New AI Forecasts Which Patents Will Matter — With a Big Catch

Every year, tens of thousands of patents describe things that never existed before. The question that matters to investors, companies and research labs is simple: which of these will other inventors actually build on? A research team analyzed 201,710 first-of-their-kind patent technologies from US filings between 2002 and 2022. They used AI to sort those technologies into groups based on how their reuse unfolded over the following 20 years, then asked whether the first year of a patent's life contains enough clues to predict which group it lands in.

The answer was mostly yes. Seven signals visible in year one — things like how crowded the technology area already was and who was filing — predicted the group with 91% accuracy on a standard scoring measure. But here is the uncomfortable part. Simply knowing the calendar year a patent was filed scored 87%. In other words, a model that just memorizes "patents filed in 2015 tend to behave like this" looks almost as smart as one that studies the invention itself. The authors checked this and concluded it mostly reflects how patenting patterns changed over time, not something special about each technology.

They also tried a different approach: grouping technologies by their year-one characteristics alone, then checking whether those groups matched the long-term reuse patterns. They barely did — only slightly better than random guessing. That is a useful warning. A snapshot of how a technology looked at birth is not the same thing as a forecast of its future.

For anyone who consumes these kinds of AI predictions — patent analysts, venture investors, corporate research teams deciding where to spend money — the lesson is about healthy skepticism. A high accuracy number can hide the fact that the model is leaning on an easy shortcut. Big forecasts about which technologies will take off should be read with the question: is this really about the invention, or just about when it showed up?