Head of Data Science bei European Patent Office, Vectorizer of the grand Prior Art Corpus
I moderated a thought-provoking panel on Generative AI's impact on patents at the PatentSemTech workshop, SIGIRConf 2025 in Padua. Panelists included Eleni Kamateri, Allan Hanbury, Linda Andersson, Sebastian Björkqvist, and Gabriella Pasi. Our discussion centered on "Human in the Loop" as crucial. Generative AI, particularly Large Language Models (LLMs), offers powerful capabilities for tasks like prior art search and drafting. Our panelists emphasized these tools are not infallible, likening LLMs to "teenagers" – brilliant but unpredictable. Human oversight is vital for accuracy in this precision-demanding field. We explored the "paradigm change" in patent search, from Boolean to relevance ranking, and now to vector search. Notably, while the shift from Boolean to relevance took a considerable time to be widely accepted in patent search, the current rapid adoption of vector search and Generative AI-driven search hints at a completely different pace. Yet, integrating LLMs presents hurdles: inherent biases, high costs, and limited AI democratization due to resource concentration in large companies. Beyond search, AI also holds promise to smoothen the overall patent process, and crucially, help scientists better understand the complex language of lawyers and patent examiners, bridging vital communication gaps. Panelists agreed "finetuning is not the solution" for patent accuracy, as it's often too slow and expensive to maintain. Instead, the future lies in "small models with domain knowledge." These models are less resource-intensive, making them more accessible and controllable for researchers. We also debated "Why do we need long patent text?" suggesting better, more symbolic information representations. The patent system, though complex, is here to stay. The challenge lies in enhancing prior art search quality amidst industry obfuscation and AI-generated "prior art." Ultimately, AI literacy is crucial. Professionals must understand these tools and combine them strategically with their indispensable domain knowledge. The path forward requires cautious, informed adoption, integrating diverse tools, and maintaining expert human judgment at the core of the patent process. It was an enlightening discussion on AI's exciting, complex journey in intellectual property. #SIGIRConf #PatentSemTech #AI #Patents
In my view, using AI for searching information is completely different from using AI for drafting patents, and these two moments should not be confused but kept separate. Further, filing patents just for filling gaps provides value only to very few people: the inventor has to invent and solve real problems, if we want the patent system to thrive. Otherwise it will become only an academic exercise, and this is not the aim of the patent system. We have to maintain the patent system relevant for society: a patent is a "social" right disguised as capitalism.
Thanks Alex. Your insights are always helpful and I support your conclusion that: “AI literacy is crucial. Professionals must understand these tools and combine them strategically with their indispensable domain knowledge. The path forward requires cautious, informed adoption, integrating diverse tools, and maintaining expert human judgment at the core of the patent process.”
Thanks Alexander Klenner-Bajaja for this report, I hope that EPO will work in applying my personal two take-home messages ("small models with domain knowledge" and "AI literacy for patent professionals") by collaborating with patent professionals and hearing their views (not only by exposing EPO views and programs at virtual seminars)
Head of Data Science bei European Patent Office, Vectorizer of the grand Prior Art Corpus
3moHidir Aras & Ralf Krestel thank you for organising and hosting the workshop 👍🏽 And a thank you to the engaged audience - we concluded that almost everyone in the audience participated in the discussion!