AI is Excellent at the Discovery and Identification of Unrecognized Scientific Knowledge but There Is No Evidence That AI Can Create New Scientific Knowledge

Abstract

AI has the ability to identify scientific and clinical relationships previously not recognized, but there are no examples where AI can take this information to create new scientific knowledge. New scientific knowledge is created by assembling identified relationships that lead to hypotheses, using this information to design experiments and learning from the results of these experiments. AI can greatly facilitate the assembling of the identified relationships, and in science perhaps, AI should be the abbreviation for “Artificial Identification.” The above prototype served as the basis for the founding this past Fall (2025) of Periodic Labs (Periodic.com) by a team that “co-created ChatGPT, DeepMind's GNoME, OpenAI's Operator (now Agent), the neural attention mechanism, MatterGen; have scaled autonomous physics laboratories; and have contributed to important materials discoveries of the last decade.” The team writes “We've come together to scale up and reimagine how science is done” and they have received over $300 million in seed funding.

A September 30, 2025, New York Times article titled “AI Scientific Discovery and Periodic Labs,” states that so-called neural networks, that is, AI systems that facilitate chatbots like ChatGPT, identify patterns in vast amounts of information on the internet to mimic the manner that words are put together and “can even write computer programs and do math problems.” The article then cites two of Periodic's co-founders Liam Fedus (co-creator of ChatGPT and former VP of OpenAI) and Dogus Cubuk (past lead of materials science and chemistry at GoogleDeepMind) as believing that “no matter how many textbooks and academic papers these systems analyze, they cannot master the art of scientific discovery. To reach that, they say, AI technologies must also learn from physical experiments in the real world.” Dr. Cubuk states that a chatbot “can't just reason for days and come up with an incredible discovery. Humans can't do that, either. They run many trial experiments before they find something incredible — if they even do.” Periodic plans to build its own lab “where robots — physical robots — will run scientific experiments on an enormous scale.” Periodic.com asserts, “We're starting in the physical sciences.”

Will Periodic's paradigm be successful in generating new knowledge in clinical pharmacology and clinical medicine? Can physical robots run scientific experiments that will have relevance in clinical pharmacology and clinical medicine? To do so, one must assume that in vitro–in vivo extrapolation (IVIVE) will work. My position is that Periodic's paradigm will frequently not be successful in creating new knowledge in clinical pharmacology and clinical medicine since in vitro studies, a powerful tool in the physical sciences, may not recognize the inherent differences between chemistry and pharmacokinetics.

Since mid-2022, we have published a series of papers demonstrating that all drug metabolism and pharmacokinetic clearance and total rate constant relationships could be derived independent of differential equations and mechanistic models of hepatic elimination using parallel and in-series rate-defining processes adapted from Kirchhoff's laws in physics.¹ Our most recent publication² “Mathematical Modeling in Drug Metabolism and Pharmacokinetics: Correct In Vitro, Not Always Valid In Vivo,” addresses the topic of this Commentary. Historically, rates of chemical reactions and subsequently in vitro drug metabolism rates were derived using differential equations based on mass balance considerations for measurements in a beaker with a fixed volume. Thus, mass balance is a necessary and sufficient criterion to describe these reactions. However, drug dosing decisions in clinical pharmacology and medicine are almost always based on systemic concentrations measurements, not amounts, so that mass balance is a necessary but not a sufficient criterion. A very slow rate-defining clearance for a drug delivery process may not occur in the same volume of distribution as systemic drug clearance following an iv bolus dose.^{2, 3} Therefore, driving force concentrations at the delivery site and in the measurable systemic fluids may occur with different distribution characteristics, and thus a different relationship between the amount and concentration of drug at the different sites. Differential equation derivations can only be converted from amounts to concentrations using one volume value^{1, 2} and thus cannot correctly characterize the relationship. However, prior to our introduction of the derivation of clearance and rate constant relationships based on correctly summing parallel and in-series rate-defining processes⁴ all pharmacokinetic relationships were derived, as in chemistry, based on differential equations.

Returning to the title of this Commentary, I strongly agree with the founders of Periodic Labs that AI cannot create new scientific knowledge. Creation of new scientific knowledge results from assembling identified relationships that lead to hypotheses, using this information to design experiments, and learning from the results of these experiments. AI can only facilitate the assembly of the identified relationships. However, the Periodic paradigm for creating new knowledge in the physical sciences, using robots to run scientific experiments on an enormous scale, will not necessarily lead to success in clinical pharmacology and clinical medicine where IVIVE cannot be assumed. Thus, the necessary in vivo experiments will be much more laborious and time consuming but they do provide the avenue for creating new scientific knowledge in clinical pharmacology and clinical medicine that cannot be accomplished by AI alone. But as per the quote from Dr. Cubuk above: Humans may also not possess the necessary ingenuity required to speculate on the yet-to-be identified, unidentified relationships from these many in vivo trial experiments. That is science and why it may progress slowly.

The author declares no conflicts of interest.

No funding was received for this Commentary.