Artificial intelligence (AI) is driving innovations in numerous sectors. Whether it is advanced diagnostic systems in healthcare, autonomous vehicles or intelligent automation in industry, AI often plays a leading role in developing new solutions.
However, the integration of AI into various solutions also raises questions of patentability. The main question here: when is an AI invention sufficiently technical in nature to be eligible for patent protection?
Solution to technical problem
According to the European Patent Office (EPO), an AI invention is patentable only if it demonstrably solves a technical problem. While AI models, such as neural networks, are essentially mathematical and abstract, they are no longer so once they are applied for concrete technical purposes. For example, when AI is used in medical devices to detect abnormalities such as cardiac arrhythmias, or when it contributes to image and audio classification via signal processing.
Technical effect
However, not all AI innovations are eligible for patent protection. Innovations that are purely abstract, cognitive or economic in nature are out of the question. An algorithm that classifies documents by content, without a technical application, is considered intellectual activity. Nor are purely abstract improvements to an AI model, such as higher accuracy in predicting preferences or economic trends. This distinction is enshrined in the COMVIK approach, by which the EPO assesses whether a technical effect and an inventive step have actually been achieved. According to this approach, only the technical characteristics of an AI innovation should be taken into account when assessing inventive step.
Specific use
Recently, the EPO raised the bar. It is no longer sufficient to simply mention an AI technique; patent applications must demonstrate how the model is technically integrated and functionally bounded to a specific use. Specifically, patent applications must clearly describe how and in what context the AI model is used. For example, how exactly it contributes to better signal processing or more efficient use of hardware. Lacking these details, an application risks being rejected due to insufficient technical contribution.
Reproducible
In addition, there is a requirement of reproducibility of the technical effect. This means that patent applications must include sufficient details about how models are trained, including specific properties of the training data. In practice, this leads companies and researchers to document and explain their AI methods more precisely.
Europe vs. United States
Although the EPO and the USPTO (United States Patent and Trademark Office) largely agree on the requirement that AI inventions have a technical application, there are differences in emphasis. These make the U.S. system more friendly to AI-related applications. For example, the USPTO looks more at practical applicability and the tangible improvement AI offers in the real world, while the EPO looks specifically at the technical nature and implementation details of the AI application.
Thus, an AI invention is patentable if it explicitly realizes a technical use or improvement. Successful patents in the AI sector are thus no longer just a matter of clever algorithms, but also of demonstrable technical improvements in the “real” world.
Read our dossier Patenting AI inventions
