot?
Work out what type of AI invention you have and whether it is excluded from subject matter eligibility
If you decide that you do want to pursue patent protection, then work out how the invention is likely to be viewed by the EPO and whether it falls into one of the excluded subject matter categories. As we’ve written before, the EPO generally considers Applied-AI inventions, where AI is applied to a technical problem (such as image classification, telecoms, etc) more favourably than Core-AI inventions where an inventor has made an improvement to the field of AI itself (e.g. new types of models, methods of training or pre-processing data).
In the event that you conclude that your invention is a Core-AI invention, don’t despair! But do put extra time into considering how your organisation might want to monetise the invention. If there are particular licensing opportunities, or products that you have in mind that will incorporate the invention, then use cases should be added to the patent application describing how the invention can be applied to these particular technology areas. The use cases can be used as basis to narrow the claims in prosecution, in a manner that is still results in a commercially useful scope.
Put the right level of detail into the application
The EPO will only grant a patent application if the application is “sufficient”. In Europe, sufficiency is a legal requirement that a patent application must contain enough detail for a person skilled in the art to be able to re-create the invention. Thus, if an applicant wants to ensure that their portfolios don’t fall foul of this requirement, care needs to be taken to ensure applications contain the right level of detail.
To determine the level of detail required in patent applications for AI inventions, the EPO has given two main sources of guidance:
Case law in this area, such as T0161/18, suggests that the EPO considers a high level of detail to be required for AI applications. T0161/18 suggests that detailed information on training data, such as an example training data set, or detailed information on how to put together such a training data set is needed in order for an application to be sufficient. This level of detail places a high burden on applicants.
However it is worth noting that sufficiency of disclosure is assessed at the filing date of the application. The application in T0161/18 has a priority date from 2005, when commercial use of AI was arguably in its infancy, so perhaps it is arguable that in 2005, it would have been an undue burden on a skilled person to have to put together a training dataset based on limited information.
It would seem to me to be much harder to argue that the same case filed today is insufficient, however. Given the ubiquity of AI and the easy access of opensource development tools, a developer working in the field might easily be able to put together a training dataset, and train various test models, almost in real time. Thus, we might expect the standard of sufficiency to change over time.
As another source of guidance, the European patent office has previously indicated that applications describing machine learning models may be given the benefit of the doubt with respect to sufficiency if it is readily apparent that the inputs and outputs of the model are causally connected. The example that was given was that less information might be needed in an application describing a neural network trained to detect faces in images obtained using an infrared camera, compared to an application describing predicting IQ from fingerprints.
This makes intuitive sense, because if a causal connection isn’t known, or at least plausible, between two types of data, then trying to patent a ML model that can convert one type of data into the other is not so dissimilar to trying to patent the proverbial perpetual motion machine – it isn’t clear how that would work either.
As a general guideline therefore, if there is any doubt as to how the inputs and outputs of a model are related, then more information needs to be added into the application in order to provide enough detail for someone else to recreate the claimed model.
We would go so far as to suggest that this applies more widely, for example, for Core-AI inventions where an inventor has made an improvement to a type of model (e.g. they have invented a new type of neural network, or a new method of training) then it very often isn’t clear (on an intuitive or layman level at least) why the particular change that has been made works in the manner that it does. New model architectures, for example, can be opaque as to how the particular configurations achieve the effects that they do.
In such situations, if a patent application is to be granted, then it is a good idea to provide more information in an application than you might otherwise consider to be necessary.
What information might be provided to improve the chances of grant?
A detailed example should be provided, including, for example, a fully described architecture of a particular model type, set up parameters, and the training data that might be used. You could point to an open-source dataset or reference a suitable dataset known in academia.
If a causal connection isn’t readily apparent, then any information that supports the presence of one might be added. Academic papers could be a source of reference here too.
Experimental data might also be added demonstrating that the claimed effect can actually be achieved. In the case of an improved model, comparative data might be added to an application, showing, for example, a percentage improvement compared to a prior art example trained in the same way on the same data.
In summary, it is possible to obtain commercially useful patents for AI applications in Europe. It can be a balancing act between choosing inventions that are detectable, and also patentable before the EPO. When you find an application that strikes that balance, be sure to put in enough information to give your application the best chance of a good outcome.