MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training issued a report that states that” There are early signals, however, that overreliance on chatbots can have a range of significant negative consequences – diminishing critical thinking, weakening memory, eroding confidence, and undermining mastery. Getting the right answer from a chatbot can create the illusion of learning – but it can also trigger “cognitive surrender”, where students fall back on AI at the first hint of struggle.”
Here is my modest proposal to prevent the “cognitive surrender” mentions in the report.
Proposal: Keep the Human in the Loop on Purpose
Problem. Generative AI can already finish most written schoolwork and a lot of office work. The rational short-term move is to hand it the hard part. The long-term cost is “cognitive surrender”: weaker memory, thinner judgment, and confidence that belongs to the model, not the person.
Goal. Not ban AI. Not pretend everyone will self-discipline. Redesign work so the human still has to *start, check, and own* the thinking.
Rule. Augment. Do not automate the part that makes you better.
How.
1. Split every task into two layers.
Layer A is allowed to be AI: first draft, search, format, summarize sources you already found.
Layer B is not: the question, the outline, the claim, the error hunt, the “what would change my mind” paragraph. Grade or pay for Layer B.
2. Require a visible struggle log.
Before the model is opened: write the problem in your own words and your first attempt. After: list what you accepted, what you rejected, and why. No log, no credit. This is cheap and it forces the brain to stay in the room.
3. Test with the tool off.
Closed-book, no-laptop slices of the same skill. If you cannot reconstruct the argument without the chatbot, you did not learn it. Institutions should treat that as the real exam; take-home work is practice.
4. Make the human the scarce resource.
Assignments and jobs should reward catching the model’s confident error, not producing more fluent pages. “Find the hole” is a better use of AI than “write the paper.”
5. Stop pretending detection is a strategy.
Detectors are weak and poison trust. Policy should be: permitted uses named in advance, prohibited uses named in advance, and assessment designed so cheating is pointless.
6. Train the skill, not the slogan.
“AI literacy” that only means prompting is how atrophy happens. Teach: when to distrust fluency, how to verify a citation, how to keep a working memory of the argument, how to notice you have stopped thinking.
**What success looks like.** People still use the model. They just cannot hide behind it. Output may be faster. The person should still be able to explain, defend, and redo the work without it.
**What failure looks like.** Faster essays, faster amendments, emptier heads — then surprise that nobody can tell when the machine is wrong.
The conundrum is not AI. It is work designed so the easiest path is not thinking. Change the work.