The AI Chemist is a column covering what artificial intelligence technologies can do now, what they could do in the future, and what they shouldn’t tackle—all written by expert contributors.
We warn teenagers that their mistakes and lapses in judgment can follow them forever because of the internet. Somebody should have told AI. Chemical errors made by generative artificial intelligence are commonplace and have a habit of sticking around just as long as a photo of a bad haircut or an ill-considered remark, and they can be as laughably bad or even offensive. But there are good reasons to take a second look at how AI handles molecular structures. And it might be time to wipe the record clean and give AI a chance for a fresh start.
One way to ensure that AI continues to improve is for chemists to stay engaged in what is happening in that fast-moving area. If chemists do not work to make AI better, the tools will not improve as fast as they could. There are signs that the AI companies are starting to pay more attention to science, such as the recent creation of science-specific models, including OpenAI’s GPT-Rosalind, Amazon Bio Discovery, and Nvidia BioNeMo. Anthropic’s new Claude Science even incorporates a chemical drawing canvas that allows you to chat with your molecules (ChemIllusion, a company I founded, has offered this feature since its inception, and we recently open-sourced it so you can have this capability without having to trust an AI company with your lab data).
Many chemists might not see an advantage to using AI for workflows they learned to do without AI a long time ago. But the next generation of chemists will expect more. They will rely on AI, sometimes too much, and they will expect AI to be able to help them. New chemistry students will expect an AI tool that can explain anything about any molecule they can imagine—not just the examples from Wikipedia that are currently covered well in the training data of existing models.
Designing new drugs is one process that AI tools could help accelerate. There are more potential drug-like molecules than there are atoms in the solar system, but AI models can stumble on tasks that involve molecules that are not already described in their training data. AI models could handle molecules outside of their training data but only if they are given the ability to execute other programs (called tool-calling capabilities) and to access external databases.
The most recent models are getting better at chemistry, but the jagged nature of AI means we should expect chemical errors to continue. AI recently solved a fundamental long-standing Erdős problem in math, yet AI still cannot accurately add long lists of numbers. You rarely see simple math errors in chat anymore because AI models now call tools that do math deterministically rather than via next token prediction. AI is ready to contribute to fundamental chemical research, but we should still expect it to be occasionally very wrong. And just like AI does better at math when the model is connected to tools like a calculator, AI can do better at chemistry when the model is connected to accurate chemistry tools. That principle is why we need to keep chemists in the loop in AI workflows. Human oversight can ensure that the right tools are used for the right jobs. The ideal AI workflow is not having the AI model do everything. The goal should be having human eyes on key steps. Human judgment is still needed to decide what things can be calculated, what things require a specific machine learning model and where a large language model can be suitable.
I started the company ChemIllusion to help chemists get the most out of AI—not just the cheminformatics specialist but everyone who works with molecules. The tools at ChemIllusion keep the chemist in the loop. When chemists provide feedback to an AI tool about its results and the AI improves as a result, they’re not just helping it avoid future mistakes. A more reliable tool can empower chemists to spend more time doing the things they enjoy and letting AI handle the tedious parts.
When you ask an AI tool to do something and it gets it wrong, you have options. Ridiculing the AI output on social media is one option, and I too do that occasionally. What else could I do when presented with a picture of pentavalent aromatic hydrogen? But you can also correct the mistake and help the AI model “learn” from the mistake by injecting the skills or agent files back into future work. This approach is one of the ways ChemIllusion enables better chemistry than you can get from simple chat. This process is important to making any AI tool better and more reliable.
Chemists do not need to hand chemistry over to AI. But we also should not stand outside the process while other people decide how these systems will work. At the level of individual use, having a chemist in the loop means getting support while letting the AI tool do the monotonous parts, with deterministic tools checking the molecular structures and human chemical judgment guiding the key steps. At the larger level, being in the loop means chemists stay engaged in shaping the questions AI researchers ask and the tools companies build.
AI will get better at chemistry with time. Waiting for perfection is not the right move. ChemIllusion has over 40 skills available, and anyone can submit a custom chemistry skill. Other tools, from both tech giants and small teams, are also pushing the field forward. Now is the time to see what AI can do and to push AI to be better at chemistry.
Credit:
Courtesy of Scott Reed
Scott Reed is a professor of chemistry at the University of Colorado Denver and the founder of ChemIllusion, a website where chemists can draw structures and create figures, slide decks, graphical abstracts, and entire chemistry courses powered by AI tools that keep the chemist in the loop.
Views expressed are those of the authors and not necessarily those of C&EN or ACS.
