When navigation programs such as Google Maps debuted, they transformed our ability to navigate quickly and accurately to a destination. At the same time, this convenience has reportedly led to a decline in our spatial memory and learning skills.
Today, large language models (LLMs) present a similar conundrum, promising to rewrite the ways we navigate the landscape of scientific research but threatening to degrade our human thinking skills. Already, LLMs can speed computer coding by an order of magnitude, automate data analyses and accelerate the process of scientific writing and editing. Artificial-intelligence (AI) agents have broken records for analyses of biological data and shown promising results in automated hypothesis generation and scientific discovery.
Going forward, improved AI architectures and implementations will only increase AI’s capabilities and accuracy. But, unlike with automobile driving, for which we can just provide a final destination and then hand over the steering wheel, scientists will for the foreseeable future continue to be the ones who steer AI. How can we train the next generation of neuroscientists with the ability to perform such AI steering? What knowledge from traditional training protocols needs to be maintained, what can be discarded, and what new types of training should be introduced?
In addressing such questions, we first need to consider a more fundamental question: “What do future neuroscientists need to know?” Knowledge comes at different levels. At lower levels, we teach facts and basic concepts, and simple problem-solving based on these. In teaching the Hodgkin-Huxley model, for example, we present facts about the voltage dependence and kinetics of sodium and potassium channel gating, connect these to concepts such as positive and negative feedback, and ask students to apply these facts and concepts to simple problems, such as why a brief hyperpolarizing pulse of current can trigger an action potential. But the ultimate goals of neuroscience research and education sit at a higher level: Analyzing complex scenarios, such as the voltage dynamics of real neurons with many channels; evaluating different courses of action, such as selecting among proposed drugs to treat a channelopathy; and generating new ideas, such as designing one’s own candidate drug. Framed this way, the key questions become what lower-level knowledge is needed to address higher-level questions, how to impart this knowledge effectively, and how to provide training in higher-level problem-solving.
With regard to lower-level problem-solving, recent studies have found that those who use AI regularly perform worse on the same tasks when AI is removed. However menacing this sounds, it may miss the point. For example, when we teach children how to multiply large numbers and perform long division, we have no intention that they will continue to do so rather than using calculators (or that as adults they will still remember how to do long division). Rather, we teach these to build the numerical literacy that underlies quantitative reasoning and the “quantitative common sense” that allows us to identify errors when mistyping a number into a calculator.
Likewise, in analyzing neural data, remembering precise syntax or how to program low-level routines is of little importance when working with an AI copilot. But having the rigorous statistical and algorithmic reasoning needed to formulate a problem precisely and evaluate the plausibility of an AI-assisted solution remains crucial. Such “formulate-and-evaluate” cycles are critical to working with AI across realms and should be added to courses to refine student thinking, replacing presentation of factual details or mechanical calculations beyond that needed to deliver concepts that students can work with in formulate-and-evaluate exercises. For example, when learning Hodgkin-Huxley modeling, students could steer AI to design a model that captures the voltage dynamics of an instructor-provided simulated or experimentally recorded neuron. Alternatively, AI can be the evaluator that provides feedback on student-proposed ideas.
How AI is implemented in the classroom is key to both developing foundational skills and learning to solve higher-order problems. Education research shows that we do not learn well by watching someone else solve a problem. Thus, if students use AI too extensively to help them answer questions and essentially only observe the solutions, they won’t build and retain foundational reasoning skills. However, used correctly, AI can dramatically increase the educationally proven practice of “active learning,” in which we are repeatedly given problems to solve and rapid feedback on our answers.
Further, AI can set up curricula tailored to each student, with progressively challenging problems, appropriate hints when students become stuck, and the ability to diagnose when a core skill is missing and offer a tutorial to address it. For example, in preparing this article, I asked an LLM to teach me why ANOVA is equivalent to a regression model. I further instructed it to not give me the answer but rather to step me through how to come to the answer myself. Within a short period of time, the LLM guided me through a progressively more sophisticated set of examples and exercises, identified when one of my answers was imprecise, and provided me with novel insights when it made me reason through a seemingly paradoxical result caused by statistical undersampling.
AI may have even more utility in training higher-level reasoning. Because AI can solve problems requiring higher-order reasoning, it can serve as a personalized tutor that facilitates and guides students through project-based learning activities involving, for example, hypothesis development, experimental design, or analysis and interpretation of complex datasets. Alternatively, entire courses can be designed around formulating, analyzing and evaluating AI-assisted solutions to challenging problems. At the highest level, AI agents have already been able to develop new ideas and approaches, and this capability will only grow as AI matures.
But it remains the job of human scientists to identify the most important questions and gaps in knowledge in need of new ideas. Acknowledging this shifting landscape, the course “Genuinely Hard Problems in Science,” taught by neuroscientist Jeff Lichtman at Harvard University, introduces students to grand challenge problems and engages them to design an undergraduate curriculum that provides the knowledge and skills to solve them.
Finally, the advent of AI as a personalized learning system means that faculty need to reconsider their role as teachers. Despite decades of research on the benefits of active learning, faculty and students alike remain enamored of a beautifully delivered lecture. AI raises the stakes for continuing this form of passive learning, given the opportunity AI offers for active, personalized instruction. Educational research suggests that teaching expert thinking is most effectively accomplished by leading students to discover the structure of knowledge through hands-on problem-solving. Faculty will need to learn to lecture less and serve more as the human facilitators who introduce, motivate, guide and reinforce problem-solving lessons hand-in-hand with AI tutors. Generating and broadly disseminating course materials that foster such AI-integrated training is a needed next step. Doing so in a deliberate and thoughtful manner will foster a new generation of students with both deep-seated knowledge and the ability to fluidly work with increasingly intelligent AI agents to solve the greatest problems ahead of us.
To understand how others are thinking about the issues discussed in this article, I asked six neuroscientists the following question: What knowledge from traditional training protocols needs to be maintained, what can be discarded, and what new types of training should be introduced?
