AivexaNewsSearch
AI news for builders and product teamsChecked every hour

3 Questions: What is the best path forward for AI in academia?

Collected Oct 13, 2026

MIT Statistics and Data Science Center Director Sasha Rakhlin, the Distinguished Professor in Data, Systems, and Society, has published a new essay, "AI and Academia," synthesizing recent discussions and readings on how artificial intelligence is reshaping academia, with particular focus on mathematics, statistics, machine learning, and engineering. The essay, framed as three questions and answers, concentrates mainly on graduate research and education. It follows an MIT report on AI and education published this summer.

Rakhlin opens with mathematics as his illustration of the pace of change. A model reached gold-medal level at the International Mathematical Olympiad last year. Only a year later, he writes, models are producing new research results, including a recently proposed solution to one of the Millennium Prize Problems. Tasks that once demonstrated advanced mathematical ability, he notes, can increasingly be performed by AI.

The mechanism he identifies behind that speed is verification. In disciplines where results can be checked quickly and reliably — formalized proofs verified automatically, programs run and tested — systems can generate candidate solutions, learn from outcomes, and improve. That same loop applies to AI research itself: better models, training procedures, and supporting tools feed the next round of development, producing a compounding process. This is my own gloss on a familiar pattern in machine learning: the tighter the feedback signal, the faster the iteration, which is why math and code have advanced faster than fields where judging a result takes months of experiment. Rakhlin adds that the shift creates opportunities for experienced researchers to pursue questions whose technical demands previously placed them out of reach, while sharpening the distinction between getting a solution and understanding why it works, what generalizes, and what to ask next. He cautions against assuming abstraction, judgment, or problem formulation will remain beyond AI, and argues universities should prepare for a near future in which AI is smarter than all of us in many, perhaps most, aspects of intellectual work.

The second question concerns credit and graduate training. Rakhlin argues that in a growing number of fields a polished paper is becoming a weaker signal of individual expertise, and that departments should reconsider what they reward — without simply defining valuable work as whatever AI cannot yet do. He suggests greater recognition for asking good questions, replication, synthesis of ideas, informative negative results, and shared datasets. Evaluation, in his view, should establish what a researcher contributed and takes intellectual responsibility for, including when substantial parts of the work were performed by AI, and those expectations should guide hiring, promotion, and funding and be made explicit for current and incoming PhD students.

Training is harder, he writes, because intellectual muscles develop through exercise. Routine calculations, coding, failed approaches, and small discoveries have long helped students build intuition and judgment, and delegating that work can remove formative experiences even as AI lets students attempt more ambitious projects. The task is to distinguish unnecessary friction from the work through which expertise develops. Students should learn to formulate problems, audit model outputs, reproduce results, and defend their choices. Fundamentals, he suggests, may become more valuable as the basis for using these tools well.

The final question addresses what universities should build. Rakhlin states the goal plainly: universities should pursue questions over long horizons, share results openly, and evaluate claims independently. Industry partnerships will be essential, he writes, but commercial priorities should not be assumed to cover the breadth of science or stay aligned with it over time, so some degree of technological independence is necessary. A likely trade-off in practice is that independence costs money and talent at a moment when both are scarce in academia relative to industry labs.

His central strategic claim is that academia's most important asset may be the accumulated knowledge and experience inside its laboratories. Published papers are a selective record, he notes: failed experiments, abandoned directions, and the reasons an approach did not work often go unpublished. Scientists and engineers hold tacit knowledge about which interventions are likely to fail and why. That missing context, he suggests, may help explain why models currently struggle in some domains — especially the empirical sciences — to anticipate consequences that are obvious to an expert. He adds that these limitations may be temporary, and that capturing negative results and scientists' interpretations could make models much better at scientific exploration.

Rakhlin sketches a future in which MIT's laboratories function more like a single, living scientific organism connected by 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 make lessons from one lab available to others. Enabling that 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. It would require substantial public and institutional investment now in compute, secure data systems, and expertise in adapting and post-training AI models.

Tracing that process, he writes, 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 confidence that their work would be acknowledged, which also bears on the earlier question of how to reward intellectual contributions. He closes by noting that AI systems can already synthesize and reason about information from more sources than any individual researcher could absorb, and that this ability will keep improving. In his view, that capacity should be used to bring researchers together around hard scientific and engineering problems. He calls the path promising but says universities need to invest now to make it possible.

Why it matters: The essay is aimed at departments and university administrators, but it has direct implications for developers and product teams working with academic labs. If universities adopt shared infrastructure for capturing hypotheses, experiments, failures, and interpretations, that data becomes a potential training and evaluation resource for scientific AI. Graduate students and researchers would face revised expectations around credit, AI disclosure, and demonstrated contribution. The prediction of continued rapid progress in verification-rich fields like mathematics is my inference from the stated trend, but it suggests tooling for formal proof checking, benchmark design, and research provenance is worth watching closely.

Read at MIT News · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

MIT Statistics and Data Science Center Director Alexander (Sasha) Rakhlin shares important considerations for departments and institutions.