assessment
MIT Is Asking the Same Questions We Are, and That's Good News
MIT's new report on AI in teaching names the problem we have been writing about: when AI can do the homework, how do we know students learned? Having colleagues with that depth of expertise working on it makes us optimistic.
Good company
When AI can do most intro programming homework, how do we know what students can do on their own? That question is the reason CodeReps exists, and it turns out we are in very good company.
On August 13, 2026, MIT released the Report of the Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, co-chaired by Eric Klopfer and Sam Madden. The committee spent the spring of 2026 studying how generative AI is changing teaching and learning at MIT, and many of its conclusions read like the conversations we have been having with instructors all year.
We find that genuinely encouraging. When people with MIT's depth in learning science, computing and teaching take a problem this seriously, good answers tend to follow.
What MIT found
Three points from the report stood out to us.
- Take-home work is losing its value as evidence. The report says AI "saps" problem sets, long-form writing and take-home exams "of their value for both teaching and assessment." It does not tell instructors to drop homework; it asks them to redesign it so it still builds durable understanding.
- Cognitive offloading is real. The committee warns that leaning on chatbots can reduce critical thinking and weaken memory and mastery, and it names a pattern every programming instructor will recognize: "cognitive surrender," turning to AI at the first sign of difficulty.
- Assessment has to change. The report points instructors toward forms of assessment that are less vulnerable to AI, such as oral exams, portfolios, and out-of-class work paired with in-class conversation.
None of this is framed as a crackdown. The report is about design, not detection, and that is exactly the framing we think the field needs.
Augmentation, not automation
One of the report's eight guiding principles is "Augmentation not automation." If we had to pick one sentence to put above every programming course right now, it might be that one.
There is a real difference between asking AI to explain why a loop never terminates and asking it to write the loop. Both use AI. Only the first leaves the student having done the thinking. MIT calls the good kind "pro-learner" AI, and we like the term.
This is the line we try to hold at CodeReps: AI makes practice abundant, and the student still has to show, alone with a problem, that the skill is now theirs.
Where our idea fits
The report also adds a caution we take to heart: leaning too heavily on short, high-stakes in-class exams can weaken students' motivation for sustained work. Splitting open practice from supervised assessment only works if the assessment side respects that. We think the answer is frequency and familiarity: many small, low-stakes checks on problems students have already practiced, rather than one high-pressure exam full of unknowns.
MIT has not evaluated this particular design, and we will not pretend otherwise. It is one concrete, testable response to the problem the report describes, and we are measuring how it works in real courses.
Why we're optimistic
It would be easy to read a report like this as bad news. We read it the other way.
Two years ago, much of the discussion about AI in education was about catching students. Now a committee at one of the world's leading technical universities is asking the deeper questions: what kind of effort students need for learning to stick, where AI helps and where it gets in the way, and what evidence instructors should ask for before they call something mastered. Those are the right questions, and they can be answered.
The tools that make the problem hard are also the ones that make good practice cheap. AI can generate and check far more exercises than any instructor could write by hand and give students immediate feedback at any hour. If we design courses so that AI supports the practice and the student does the thinking, students may well end up practicing more than they ever did.
We are glad to be working on this alongside colleagues like the MIT committee, and we are looking forward to what the next few years of evidence show.
If you teach programming and are rethinking your assessments this year, we would love to hear what you are trying.
Reference: MIT, Report of the Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, August 13, 2026 (PDF).
Useful? It will be to a colleague too. Pass it on.