One Year of 'Study Mode': The AI Industry Admitted the Problem. It Hasn't Solved It.
In a single stretch of mid-2025, the three biggest AI labs did something remarkable: they all conceded the same point. OpenAI shipped Study Mode in ChatGPT at the end of July. Google followed within days with Guided Learning in Gemini. Anthropic added Learning Modes to Claude the same month. Three companies locked in the most expensive product race in tech history paused to ship, essentially, the same feature — a version of their chatbot that withholds the answer and asks questions instead.
Read that as what it is. The companies that know these models best looked at how students were using them and concluded that the default product — the answer engine — was hurting learning badly enough to warrant a separate mode that behaves like the opposite of the product. Nobody ships a mode that makes their product slower and more effortful unless the fast version is causing damage they can no longer ignore.
We've been making that argument since before it was fashionable. A year later, it's worth asking the question that matters for anyone writing a fall syllabus: did the toggle fix it?
The research got darker, not better
While study modes were rolling out, the evidence on unstructured AI use kept arriving — and it points one direction.
An MIT study put EEG caps on students writing essays with ChatGPT, with a search engine, or with nothing. The ChatGPT group showed reduced neural connectivity in networks tied to memory and creativity — and minutes after finishing, many couldn't recall what they had just written. The researchers coined a term for the long-run version of this: cognitive debt. Work by researchers at UT Austin, Georgia Tech, and Hugging Face found the same trade at the behavioral level: AI assistance boosts short-term performance while weakening long-term retention, critical thinking, and problem-solving. A Microsoft and Carnegie Mellon survey of 319 knowledge workers added the mechanism — the more people trusted the AI, the less critical-thinking effort they reported spending. And the reliance skews young: participants aged 17 to 25 showed the highest AI dependence and the lowest critical-thinking scores of any group studied.
None of this says AI can't teach. It says unstructured, answer-first AI reliably substitutes for thinking rather than provoking it. Which is exactly the diagnosis the labs made themselves when they built the modes.
Why the toggle didn't fix it
So the industry named the disease and shipped a treatment. The trouble is structural, and it comes down to three words: the mode is optional.
Study mode sits one tap away from the answer engine, inside the same product, behind the same login. The student who opens it at 11pm with a problem set due at midnight is being asked to choose the slow, effortful version over the instant one — voluntarily, alone, every single time. That's not a learning environment. That's a diet app installed on the fridge.
The instructor, meanwhile, is nowhere in the loop. A general-purpose study mode doesn't know what course the student is in, what's been taught so far, what notation the professor uses, or what this week's assignment is actually testing. It can't tell the difference between productive struggle and a student quietly drowning. And it reports nothing back — no signal about where the class is stuck, no visibility into how help was used, no way to set different rules for homework versus exam week. The labs built a tutor persona. They did not build any of the connective tissue that makes tutoring part of a course.
That's not a criticism of the people who built these modes. It's a recognition of what they're for. A consumer product with a billion users cannot be structured around one chemistry professor's rules. Something built for the classroom can be.
The mode was never the product. The architecture is.
Here's the distinction we'd offer anyone evaluating AI for their department this fall: ask not whether the tool has a learning mode, but whether learning is the only mode it has.
When refusal to hand over answers is a setting, it's a suggestion. When it's the architecture, it's a guarantee. A purpose-built teaching tool should be grounded in the actual course — the professor's materials, notation, and pacing — so its help is scaffolding for this class, not generic tutoring for the median student on the internet. Its behavior should be set by the educator, per assignment: more guidance early in the term, less as competence builds, none at all during assessments. And it should close the loop — surfacing where students are stuck, in aggregate, before the midterm does it for you.
Under that architecture, the 11pm problem doesn't require a heroic act of student willpower, because the legitimate help is the only thing on offer — and it's better help, because it knows the course. The cognitive-debt studies all describe what happens when AI does the thinking. A tool that structurally cannot do the thinking, and instead makes the student do it with support, is on the other side of that ledger.
One year in
The study-mode launches were the AI industry's concession that we were all right to worry. The year of research since is the receipt. What's left is the part a toggle can't do: making the learning-first behavior mandatory, course-aware, and visible to the person responsible for the learning.
That's what we're building — AI tutoring where the refusal to just give answers isn't a mode you hope students pick, but the foundation everything else sits on. If your department is planning for spring pilots, we'd love to talk.