AI for Science Has a Bad-Data Problem w/Intellicat
At our last salon, Paul Litvak spoke about the urgent need for a verification layer for science. On October 14th, come listen to Intellicat's CEO, Valentin Rodionov, explain how they are building exactly that!
Abstract: Government and industry are putting AI agents to work on science. These systems read the literature, generate hypotheses, and design experiments. But they cannot reliably tell good science from bad.
At Intellicat, we have built TRACES, an open benchmark that tests exactly that. We gave 30 leading models 42 known-unreliable papers and asked them to design follow-up research. They went along 93% of the time. Some identified the paper as discredited, then built on it anyway.
The problem is bigger than retractions. Millions of unreliable papers remain in the scientific record and are treated by AI systems as ordinary evidence. Intellicat Runtime adds the missing credibility layer: scoring roughly 500 million documents and making those signals available to AI systems at inference time.
This talk presents the failure, the evidence, and a practical fix.
Speaker Bio: Valentin Rodionov is founder and CEO of Intellicat and a Research Associate Professor at CWRU. A chemist by training, he helped uncover the mechanism of the copper-catalyzed click reaction at Scripps Research with M.G. Finn and K. Barry Sharpless, work underlying the 2022 Nobel Prize in Chemistry. In 2022, his group reported the synthesis of graphyne, a new carbon allotrope.
Rodionov has spent years confronting unreliable science firsthand: as a domain expert trying to reproduce published chemistry, as a scientific sleuth whose investigations have contributed to multiple retractions, and now as an AI researcher. He built TRACES, an open benchmark showing that leading AI models accept flawed science 93% of the time. Intellicat scores roughly 500 million scientific documents for credibility before AI systems build on them.
About the DC Industrial Policy Salon
The DC Industrial Policy Salon was started by Nicholas Wagner, Founder of Learning Journey AI, in 2026. It convenes leading experts and policymakers to discuss issues of science, technology, and industrial production in Washington, DC. To learn more about our events, please visit https://dcips.net.