AI trademark search

Is your proposed trademark logo infringing or not? The WIPO Artificial Intelligence (AI) trademark search engine can provide helpful insight.

Logos (images associated with a product or service) are essential to trademark law. It is easy to tell when an infringer uses an exact copy of a registered mark. However, the process gets more and more subjective as the marks differ. Another problem is that companies often differ in legal budgets and aggressiveness. The history of trademark enforcement shows many examples of trademark bullying, where a given mark holder may act unreasonably.

Patent offices are well aware of the increasing effectiveness of machine vision and AI in helping their examiners. Several international patent offices, such as the World Intellectual Property Organization (WIPO), have recently introduced user-friendly interfaces that allow the general public do such searches free of charge.

Do these AI systems work?

These AI trademark logo search systems are new and come with little or no documentation or validation. There are also not many published papers. Here, I report a few tests of the WIPO system that confirm that the system does work, at least to some extent. In this quick initial study, I challenged the WIPO system with several types of trademark logos. I took some from established trademark infringement trials, and some were taken as “controls” to check system function. I recorded these trials’ first three hits (out of hundreds of reported hits). The results are shown below.

Image recognition ability. How well does the system understand what it is looking at? Can it distinguish between images like the free software GNU (with horns) and images like the open BSD daemon (with horns)?

AI check of trademark logos similar to GNU and BSD

Here, the WIPO system scanned 222,058 results and clearly understood that the GNU image had the best similarity to other bovine-face-looking-at-you-like trademarks. The system also inspected 11,464 results and understood that the BSD daemon had the best resemblance to other non-threatening devil or monster type trademarks. So, not bad. The system got the general category and features about right.

What about known trademark disputes? How well does the system anticipate the outcomes?

Apple is an 800-pound gorilla with a distinctive “apple with a bite missing” trademark logo. Apple usually prevails in its trademark disputes. But not always. In 2013, Apfelkind, a small German cafĂ©, succeeded vs Apple. Their mark was an image of a child’s face framed by an apple, with the words “apfelkind” below in cursive. What does the WIPO AI system think?

AI check of trademarks similar to apfelkind logo

The results show that more image data is better. The WIPO AI system scanned 1,005,327 results and correctly identified that the “face in apple” part of the logo was some framed face and even suggested a different human-face-in-apple design to consider. However, when fed the entire logo, the system scanned 148,959 results and immediately picked the correct identification. In neither case did the WIPO AI system suggest a logo owned by Apple computer.

What about a case where the trademark owner (here, Disney) mostly prevailed?

AI check of trademarks similar to deadmau5

A few years ago, “deadmau5” attempted to trademark a caricature of the distinctive Disney Mickey Mouse ears mark. Although the Mickey Mouse copyright will expire shortly, trademarks can potentially last forever. Further, “satire” or “parody” defenses don’t work for trademarks. Thus, Disney was generally successful in limiting the use of this mark, and the AI study suggests that their objections may have been reasonable.

Note, however, that the search settings are important. When we restrict the search to just active US trademarks (top), the Disney marks immediately appear in two of the first three hits. By contrast, if we open up the search to any status mark (including inactive) worldwide, then the first active Disney marks only appear later in the search. This shows the importance of proper search parameters. It also shows the importance of examining at least the first few hundred hits rather than terminating the search at the top three.

However, again, the WIPO system appears to be functional.

Suggested uses:

It looks as if AI trademark image systems are effective. Although human judgment is still needed, AI searches should become a routine part of standard trademark work. Some final thoughts and suggestions:

  • AI searches can be a valid “sanity check” for trademark applications.
  • Consider AI searches as a way to help respond to allegations of trademark infringement. Here the response strategy can vary depending on the AI findings. Further, at least in cases where the AI findings suggest that there is no close match, consider using such findings as part of a response.
  • Since online AI systems are still in their early stages, consider running various control tests with your analysis to help demonstrate that the AI system is working adequately.