The AI Tutor That Refuses to Answer Questions Got Better Results

The AI tutor that worked best was the one that refused to give students answers.

Researchers at the University of Wisconsin, La Crosse built an AI tool called Macro Buddy specifically for their undergraduate macroeconomics class. It had internet access turned off so it couldn't look things up. And whenever a student asked it a question, it didn't answer. Instead, it asked one back.

Students who used Macro Buddy alongside group discussion with classmates scored higher on in-person, closed-book exams than students who studied alone with no AI at all.

That sounds like a small finding. It isn't.

The problem everyone's arguing about

Roughly 90% of US college students now use generative AI for their coursework, according to a 2025 survey of 1,100 students across two- and four-year institutions. The debate in education has mostly been framed as: is this cheating, and how do we stop it?

The Brookings Institution spent a year gathering a different kind of evidence. Drawing on more than 400 studies and hundreds of interviews with teachers, students, and experts, they published a report in January 2026 that reframed the problem entirely.

Cheating isn't the real issue. Learning is.

"Students can't reason. They can't think. They can't solve problems," one teacher told Brookings researchers. The report described AI as the "fast food of education" — convenient and immediately satisfying, cognitively hollow over time. The term they used for what's happening to students' thinking skills: "great unwiring."

The OECD reached a similar conclusion in its Digital Education Outlook 2026, released the same month. "Offloading cognitive tasks to general-purpose chatbots creates risks of metacognitive laziness and disengagement that may deter skill acquisition in the long run," the report states. More damning: students who use general-purpose AI tools tend to produce better assignments — but when AI is removed at exam time, that advantage disappears. Sometimes it reverses.

Better outputs. Less actual learning. The AI was doing the thinking for them.

What Macro Buddy did differently

The key design decision behind Macro Buddy was simple and uncomfortable for any student hoping for a quick answer: it wouldn't give one.

If a student asked why falling prices might increase consumer spending, Macro Buddy wouldn't explain. It would ask: what happens to your purchasing power when the price of something drops? The student would have to work it out.

This approach — Socratic tutoring — is well-established in learning science. The struggle to retrieve and apply knowledge, rather than receiving it passively, is where durable memory forms. Macro Buddy was essentially a structured version of that struggle, available at any hour, trained on the specific course materials, unable to go beyond them.

The study enrolled 140 first- and second-year undergraduates across four sections. After an initial shared exam, researchers split the sections into four study conditions: alone without AI, in groups without AI, alone with Macro Buddy, and in groups with Macro Buddy. All exams were in-person, closed-book, no AI.

The group that combined Macro Buddy with peer discussion came out ahead. Students in that condition had to articulate their reasoning to the AI and then explain it again to classmates — two rounds of effortful retrieval. By the third exam, the advantage widened.

Why the combination mattered

The peer discussion piece isn't incidental. Teaching something to another person is one of the most reliable ways to solidify understanding. When you explain a concept, gaps in your own knowledge become visible. Macro Buddy surfaced those gaps; the peer conversation closed them.

Students who only used Macro Buddy without group work showed more modest gains. The AI alone wasn't a miracle — it was a better study partner than staring at notes solo, but it worked best as a complement to human interaction, not a replacement for it.

This matters globally. AI tutoring tools are being deployed at scale, often positioned as equalizers — giving students in low-resource schools access to something like one-on-one instruction. The AI in education topic is increasingly central to how countries think about closing learning gaps. But the Macro Buddy research suggests the design of those tools matters as much as their availability.

An AI that hands a struggling student the answer is training them to expect answers. An AI that asks a struggling student another question is training them to think.

The design gap nobody's closing fast enough

Most AI tools available to students right now are built for output quality, not learning. ChatGPT, Claude, and their equivalents are optimised to produce useful responses — that's precisely why they're so popular for homework, and precisely why they're so dangerous for cognitive development when used without structure.

The OECD's 2026 report identifies what purposeful educational AI looks like: it serves as a peer-like contributor in collaborative tasks, generates personalised materials tied to specific course content, provides feedback tied to learning goals rather than output polish. Macro Buddy fits that description. Most tools don't.

Schools and universities are largely responding to AI by banning it, monitoring it, or — less commonly — trying to redesign how learning is assessed so that AI shortcuts stop working. The deeper redesign problem, which the Brookings report flags but education systems have barely started to address, is: what should AI assistance look like during the learning process, not just at submission time?

The learning science research at Albis points toward an answer that's counterintuitive: the AI that makes learning harder in the moment makes it stick. Friction is the feature, not the bug.

What this means for the AI tutoring boom

The e-learning market is projected to grow from $276 billion in 2026 to $462 billion by 2031. A significant portion of that growth is AI tutoring. Investors, edtech companies, and governments are all pouring resources into AI tools that promise personalised learning at scale.

The Macro Buddy study is a reminder that the question isn't whether AI can tutor — it clearly can. The question is whether the people building these tools understand enough about how learning actually works to build them well.

An AI that gives students answers on demand will be popular. It will feel helpful. Usage metrics will look great.

An AI that asks students better questions, designed around specific course content, integrated with peer interaction — that's harder to build, slower to use, and less immediately satisfying. It's also the one that made exam scores go up.

Education has always been about the productive struggle. The best AI tutors, it turns out, have figured out how to preserve it.

---

Albis tracks how knowledge and information travel across regions and disciplines. Explore the AI topic hub for more on artificial intelligence and learning.