Tool Reviews

Three Questions on the Optimal Way Forward for AI in Academia

AI is changing academia by accelerating research, challenging credit, and affecting training; universities need standards and shared infrastructure.

Three Questions on the Optimal Way Forward for AI in Academia

Key Takeaways

  • AI is rapidly reshaping academic research, especially in mathematics and engineering, because machine-verifiable proofs and testable software allow fast feedback and iterative improvement.
  • A polished manuscript is becoming a weaker signal of a researcher’s knowledge, so universities should rethink how they credit contributions such as asking strong questions, replicating studies, and sharing data.
  • Graduate training still needs productive struggle: students must retain core skills like problem formulation, result checking, and judgment, even as AI handles more routine work.
  • Universities should build shared AI-driven scientific infrastructure to connect labs, preserve tacit knowledge and failed experiments, and support long-term, open, independent research.

Meta description: MIT's Sasha Rakhlin explains how AI is reshaping graduate research, credit, training, and university science infrastructure for mathematics and engineering.

Three Questions on the Optimal Way Forward for AI in Academia

Tags: AI in academia, graduate research, MIT, Sasha Rakhlin, scientific infrastructure

![Abstract friendly illustration of AI tools connecting university labs, equations, code, and research graphs, with no people, logos, or text](assets/ai-academia-featured.png)

A new essay by Sasha Rakhlin, director of the MIT Statistics and Data Science Center and the Distinguished Professor in Data, Systems, and Society, IDSS, and Brain and Cognitive Sciences, explains how AI in academia is changing graduate research and education. It follows an MIT report on AI and education published this summer and focuses on mathematics, statistics, machine learning, and engineering.

Why mathematics shows the speed of change

Rakhlin points to mathematics because it makes the pace of change easy to see. Last year, a model achieved gold-medal-level performance at the International Mathematical Olympiad. Just a year after that, models are producing new research results, including a recently proposed solution to one of the Millennium Prize Problems, a set of well-known unsolved mathematical challenges. Tasks that once showed advanced mathematical ability can increasingly be handled by AI.

The central condition, he argues, is how fast and how reliably work can be checked. A proof written in a formal language can be verified by a machine, while software can be executed and tested. When feedback is quick and trustworthy, systems can create candidates, learn from the results, and improve. That loop also applies to AI development itself: better models can help improve training procedures and tools, creating a compounding effect.

Three Questions on the Optimal Way Forward for AI in Academia

For experienced researchers, this opens paths to questions that were previously too hard to pursue. But Rakhlin warns that getting an answer is not the same as explaining why it works, knowing what transfers, or choosing the next question. He says universities should prepare for a near future in which AI may be smarter than humans in many, perhaps most, areas of intellectual work.

Why a polished manuscript is no longer enough

In more and more fields, he says, a polished manuscript is becoming a less reliable sign of what a person knows. Departments should start now rethinking what counts as valuable scholarship. The answer, in his view, is not to define valuable work as whatever AI cannot yet do. Instead, universities may need to give more recognition to asking good questions, repeating studies, bringing ideas together, reporting results that rule out a path, and sharing datasets.

Evaluation should also clarify what a researcher actually contributed and what intellectual responsibility they are taking on. That becomes especially important when large parts of a project are produced with AI assistance. Rakhlin says these standards should guide hiring, promotion, and funding decisions. They should also be made clear to students already in doctoral programs and those about to start.

Why graduate training cannot skip the hard steps

Training is the harder problem. Rakhlin compares intellectual skill to muscle that grows through repeated effort. Ordinary calculations, programming, dead ends, and small wins have long helped students build intuition and judgment. AI can remove some of this work, which may let students attempt more ambitious projects. The risk is that it also removes the experiences through which expertise forms.

Departments, in his view, need to separate unnecessary friction from productive struggle. Students should still learn to state problems clearly, check what models produce, repeat published findings, and justify their decisions. As tools become more powerful, the basics may become more valuable, not less, because they are what allow researchers to use those tools well.

What universities should build: shared scientific infrastructure

Rakhlin sees an opportunity for MIT and other universities to create infrastructure that helps science move forward on its own terms. The goal, he says, is for universities to work on questions over long time horizons, publish results openly, and assess claims independently. Industry partnerships will matter, but commercial priorities may not cover the full breadth of science or stay aligned with academic needs over time. Some technological independence is therefore necessary.

Three Questions on the Optimal Way Forward for AI in Academia

He argues that one of academia's most valuable assets is the knowledge held inside laboratories, not just in published papers. Papers are selective. Experiments that failed, directions that were dropped, and the explanations for why something did not work are often absent from the public record. Scientists and engineers also carry tacit knowledge, or hard-to-write-down know-how, about what is likely to fail and why. This missing context may help explain why AI models currently struggle in some domains, especially the empirical sciences, to anticipate consequences that are obvious to experts. Those limitations may be temporary, and capturing negative results and expert interpretations could make models better at scientific exploration.

He imagines MIT's laboratories working more like one connected, evolving scientific system, linked through shared AI research infrastructure. In one example, a neuroscience lab wants improved tools for separating neurons in microscope images. An AI assistant could spot a useful development from a computer vision team, introduce the groups, suggest evaluation tests, and guide repeated improvement. It could also surface unresolved questions for students and faculty, while making lessons learned in one laboratory available to others.

To build that future, universities need workflows that capture what was proposed, what was changed, what happened, what failed, and how experts read the result. Shared systems should connect those workflows so AI agents can use tools and information across laboratories under appropriate permissions. This will require large public and university-funded support today for computing capacity, protected data systems, and people skilled at customizing and post-training AI models, meaning adapting them after initial training.

What comes next for researchers and students

The process of tracking this work could also preserve the lineage of ideas. It could make contributions easier to recognize, including those of graduate students. With agreed rules for consent and credit, researchers could collaborate more openly, with more confidence that their work will be acknowledged. That, in turn, could help solve the earlier problem of how to recognize and reward intellectual contributions in an AI-assisted research environment.

Rakhlin's broader point is that AI systems can already gather and reason across more information than any individual researcher could absorb, and their ability to make useful connections should continue to improve. Universities should use that capacity to bring people together around hard scientific and engineering problems, letting researchers build on one another's strengths and insights. The direction he outlines is not just about using AI tools. It is about deciding what universities want to protect, what they want to measure, and what they need to build now.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.