MIT's Rakhlin: Universities Must Prepare for AI Smarter Than Us in Most Intellectual Work

Why research is changing so fast

Rakhlin, the Distinguished Professor in Data, Systems, and Society, IDSS and Brain and Cognitive Sciences, has written a new essay called "AI and Academia". It synthesizes recent discussions and readings about how AI is changing academia, particularly mathematics, statistics, machine learning and engineering. The interview focuses mainly on graduate research and education.

He uses mathematics as an example of the pace of change. According to Rakhlin, a model reached gold-medal level at the International Mathematical Olympiad last year. A year later, models are producing new research results, including a recently proposed solution to one of the Millennium Prize Problems.

A central factor in how fast AI advances in a discipline, he says, is the speed and reliability of verification. Formalized proofs can be checked automatically, and programs can be run and tested. Where evaluation is fast and reliable, systems can generate candidates, learn from outcomes and improve. The same applies to AI research itself, where better models can contribute to the next round of development and create a compounding process.

For experienced researchers, Rakhlin sees opportunities to pursue questions that were previously out of reach for technical reasons. He also says the distinction sharpens between obtaining a solution and understanding why it works, what generalizes and what to ask next. He cautions against assuming that abstraction, judgment or problem formulation will stay beyond AI.

Rethinking credit and graduate training

Rakhlin argues that in a growing number of fields, a polished paper is becoming a weaker signal of individual expertise. Departments should begin now to reconsider what they reward, he says, without simply defining valuable work as whatever AI cannot yet do. He suggests that asking good questions, replication, synthesis of ideas, informative negative results and shared datasets may deserve greater recognition.

Evaluation, in his view, should establish what a researcher contributed and takes intellectual responsibility for, including when substantial parts of the work were done by AI. These expectations should guide hiring, promotion and funding, and be made explicit for current and incoming PhD students.

Training is the harder problem, according to Rakhlin. Routine calculations, coding, failed approaches and small discoveries have long helped students build intuition and judgment. Delegating that work can remove formative experiences, even as AI lets students attempt more ambitious projects. He calls for distinguishing unnecessary friction from work through which expertise develops.

• Students should learn to formulate problems.
• They should audit model outputs and reproduce results.
• They should be able to defend their choices.

Fundamentals, he says, may become more valuable as the basis for using these tools well.

Building shared AI infrastructure across laboratories

Rakhlin says universities should be able to pursue questions over long horizons, share results openly and evaluate claims independently. Partnerships with industry will be essential, he says, but commercial priorities may not cover the breadth of science or stay aligned with it, so some degree of technological independence will be necessary.

He describes academia's accumulated laboratory knowledge as its most important strategic asset. Published papers are a selective record, since failed experiments, abandoned directions and the reasons an approach did not work often go unpublished. This missing context may help explain why models currently struggle in some domains, especially the empirical sciences, to anticipate consequences that are clear to an expert. Rakhlin notes these limitations may be temporary, and that capturing negative results and scientists' interpretations could make models much better at scientific exploration.

His vision is for MIT's laboratories to function more like a single, living, scientific organism, connected through shared AI research infrastructure. In his example, a neuroscience lab needs better methods for segmenting neurons in microscopy images. An AI agent could recognize a relevant advance from a computer vision group, connect the researchers, propose benchmarks and help them iterate. It could also surface unresolved questions to students and faculty and share lessons between labs.

Achieving this would require workflows that capture hypotheses, interventions, outcomes, failures and interpretations, plus shared systems that let agents use tools and information across laboratories with appropriate permissions. Rakhlin says this needs substantial public and institutional investment now in compute, secure data systems, and expertise in adapting and post-training AI models.

What it means for universities

Rakhlin says tracing this process could preserve the lineage of ideas and make contributions, including those of graduate students, easier to recognize. With agreed rules for consent and credit, researchers could collaborate more openly, with greater confidence that their work would be acknowledged. He adds that this could help answer the question of how to recognize and reward intellectual contributions.

He calls this a promising path forward, but stresses that universities need to invest now to make it possible. The interview also notes that MIT published a report on AI and education this summer, a sign of the attention the topic is receiving.