par Vandamme, Thomas
Président du jury Puttemans, Andrée
Promoteur Cabay, Julien
Co-Promoteur Debeir, Olivier
Publication Non publié, 2025-04-17
Thèse de doctorat
Résumé : The ever-increasing number of Trade Mark (TM) applications, combined with the complexity of the legal standard of the Likelihood of Confusion (LoC) governing TM comparisons, creates a significant administrative burden. Recent advances in Artificial Intelligence (AI) have led to the development of numerous tools aimed at alleviating this burden. Among the most promising are TM search engines, which allow users to search for potentially confusingly similar marks. Several public Intellectual Property (IP) offices, responsible for administering TM registries, have made such tools available to the public. Despite their practical relevance and the critical legal stakes involved, the capabilities of these search engines have seldom been rigorously assessed.This thesis investigates the capabilities of two such tools: the one of the Benelux Office for Intellectual Property (BOIP) and the one of the European Union Intellectual Property Office (EUIPO). Two experiments, grounded in settled case-law, are conducted to assess their effectiveness. The results are underwhelming: neither system successfully retrieves confusingly similar TMs in more than 10% of cases. Furthermore, comparative analysis reveals significant methodological flaws in the development and testing of the EUIPO’s system.To bridge the gap with academic literature, a further experiment examines a state-of-the-art algorithm for TM retrieval. This analysis exposes similar shortcomings, notably that the algorithm is overly tailored to the standard benchmark dataset (METU), which fails to accurately reflect the legal and practical realities of TM retrieval. A direct comparison of the BOIP and EUIPO systems with this algorithm confirms the latter’s sligthly inferior performance—partly attributable to the misaligned objectives embedded in the METU dataset.Overall, this thesis shows that none of the tested algorithms are sufficiently able to implement the relevant legal standard of the LoC, leading to the conclusion that algorithmic confusion remains an unsolved problem. In light of the methodological flaws uncovered in the development of these algorithms — and, more importantly, the general absence of adequate evaluations — we propose several avenues to reduce the confusion that surrounds them. Chief among these is the introduction of two purpose-built datasets, specifically curated for benchmarking trademark retrieval algorithms using opposition cases. We hope these datasets will encourage practitioners and stake-holders to engage more critically with performance standards and to foster the development of more robust, legally grounded systems.

Documents en relation