Decoding DuPont: A Large-Scale Empirical Study of Trademark Confusion at the TTAB

Abstract

For fifty years, the Trademark Trial and Appeal Board has applied a thirteen-factor framework to determine likelihood of confusion in trademark registration disputes. Prior scholarship has suggested that only a few factors meaningfully influence outcomes, but the hypothesis has never been tested at scale. This study presents the first large-scale computational analysis of TTAB likelihood-of-confusion decisions, using a large language model to extract scored findings for all thirteen DuPont factors from 4,003 inter partes decisions spanning 2000 to 2025.

The results are striking. A simple rule asking only two questions: “are the marks similar, and are the goods or services related?” correctly predicts the outcome in 99.55% of the 4,651 trademark comparisons in the dataset. Adding the remaining eleven factors does not meaningfully improve that number. No secondary factor contributes more than a fraction of a percentage point to accuracy, and the two-factor pattern holds at above 99% in every five-year period studied.

These findings demonstrate that the TTAB’s multifactor framework operates, in practice, as a two-factor test, with consequences for litigation costs, access to justice, and doctrinal transparency. The study advances a broader “multifactor collapse” hypothesis and outlines a research agenda examining whether other legal balancing frameworks exhibit similar patterns.

Keywords: Trademarks, Empirical, Confusion, TTAB, Trademark Trial and Appeal Board, Intellectual Property, likelihood of confusion, DuPont, large language models, multifactor balancing tests, computational legal research, legal reform

JEL Classification: K11, K40, K41, C45, C81, D73, O34, C25