Indianapolis—Artificial intelligence wasn’t explicitly the theme of the Chemical Innovation Exchange (CIEX), but it may as well have been.
The conference, which was held in Indianapolis Sept. 9–10, hosted companies offering AI-enhanced software intended to help chemical makers discover new materials and run their laboratories. Also there were chemical company chief technology officers and research executives who are grappling with how to integrate AI into the innovation process. As scientists, they are often skeptical about the power attributed to AI, but as executives they know it will be a big part of their industry in the future.
“The biggest risk of AI is that it can create the illusion of wisdom, and this is what scientists hate.”
One of the event’s sponsors was the American Chemical Society, which publishes C&EN but is not involved in editorial decisions.
The software providers use both generative AI and machine learning to offer a variety of research services. Basetwo, for example, layers AI on a physics-based understanding of how chemicals and equipment interact to reduce experimentation time and manufacturing costs. Cypris and GetFocus apply AI to patent and other data to predict technological trends and help researchers decide which ones to pursue.
Albert Invent and Uncountable are all about execution once a research direction is decided. Their software can digest a customer’s internal R&D data, layer external data, then use AI to help interrogate it all. CAS (a division of ACS) and Elsevier, meanwhile, marry AI with proprietary patent and academic publishing data from their organizations’ publishing arms.
In presentations, as well as in conversations during breaks, executives from these firms took pains to distinguish what their tools deliver from what a chemist can retrieve by querying a general-purpose AI model like OpenAI’s GPT products or Anthropic’s Claude. Dan Wilson, vice president of sales at Albert, noted that those models are notoriously bad at answering chemistry questions. Even specialized versions of those models, like Claude Science, often don’t get it right, he said.
Building trust in chemistry AI
Trust in the results that chemistry AI products provide was a big theme in talks at the event. Stuart Whayman, president of Elsevier’s corporate markets unit, pointed to a study conducted by his firm (PDF) that found only one in five researchers trust AI for performing critical tasks like new product formulations or regulatory submissions. “It’s not good enough to hoover up as much information as you can and run AI on that,” Whayman said.
Andrea Jacobs, senior director for AI at CAS, argued in a presentation that AI-guided research needs a process equivalent to peer review to earn the confidence of scientists.
She proposed that AI peer review start with data: knowing where information comes from and who is accountable when it’s wrong. Next, researchers must be involved in any task AI is asked to perform; they need to determine what an AI agent can decide on its own and where it must stop. Finally, AI peer review should include human judgment to decide what to trust and whether a tool is meeting an organization’s needs.
“At the end of the day, you and the scientists on your team are the ones who hold the judgment,” Jacobs told the audience.
In their talks on stage, both Whayman and Jacobs emphasized the pedigree of the scientific information their tools can draw from. When it comes to scientific data, the important factor is not the quantity, Whayman told C&EN before the talk. Service firms can’t be “just aggregating everything you can find on the web but [should be] having a mind of curation and quality as to what it is you are bringing into your databases and into your tools.”
Regardless of where they draw their data from, all the software firms at the conference emphasized the importance of accurate and verifiable data. “We are data first,” said Timothy Donaghy, a business development leader at Uncountable. “AI is useless if you don’t have good data.”
Getting skeptical scientists onboard with AI
Despite such assurances, researchers at the bench can be skeptical about the promise of AI. In a talk on using generative AI in research, Jayshree Seth, a corporate scientist at 3M, cautioned software companies and corporate executives against pushing too hard or promising too much. “The biggest risk of AI is that it can create the illusion of wisdom, and this is what scientists hate,” she said.
Seth said 3M scientists are using generative AI for practical purposes, such as synthesizing technical literature, summarizing patent information, querying and distilling internal technical reports, and finding in-house subject matter experts.
“Generative AI is a very powerful tool for very specific things,” Seth said. “You need to be very, very judicious about where you use it.”
Speaking to the executives in the audience, Seth said one way to encourage the adoption of AI in the lab is to find champions like her who are in the trenches with other scientists rather than upstairs in the C-suite with the executives.
Researchers won’t embrace AI blindly, Seth said, but executives may be surprised at how readily they do adopt it to speed through the more onerous parts of their jobs. “Reading literature, figuring out how to boil it down into a PowerPoint, looking at the patent landscape to understand what to avoid and what we’ve already said”—these are the kinds of tasks researchers want help with, she said.
In fact, Seth said, the uptake of AI tools has been fairly quick at 3M. “Eventually, what we got is people saying things like, ‘This is a game changer,’ ” she said. “If you work with people’s pain points, they are likely to follow.”
