在 AI 时代重新思考神经科学研究生教育

What knowledge from traditional training protocols needs to be maintained, what can be discarded, and what new types of training should be introduced? Six neuroscientists weigh in.

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?

How do we decide what knowledge and skills are worth keeping? The question of how technology renders skills obsolete is not new. We no longer teach numerical integration on a hand-cranked calculator, as Andrew Huxley did when modeling the action potential; and engineers no longer solve equations with nomograms, charts that let users read off answers with a straightedge. Automation can replace the need for a skill, but only if it fails visibly. Programmers no longer learn to write machine code, because when a compiler can’t translate a program, it stops and reports the problem. But an AI system that confidently reports a new type of predictive neural activity needs a human monitor who can recognize that the stimulus and response were simply misaligned in time, and that recognition requires understanding the analysis.

A skill, even an obsolete one, should also be retained if it is the foundation for deeper understanding. If it is, it can often be taught in a reduced form. Experimentalists who take an introductory machining course and build something by hand often become better designers. Similarly, computing backpropagation by hand is still taught because it helps students understand why models fail, for example from vanishing gradients, and how to fix them. AI systems may diagnose such failures better than most people can, but a person cannot judge a diagnosis they don’t understand, and researchers designing new model architectures will face problems no one has seen before.

In this regard, AI is not a new kind of problem but the newest instance of an old one. Every tool that has taken over human work, from the calculator to the compiler, has forced us to decide which skills we can safely let go, which we must keep in order to know when the tool is wrong, and which we must preserve as a foundation.

For the skills we choose to preserve, AI also opens exciting new possibilities for teaching, provided it is used well. Students given unrestricted access to AI can end up learning less, because the AI does the thinking for them. But AI tutors that can draw on many proven teaching methods and tailor their approach to each student offer real promise. Because explaining material clearly reveals gaps and deepens understanding, learning-by-teaching agents and Socratic tutors could help solidify concepts and skills in ways that are difficult to achieve with large classes and traditional methods.

Human teachers nonetheless remain essential. They offer individual perspectives and connect material to other disciplines and to students’ future goals. They also provide mentorship, motivation, judgment about what matters, and a model of how an expert thinks. To make the most of these strengths, lecturers could often benefit from enlisting a probing AI to improve the clarity of their lectures.

Knowing low-level facts, such as sodium channel kinetics, and mastering low-level techniques, such as writing a thresholding function in Python, shape our research in two ways. Occasionally, they directly support our work, but more often they help indirectly, by contributing concepts to our intuition: our internal model of the world and the logic with which we reason from it. Which is why we believe it important to teach math to all children, whether they’ll use those skills or not in their career.

Whereas search engines removed the need to memorize facts, LLMs remove the need to execute techniques: Derivations, coding and data-wrangling will be done faster, and often better, by an LLM. This means we do not get to “learn on the job” anymore: In graduate school, most of us learned statistics and programming while wrangling with data and debugging code. Our intuition grows as a byproduct of doing the work; when agents do the work, intuition stagnates. As Terence Tao has put it, LLMs are like a helicopter flying us to our objective; and we learn a lot through the process of reaching our objective (for instance, that there may be a better objective!).

Yet our intuition is what inspires the research question we’ll ask next and what enables us to evaluate the output of LLMs. As the craft of a scientist slowly turns into smart agent management, our ability to choose questions and evaluate agents will arguably become all that matters.

So we must develop the discipline to carve out protected time for intentional intuition-building, through learning facts and techniques—for instance, by using LLMs as tutors rather than ghostwriters, as Goldman suggests. Without that discipline, the knowledge our agents generate will outpace our ability to understand it, and we humans will lose the ability to steer them.

I really love the idea of this article, and, indeed, I have been struggling with exactly the question you ask. Like, really struggling. I think it is unclear what knowledge from traditional training protocols needs to be maintained. The reason is that, as someone whose training is long since complete, it is hard for me to know which training experiences are critical for my current expertise. For instance, it seems at first crazy that students could, thanks to LLMs, now only learn the basic rudiments of writing code; surely failing to gain deep expertise would mean they missed out on the critical thinking skills that I feel I gained from that experience. But I must doubt myself, because I have seen too many times that principal investigators argue for retaining skills that turn out not to be essential. For instance, I was berated by a principal investigator when I was in graduate school for not having built my own amplifier in my electrophysiology rig. They said, “Building the amplifier is a key part of the science. Without having built it yourself, you don’t truly understand the signals you are measuring!” But actually, I have done OK without ever learning to build my own amplifiers. 

Still, we need to be continually thinking deeply about what skills can and cannot be abandoned in the age of AI. And we need to keep revisiting this once we see the consequences of opting out of certain training skills. 

As we all know, error signals are the drivers of learning. This requires training or learning in an active mode: taking on a question, attempting an answer, making mistakes, being confounded by them, experimenting, and correcting. In whatever mode AI-driven learning occurs, making sure that attempt, confusion and resolution remain in the human part of the cycle seems vital. Each of us must eventually learn to recognize errors, one’s own and others’, as best we can without external feedback; otherwise we risk weakening the ability to take responsibility for the work. To preserve this, we need to better identify how the ability to recognize error, or critical thinking, emerges. I think it remains to be seen what level of engagement with the solution process builds this skill.

Neuroscience graduate education hasn’t changed very much since I was a graduate student 50 years ago. Scientific knowledge has grown dramatically, and today’s students have much more to learn. Despite advances in technology that have made journals and books more accessible, classroom teaching is pretty much the same. AI is impacting many areas of science and society. How will the recent advances in AI make a difference in education?  We are at the beginning of a long process of integrating AI into education at every level, from K-12 to graduate training. 

When LLMs were introduced by OpenAI in 2022, New York City Public Schools—the largest school district in the United States—quickly implemented a strict ban on them on its networks and school-issued devices. The danger was cognitive offloading: If AI can supply answers and write essays, student learning would suffer. Enforcing this ban was impractical, and the school district soon reversed it. However, managing the harm cognitive offloading can do to learning remains central to AI’s future use in educational settings.

I helped organize a workshop sponsored by the Ernst Strüngmann Forum in August 2026, called “Memory, Learning, and Neural Adaptation in the Age of Cognitive Offloading.” We invited 50 experts in education, neuroscience, cognitive science and educational technology to meet for a week and discuss cognitive offloading from many angles in depth. Cognitive offloading has a long history in human culture. Books, for example, reduced the need to memorize. More recently, the internet has made the world’s knowledge even more accessible. AI is just another step in the same direction. Education accommodated books and the internet, but AI is so new that it’s not yet clear how best to integrate it. 

One recent randomized controlled study explored how different uses of AI affect student learning. When students used AI to get answers and write essays, they learned much less than when they used AI more like a tutor to explain complex concepts or clarify ideas; their test scores were higher, and their essays were higher quality. We need to teach students best practices for how to use AI to help them become better learners. Once a student masters using AI to enhance learning, it will support lifelong learning.

The most efficient way to teach a child is one-on-one tutoring with an adult who knows the student and is well qualified to help the student make progress. Tutoring is much more labor-intensive than teaching in classrooms with dozens of students from different backgrounds. Incorporating AI into the curriculum will eventually require rethinking the current assembly-line model in schools. For graduate students, an AI tutor might become an AI mentor.

In a famous song from 1965 entitled “New Math,” satirical singer-songwriter Tom Lehrer parodies the then-new approach to mathematics education that ensured that students “understand what you’re doing” rather than “get the right answer.” The song was reportedly “enjoyed by new math proponents and critics alike,” perhaps because both sides felt that it was a fair characterization of their position. AI can seem to pose a similar dilemma, pitting the quick-and-right-answer approach against learning the fundamentals.

I have appreciated AI as both a tool for learning and a tool to avoid having to learn (e.g., coding up a website without learning HTML). The key difference is generally my own motivation to spend time and mental effort on the topic, which is in turn informed by my beliefs about whether that knowledge will support some future creative endeavor. Unfortunately, I might be wrong about that a lot, because it’s a really hard problem to anticipate what knowledge one would most like to have in the future to solve a set of problems that haven’t yet presented themselves. 

I agree with Goldman that an important role for teachers and advisers is to help students decide what they most benefit from knowing versus what can be safely outsourced to technology, anchored in an end goal of solving big unknowns. Of course, unlike with “new math,” there isn’t really an algorithm for this. In practice, the distinction between the fundamental and the trivial can vary a lot between fields, and sometimes this is just a matter of taste. Knowing what can be safely outsourced also depends on hard-to-formalize information about limits of the technology. Whereas the scope of a calculator is limited and its behavior reliable in that scope, interpreting the outputs of an AI system is harder. The discipline in which LLMs have been probably the most prolific is mathematics, and these outputs rely heavily on the fact that proofs in this domain are separately verifiable. We have seen in our own work that wrangling LLM outputs into interpretable insights about the process can be hard to do and requires thinking hard about what form of knowledge you are trying to obtain. AI provides both challenges and opportunities for training the capability for creative thought, and I agree with Goldman’s argument that navigating this effectively will likely require thinking deeply about the role that knowledge serves in scientific practice.

tags:

注意转载授权 · 阅读原文