Inventive AI: can machines innovate?

The integration of technology based on artificial intelligence into our everyday lives continues apace, in fields such as self-driving vehicles, speech recognition, and machine translation. With the recent explosion in the collection of data and increases in computer power for processing that data, the influence and effectiveness of AI-based innovations is growing rapidly.

One area of particular interest in the intellectual property profession at present is the extent to which an AI system can contribute to the creation of an invention. Put simply, can an AI system be an inventor?

This is a hot topic which has attracted the attention of various patent granting authorities, as well as the general media.

For example, the United States Patent and Trademark Office recently put out a request for comments on the involvement of AI in patenting inventions. One topic of interest here is whether laws and regulation around inventorship need to be revised to take into account the potential contribution of an entity that is not a natural person.

Meanwhile, earlier this year the European Patent Office commissioned a study from Dr Noam Shemtov on inventorship in inventions involving AI activity. This study concluded that none of the major jurisdictions for IP protection allow for AI systems to be considered as an inventor.

A common response to the question of whether AI is capable of inventing is: “No, it’s just a tool for human use”, or “No, current AI systems are not capable of imagination, and hence cannot produce anything that is beyond what they are trained to do”.

Such views are based on a common view that AI is typically used to perform certain well-defined tasks. In this type of scenario, an AI system would typically be provided with a training set of labelled data from which patterns can be derived, for example to set the weights of interconnected nodes in an artificial neural network.

An advantage of AI systems is their ability to recognise and utilise patterns in multi-dimensional data, where corresponding human analysis would be practically impossible. Once trained, such systems can be used to recognise corresponding patterns in unlabelled data, which may find use in diagnostics or other decision-making processes.

From the perspective of innovation, an AI system used in this way would not normally generate an invention, since the task it is trained to perform is known in advance. This is an example of an AI system being used as a tool, for which the rules relating to the inventorship are likely to be the same as any over invention where a computer program was used by a human in its conception.

In his report for the EPO, Dr Shemtov identifies an inventor as an entity responsible for “the intelligent and creative conception of the invention”. He notes that “in this context intelligence should carry its day-to-day meaning as in conscious self-aware and volitional, hence excluding AI systems at present as well as in the foreseeable future”.

Where AI is involved, Dr Shemtov’s report concludes that the inventor is the human who “geared up the AI system towards producing the inventive output, taking decisions in relation to issues such as the choice of the algorithm employed, the selection of parameters and the design and choice of input data, even if the specific output was somewhat unpredictable”.

In August 2019, several mainstream media outlets reported on a project that has a stated aim of “seeking intellectual property rights for the autonomous output of artificial intelligence”. The Artificial Inventor Project has submitted two patent applications that disclosed inventions allegedly generated by a particular type of AI system, which the project has dubbed “DABUS”.

DABUS is interesting because it attempts to present an invention that is wholly created, from original conception to realised proposal, by an AI system. In this way it challenges the conclusion in the Shemtov report, by suggesting that no human is involved in the creation of the invention.

The DABUS AI system uses a type of reinforcement learning technique, in which two AI sub-systems (known as artificial neural networks) are arranged to interact with one another.

A first AI sub-system is established with a ‘world view’ made up of data contained in one or more information sources (which may be online data sources or encyclopaedias, for example). The first AI sub-system is trained using this general knowledge to ‘learn links between different topics. It is then configured to autonomously perturb the links that were established during training.

Meanwhile, a second AI sub-system is arranged to monitor the first AI sub-system to recognise the perturbations as new ideas, and then assess them for utility or some other value against the pre-existing world view. The second AI sub-system implements a form of reinforcement learning through which new ideas that exhibit utility are fed back into the world, where they are available to be developed through further perturbations. New ideas without benefit are not fed back, so the second AI sub-system therefore acts to promote the new ideas having some perceived benefit.

In the DABUS project, an AI system operating in a similar way to that described above created a design for a beverage container and a device for attracting attention. The DABUS team claims that these ideas have come wholly from the AI system, and not as a result of a human using an AI system to solve a certain problem.

The DABUS patent application do not name any human inventors. As such, they should force the issue of AI inventorship to be considered. Robert Jehan, who is one of the patent attorneys involved in the DABUS project, discusses some of the issues that would need to be addressed in a scenario where an invention was devised by an AI system.

On the one hand, it is possible to take a view that DABUS itself is a human creation, designed for the purpose of generating new, fully formed ideas. Following the conclusion of the Shemtov report, one may then postulate that all inventions generated by DABUS are invented by DABUS’s creators, since those inventions would not exist if those humans had not created DABUS. The patent system could remain unchanged if this view were adopted.

On the other hand, one could say that the complete disconnect between any human thought process and the actual invention in the two DABUS applications means that it would actually be wrong to name a human inventor, both in principle and according to current patent law regulations, as mentioned in the Jehan article. It seems likely that this disconnect will only increase as more sophisticated ways of self-improvement for AI systems are developed.

For this reason, the time seems right to grapple with the concept of non-human inventorship, so that patent law remains fit for purpose as we continue our journey through the fourth industrial revolution.