AI Patent Prior Art Search: A Feasibility Study

Can artificial intelligence genuinely support patent examiners in searching prior art? This is the question the UK Intellectual Property Office (IPO) set out to answer when it commissioned Cardiff University to run a dedicated feasibility study. The goal was to assess how realistic, technically demanding, and effective AI-based tools could be if applied to prior art searching, with a specific focus on developing a proof-of-concept for an AI-powered search or due-diligence check that could eventually sit within the IPO's online filing system and the examiner's own search workflow.

Prior art searching is far from a simple lookup task. It's an interactive, multi-step process that typically involves running several searches, applying different search strategies, and actively managing how those searches evolve. For any AI system attempting to support this work, the biggest linguistic and semantic hurdles come from the nature of patent language itself: formal legal phrasing, unusually long sentences, dense acronyms, and the technical specificity of patent claims.

The study's objectives were threefold: assess how viable various AI technologies actually are for this kind of search, compare different methods to work out which algorithms perform best, and thoroughly evaluate whichever approach came out on top. To do this, the researchers drew on a wide spectrum of modern supervised and unsupervised machine learning methods, applying them to tasks like pulling out relevant features from patent text, expanding search queries, classifying documents, clustering them, and modelling underlying topics. The toolkit included natural language processing methods such as text segmentation, normalisation, lemmatisation, stemming, identifying word co-occurrences, and detecting multi-word terms; supervised learning models including support vector machines, naive Bayesian classifiers, decision trees, and random forests; unsupervised approaches like word embeddings, distributional semantics, and neural networks including deep learning; and semantic technologies built around lexico-semantic knowledge and latent Dirichlet allocation.

One of the study's clearest conclusions is that, with the AI tools available today, a fully automated end-to-end solution for the patent filing process simply isn't achievable yet. That said, the research did uncover a genuinely promising result: while examiners currently sort patents into technology categories by hand, the automated classification approach tested in this study achieved remarkably high accuracy. That opens the door to embedding this kind of classification directly into the online pre-filing process, giving prospective applicants an easier way to run due-diligence checks before they even file.

Looking across the different AI technologies tested, the researchers found consistent evidence that no single algorithm, on its own, can handle every stage of the prior art search process, whether that's classification, building a search query, retrieving documents, ranking them, spotting similarities, or visualising topics. Importantly, the project wasn't aiming to build a complete, production-ready information retrieval system. Instead, the goal was a proof-of-concept flexible enough to let the team experimentally compare different approaches side by side. What emerged is a picture where different state-of-the-art algorithms each handle a different piece of the puzzle: retrieving the most relevant documents, ranking them, suggesting useful synonyms, recommending classifications, and clustering or visualising the resulting concepts.

The concept model itself was built around a user-centred design philosophy, shaped throughout by the actual needs, preferences, and constraints of patent examiners. This produced a human-in-the-loop system, one where AI and human judgment work together rather than AI replacing the examiner's role. The underlying algorithms are there to help examiners move through large volumes of patent data more efficiently, by surfacing plausible search terms and organising patents into topic categories that are easy to interpret. Throughout, the examiner stays firmly in the driver's seat as the actual decision-maker, with AI functioning purely as an intelligent support layer.

To put all of this into practice, the team built a working system in Python based on their concept model. Testing and validation were carried out across three technology domains, civil engineering, computer technology, and transport, chosen specifically because they represent the three highest-volume filing areas at the IPO over the past decade. The underlying proof-of-concept was trained on English-language patent data pulled from several sources: the PATSTAT worldwide bibliographic database, along with full-text patents from the UK, European, and US patent offices. From there, qualitative testing was carried out with real patent examiners, who worked through a set of representative "query" patents across each of the three domains.

The results from these examiner-led experiments were encouraging. They pointed strongly toward AI's ability to cut down both the time and cost involved in prior art searching, particularly the labour-intensive task of sifting through large volumes of retrieved patents. Precision in the experiments ranged between 30% and 50%, meaning that out of the first ten results returned, somewhere between three and five were genuinely relevant. Examiners who took part in the testing confirmed that this hit rate outperformed what they typically achieve with their existing search tools, a strong signal that AI-assisted search has real potential to support examiners going forward.

That said, the study is clear about where the limits lie. AI-assisted search still depends entirely on the examiner manually crafting the initial search statement. No AI algorithm currently exists that can reliably read a patent application and generate that statement automatically, and this remains one of the most demanding, expertise-driven parts of the entire search process. Formulating a strong search statement requires a genuine grasp of the application's core subject matter and its potential novelty, judgment calls that depend on the kind of specialised experience only a human examiner brings. The researchers are explicit that this should stay a human responsibility, both to properly scope the AI's search and because it isn't a task suited to automation.

Examiners who tested the proof-of-concept also offered a range of ideas for improving the system further. Suggestions included more flexible search strategies, such as drawing on different sections of the patent text at different points in the process, zeroing in on the paragraphs most central to the invention to sharpen retrieval, or adjusting how search parameters are weighted. Other ideas involved hybrid strategies that combine text-based and image-based search, and knowledge-based strategies that bring in structured knowledge types like method or process categories, along with domain-specific ontologies.

The experiments also revealed just how differently individual examiners approach their searches, underlining a real need for more adaptable tools going forward, ones that can flex to match varied search strategies. The researchers see room to build on the current process with new tools for tasks like searching image-based patents, gathering due-diligence evidence, catching ambiguity in applications, identifying contradictions, and visualising how documents relate to one another.

Taken together, the study offers a thorough assessment of how viable different AI technologies, spanning both supervised and unsupervised machine learning, actually are for prior art searching. The clearest finding is that no single available algorithm can independently cover the full breadth of the process. The proof-of-concept built for this research instead layers different state-of-the-art techniques across the different stages of the search. The resulting hit rate outperformed examiners' existing tools, reinforcing the case that AI has real potential to support patent examiners going forward. The study also points to promising future directions, combining AI with NLP and computational semantics, while stressing the continued importance of human-centred decision support. At the same time, the researchers are candid that this is just a starting point: larger-scale, more rigorous testing is still needed, involving a bigger set of patents, more examiners, a broader range of technology areas, and further research into algorithms capable of supporting more flexible, iterative search strategies.

If you're curious how AI-assisted prior art search has evolved since this research, WOIPS's own novelty search tool builds on similar AI-driven principles, retrieval, ranking, and similarity analysis, to give inventors and applicants a fast, practical read on their invention's novelty before they file.