On October 8, 2026, MIT News published an essay by Sasha Rakhlin, Director of the MIT Statistics and Data Science Center and Distinguished Professor in Data, Systems, and Society, that synthesizes recent discussions about how AI is changing academia. According to the essay, AI systems have reached milestones that would have seemed improbable a few years ago, including gold-medal-level performance at the International Mathematical Olympiad and a recently proposed solution to a Millennium Prize problem. Rakhlin's central argument is that universities now need to rethink how they organize research, train students, and assign credit, because the routine academic work that once served as training ground is increasingly something machines can do. The essay also points to broader institutional attention on this question, including an MIT report on AI and education published this summer.
For students, the essay raises an uncomfortable question: if AI can draft, solve, and summarize, what exactly are you supposed to be learning? The practical answer proposed here is not to avoid AI, but to change how you think about which parts of your coursework you delegate to it.
The essay's key distinction is between unnecessary friction and the work through which expertise develops. Unnecessary friction is effort that adds no learning: formatting citations by hand when a tool can do it accurately, retyping data that already exists in a machine-readable file, hunting through library stacks for a paper available online. Formative practice is effort that builds capability: wrestling with a problem statement until it is precise, checking whether a result actually holds, reconstructing an argument to see if it survives scrutiny. Both feel like work. Only one of them makes you better.
The difficulty is that AI collapses this distinction from the student's side. When a tool can do anything you ask, every task looks equally delegable. The friction you skip in one case is genuinely unnecessary; the friction you skip in another was the entire point of the assignment. You often cannot tell which is which until the skill gap shows up later—in an exam, a job interview, or a research project where the AI's output is confidently wrong and nobody in the room can tell.
So the method suggested here is to decide deliberately, before starting an assignment, which steps you will keep for yourself. The essay names four capabilities students should learn: formulating problems, auditing model outputs, reproducing results, and defending their choices. Treating these as a checklist gives you a concrete way to protect your own development.
Formulating the problem is where most of the thinking happens, and it is the step most tempting to delegate. Asking an AI to "write something about climate policy" produces generic output because the problem was generic. The discipline of narrowing a vague topic into a specific, answerable question—what population, what time frame, what would count as evidence—is a skill that transfers everywhere, and it is also the step that determines whether everything downstream is worth doing. Keep this one. Write your own problem statement before you consult any tool, then use the tool to stress-test it.
Auditing AI-generated outputs means treating the tool's answer as a draft claim, not a fact. AI systems can produce fluent, well-structured work containing fabricated citations, subtle statistical errors, or confident misreadings of a source. The professional skill is not avoiding such tools; it is verifying them. Pick one claim in the output and trace it to its origin. Check whether the cited paper actually says what the tool says it says. This habit—skeptical verification of machine output—is rapidly becoming a baseline expectation in knowledge work, and coursework is the cheapest place to build it.
Reproducing results is the traditional core of research training, and it remains so. If a calculation, a code script, or an experimental analysis produces a key result, run it yourself rather than accepting the summary. Re-deriving a result by hand, even once, teaches you where the assumptions live and where the result is fragile—knowledge that never appears in the polished output. When AI produces a proof, a dataset cleaning pipeline, or a model, the question "can I get this same answer from the raw inputs?" is the difference between understanding and delegation.
Defending your choices means being able to explain, in your own words, why you took one approach rather than another: why this method, this sample, this threshold. If you cannot articulate the reasoning behind your work, you have submitted an answer without owning it. A practical test before submitting any AI-assisted assignment: could you answer follow-up questions about your choices without the tool open? If not, you have delegated the part that gets examined—in class and in careers.
None of this means doing everything manually. The sensible division of labor is to delegate the mechanical and keep the judgment: let tools handle formatting, searching, and first-draft drudgery, while you retain the problem statement, the verification, the reproduction, and the justification. A useful rule of thumb: if a step would teach you something you would need in an exam or a professional setting where you must answer for the result, keep it; if it would only cost you time, delegate it.
Two cautions. First, follow your instructor's AI policy—courses differ on what is permitted, and the method above operates within whatever rules apply to you, not around them. Second, this is practical learning advice inspired by the essay's arguments, not an official MIT method and not a proven intervention; the essay's claims about AI milestones are attributed to the essay and are not independently verified here.
The essay's larger point is that institutions are rethinking what coursework should certify as AI capabilities advance. Students can participate in that rethinking from their side. The durable skills are the ones that survive the tooling: knowing what question is worth asking, knowing when an answer is wrong, knowing how to check it, and being able to stand behind the result.
For your next AI-assisted assignment, try this: choose one routine step the tool would usually do, and do it yourself first—re-derive a key result, re-run one calculation, or check one output against the original source. Then write one sentence on what you learned that the tool did not tell you.
中文摘要
2026年10月8日,MIT News刊登了统计与数据科学中心主任Sasha Rakhlin的文章,探讨AI如何改变学术研究、教学与学分认定。文章提出,学生应学会提出问题、审查AI输出、复现结果并为自己的选择辩护。本文据此给出实用建议:区分“不必要的摩擦”与“形成能力的练习”,在AI辅助作业中主动保留关键步骤——自己先推导关键结果、重跑一次计算或核对一处输出,再写下一句你从中学到而工具没有告诉你的东西。请遵守所在课程的AI使用政策。