Here’s something most educators get wrong about AI in the classroom. They focus on whether students should use it. The better question is how students interact with it when they do.
A study by Cheng et al. (2025) looked at one specific behavior: asking questions. They compared how students seek help from GenAI versus a human tutor during a writing task, and what they found has real implications for how we design AI-supported learning.
The short version? Students who asked AI direct, specific questions performed better on their writing. The AI didn’t make them better writers. The way they questioned it did.
Asking AI Questions
One of the clearest findings was that students interacting with GenAI asked significantly more questions than those working with a human tutor. They were also more direct about it. They asked for definitions, procedures, specific features, and concrete steps for accomplishing their writing goals.
With human tutors, the pattern looked different. Students asked fewer questions overall, and the questions they did ask were more indirect and judgment-focused. They hedged. They softened their language. They spent energy managing the social dynamics of the interaction.
As the authors put it: “These findings suggest learners may be less hesitant to admit knowledge deficits and more willing to repair them when interacting with GenAI compared to human tutors” (p. 1).
That’s worth sitting with for a moment. Students felt safer admitting what they didn’t know when they were talking to an AI. The social cost of saying “I don’t understand this” dropped to almost zero.
The Type of Question Matters More Than the Tool
Here’s where it gets interesting. Access to AI alone didn’t predict better writing. The type of questions students asked did.
Students who used pragmatic, information-focused questions performed better. They asked things like: What does this term mean? How do I structure this paragraph? What’s the difference between these two approaches? These are knowledge-deficit questions. The student recognizes a gap, names it, and asks for exactly what they need.
“Participants who sought help from GenAI by making direct requests for specific information tended to perform better” (p. 2).
And the mechanism was clear. “Question asking mediated the relationship between experimental condition and performance, meaning that how learners asked questions explained why performance differed” (Cheng et al., 2025, p. 3).
So the improvement came from the questioning behavior, not from the AI itself. That’s a crucial distinction for educators who worry that AI is doing the thinking for students. In this study, the students who benefited most were the ones actively thinking about what they needed to know and asking precise questions to fill those gaps.
Why Human Tutoring Looked Different
The study also revealed something important about how students interact with human tutors. Those conversations were longer, more back-and-forth, more focused on feedback and clarification. On the surface, that sounds like deeper engagement. But the data tells a more complicated story.
“Participants in the GenAI condition tended to ask one-off questions, while Tutor participants tended to have longer conversational exchanges” (p. 1).
The longer exchanges with tutors weren’t always about learning. A lot of that conversational energy went toward establishing common ground, maintaining social alignment, and managing how they were perceived. Students worried about sounding competent. They worried about politeness. They filtered their questions through a social lens before asking them.
That filtering has a cost. When students spend cognitive energy managing the relationship, they have less energy for the actual learning task. And some questions never get asked at all because the student decides they’re too basic, too embarrassing, or too revealing of what they don’t know.
What This Means for the Classroom
I’ve been arguing for a while now that the real impact of AI in education shows up in how it changes the learning dynamic, not just the learning output. This study adds strong evidence to that argument.
AI creates a low-friction space for help seeking. Students can ask as many questions as they want, as directly as they want, without worrying about what the tutor thinks of them. That psychological safety has measurable effects on performance.
But here’s the part we can’t ignore: the students who benefited most were the ones who knew how to ask good questions. They were specific. They were targeted. They named their knowledge gaps clearly.
That skill needs to be taught.
If we want AI to actually improve learning outcomes, we need to spend time teaching students how to question effectively. That means modeling good prompts in class. It means showing students the difference between a vague request (“help me with my essay”) and a precise one (“how do I transition between my second and third paragraphs without repeating my thesis?”). It means building question-asking into the curriculum as a skill in its own right.
The study also carries a lesson for human tutoring. If students feel safer asking direct questions to AI, that tells us something about the environments we’ve created for human interaction. Tutoring sessions, office hours, and classroom discussions could all benefit from intentionally reducing the social cost of admitting uncertainty. The best tutors already do this instinctively. The research suggests we should do it systematically.
AI didn’t replace the need for good teaching in this study. It highlighted exactly where good teaching matters most: in helping students develop the metacognitive skills to know what they don’t know, name it clearly, and seek the right kind of help.
That’s a skill that will serve them long after they leave the classroom, with or without AI.
Related: Cognitive Surrender: How AI Is Quietly Reshaping the Way We Think
Reference
Cheng, Y., Fan, Y., Li, X., Chen, G., Gašević, D., & Swiecki, Z. (2025). Asking generative artificial intelligence the right questions improves writing performance. Computers and Education: Artificial Intelligence, 8, 100374. https://doi.org/10.1016/j.caeai.2025.100374