Last year I wrote about Cognitive Time Under Tension: the amount of time a learner spends actively working at the edge of their ability, not just sitting in front of a task. I argued it was the metric we should be watching, and that AI could either extend it or destroy it depending on how we used it.
On August 13, MIT released the report of its Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training (https://aiandeducation.mit.edu/report/). Eighteen faculty, students, and staff spent five months surveying the campus, running listening sessions, and studying peer institutions. The charge was to assess AI use and propose a policy. What they delivered instead was a warning that the policy is the smaller problem.
Interestingly after reading through the report a couple of times, the report never uses the phrase Cognitive Time Under Tension, but it certainly describes it from another side of the problem.
What MIT Found
The committee’s core finding is that AI has broken the assumption behind most out-of-class assessment. Problem sets, take-home exams, essays, and coding assignments were designed on the premise that completing them required the student to think. That premise no longer holds. The report states plainly that these tools can now produce credible solutions to almost any written assignment in the undergraduate curriculum.
The second finding is the one most people are missing. The damage is not primarily an integrity problem. It is a learning problem, and it is a social one. Attendance at office hours has dropped. Study groups have thinned out. Students who used to argue through a proof with a roommate now get the answer from a chatbot alone in their room. The committee calls this an erosion of the social contract between instructors and students, and it identifies increasing isolation and undermined student confidence as concerning effects already visible on campus.
The survey data backs it up. In MIT’s spring 2026 Quality of Life survey, undergraduates were the only group on campus who reported that AI made them feel more replaceable than capable, 40 percent to 34 percent. In the Tech’s fall 2025 survey, 90 percent of undergraduates said they were concerned about overreliance on these tools, and 67 percent were very concerned. Seventy percent said AI proficiency would matter for their careers. Only 25 percent thought MIT was preparing them for it.
When the committee asked respondents to write in their own concerns, the top category was not cheating. It was the impact on learning and cognitive development.
Where the Report and CTUT Agree
Read side by side, the vocabulary lines up almost one to one.
The report talks about productive struggle, cognitive friction, and hard fun. CTUT is the time spent in that state. The report’s guiding principle of augmentation over automation is the same distinction as coach versus crutch: AI helps when it makes the human better at the work and hurts when it does the work for them. The report warns about cognitive surrender, where a student falls back on AI at the first sign of difficulty. That is CTUT collapsing to zero.
Most importantly, Section 3.1.2 tells instructors to revisit what they want students to know and devise assessments that foster, or even include, the kind of productive struggle that builds durable understanding. That sentence is CTUT written as a design requirement for the entire institution.
The committee also confirms something I have argued for a while: the easy fix makes it worse. Many instructors are responding to AI by shifting weight onto in-class exams. The report calls this out directly. Time-limited assessment reduces how much thought a student can put in, and it signals the opposite of what we want them to value. Quick, high-stakes evaluations are low-CTUT by design.
Where MIT Pushed My Thinking
I would be doing the report a disservice if I only claimed the parts that agree with me. Three places changed how I frame this.
Struggle is social. My original post treated CTUT as an individual learner working against material, with AI as the adaptive partner. MIT’s evidence says the highest-value struggle on their campus happened between people: study groups, office hours, a spirited argument with a peer. A tool optimized for one-on-one tension can quietly replace the collaborative version and nobody notices until the study rooms are empty. CTUT needs a peer dimension.
Incentives beat design. I framed the risk as a prompt-discipline problem. Frame the AI as a sparring partner and the tension holds. The committee’s data says otherwise. Students surrender to AI under deadline and GPA pressure regardless of how the tool is framed. Their proposed fix is structural: change what gets assessed, reconsider what grades signal, and rebuild the agreement about why the work is worth doing. A well-designed Socratic tutor loses to a well-designed answer machine at eleven o’clock the night before the due date.
Measure through assessment, not sensors. I pointed toward eye tracking, heart rate variability, and keystroke analysis as proxies for cognitive engagement. MIT is skeptical of anything that feels like surveillance, from lockdown browsers to unclear logging policies, and says students should not feel policed. Their route to observing struggle is assessment design: oral exams, in-class conversation about out-of-class work, version history, portfolios. That is CTUT made visible without wiring anyone up, and it is the right position for a classroom.
Five Actions for Secondary and Post-Secondary Educators
Every one of these is grounded in a specific recommendation from the report. None of them require a new tool or a budget line.
1. Decide what the assignment is for before deciding what AI is allowed to do
The committee recommends backward design as the starting point. Do not begin by asking whether AI should be permitted. Begin by defining what the student should know, be able to do, and value at the end of the unit. Then set the AI posture per assignment, and write the rationale into the syllabus so students can see it.
In my program that posture takes the form of three gates. Gate 1 is closed: no AI, because the goal is to build the mental model from nothing. Gate 2 is adversarial: AI critiques the student’s work but does not produce it. Gate 3 is open: full tooling, with a written decision log documenting what was delegated and why. MIT’s Appendix B proposes a similar four-option menu. The specific labels matter less than the requirement that every assignment carries one, and that the student understands what learning the restriction is protecting.
2. Replace detection with process evidence
The report is unambiguous: do not rely on AI detectors. They produce false positives against non-native English speakers and neurodivergent students, they invite an arms race with humanizer tools, and they build an atmosphere of suspicion that damages the relationship the whole enterprise depends on. MIT’s own Committee on Discipline will not accept detector output alone as evidence.
The alternative is to make the work show its own history. Require submission through platforms that capture version history. Set staged deadlines so progress is visible over time. Pair out-of-class work with a short in-class conversation where the student explains a decision they made. Have students keep a decision log on open-gate work. A student who submits a polished solution five minutes after the assignment posts has told you everything you need to know, and you did not need a detector to learn it.
3. Put people back in the loop, on purpose
Section 3.1.4 recommends that every course include a regular, structured, in-person social component. Not students sitting in a room quietly taking notes. Group problem-solving guided by an instructor or aide. Weekly check-ins on group projects with deliverables that assess both individual and collaborative contribution. Feedback discussions run against a shared rubric. In-class discussion where individual participation is graded.
The key word is structured. Unstructured group time collapses into one student doing the work and three watching. Explain on day one why the in-person component exists and what it is for. For high school programs this is the easiest recommendation to implement because the students are already in the building. For post-secondary, it may mean reclaiming recitation and lab time that has drifted toward passive review.
4. Scope projects up and move assessment toward judgment
The report makes a point that surprised some readers: because AI is good at coding and some kinds of design, instructors can now assign projects that were unrealistic a few years ago. MIT’s capstone software course now expects near production-quality artifacts in a single term. The learning shifts from writing every line to making and defending design decisions, evaluating what the tool produced, and observing how the system behaves in the real world.
This is the strongest CTUT lever available. A project with a real external stakeholder, a real deadline, and a real consequence for a bad decision keeps a student in the load zone for weeks. The report also recommends elevating portfolios as evidence of mastery alongside or in place of grades, noting that many employers already care more about demonstrated work than transcripts. If you teach a capstone, this is the moment to make it bigger and to make the deliverable public.
5. Disclose your own use and fix the incentives you control
Students notice when instructors use AI for slides, feedback, or grading while restricting student use, and they read it as a double standard. The committee strongly recommends transparency: if AI touched the lecture, the assignment, or the evaluation, say so and say why. Modeling responsible use does more for the social contract than any policy paragraph.
Then look at the incentive structure inside your own course. If the only feedback a student receives on a project is from a machine, the report says they are right to be angry. Use AI graders as student-facing tools for formative feedback, not as the final word. Weight process and progression, not just the end product. Create deadlines that prevent the pile-up at the end that drives cognitive surrender. The committee explicitly rejects rationing top grades as a fix, because it increases the pressure that causes the problem.
The Metric Has Not Changed
The MIT report is diagnostic. It names the disease with precision and prescribes a set of structural treatments. What it does not do is name the thing it wants more of. Cognitive Time Under Tension is that thing: the sustained, effortful, slightly uncomfortable state where learning actually happens. The report describes an institution watching that state drain out of its classrooms and deciding to rebuild around it.
You do not need MIT’s budget to do the same. You need a clear answer to what each assignment is for, evidence of process instead of detection, structured time with other humans, projects big enough to demand judgment, and honesty about your own use. Five things. Start with one.
The full report is available from MIT’s faculty governance site. It is 38 pages and worth every one of them. (https://aiandeducation.mit.edu/report/)
