The Same Chatbot, Two Very Different Students

Give two students the exact same AI tutor and the exact same twenty minutes, and you might assume they will come away with roughly the same understanding. A recent experiment says that assumption is wrong. What each student was trying to achieve, not the tool in front of them, turned out to be the thing that shaped how much they actually learned. The chatbot was a constant. The mindset was the variable that mattered.

The study, conducted by Laura Schmidt, Niklas Obergassel, and Julian Roelle at universities in Germany, recruited 104 university students for a supervised online session [1]. Everyone used ChatGPT to study the same four ideas from social psychology, including the well-known mere-exposure effect, where simply encountering something repeatedly can make us like it more. The only difference was the frame the researchers placed around the task.

Two ways to point a mind

Half the students were nudged into what psychologists call a mastery orientation. Their instruction was to deepen their own knowledge, to understand the material for its own sake. The other half were pushed toward a performance orientation, told in effect to outshine their peers. To keep the framing from fading, the researchers repeated each group's goal every five minutes throughout the twenty-minute session [1]. Then everyone took a test measuring what they had absorbed, and the transcripts of their conversations with the model were analyzed for the kinds of prompts they used.

The mastery group came out ahead on the thing that counts. They acquired more conceptual knowledge, producing better definitions and clearer explanations of the ideas they had studied [1]. The students chasing a competitive edge did not learn the concepts as solidly. Instead, their final answers were more likely to be padded with non-essential trivia, the sort of impressive-sounding detail that signals effort without demonstrating genuine understanding. They also reported feeling more pressure during the session.

That contrast is the heart of it. Same model, same content, same clock. The goal a person carried into the interaction changed the character of the exchange and, downstream, what stuck.

Why the goal leaks into the prompts

There is a neat mechanism hiding in this result. When you learn with a chatbot, you are not a passive recipient. You steer the whole thing through your prompts, and your prompts carry your intentions with them. A learner who wants to understand tends to ask the model to explain, to compare, to clarify what is confusing. A learner who wants to win tends to fish for facts they can display, the quotable morsels that make an answer look sophisticated. The tool obligingly hands back whatever you reach for. So the difference in learning was not something the AI imposed. It was something the students' goals produced through the questions they chose to type.

This is a useful corrective to the way we usually argue about AI in education, as if the technology were a single lever that either helps or harms. The same system can behave like a patient tutor or a trivia dispenser depending on who is at the keyboard and what they are after. That places a lot of weight on the human's stance going in, which is not something a better model will fix.

It also connects to a broader pattern in how our individual differences shape what we get from these tools. Who even reaches for a chatbot in the first place, and why, varies with personality and self-belief, as work on which personalities actually gravitate toward ChatGPT explores. The goal-framing result adds a layer: it is not only who uses AI, but with what intention, that decides the outcome.

Reading it with the right caution

The findings deserve a few asterisks. This was a modest sample of about a hundred students, and twenty minutes is a short window in which to build deep expertise, so the study captures early learning rather than the slow work of mastery over a term. Participants also arrived with different levels of AI experience, which adds noise. And the test measured factual and conceptual knowledge rather than the harder-to-pin-down skill of applying ideas in novel situations, so we should not overclaim about "deep understanding" from this alone.

Still, the signal is clear enough to act on. The value you extract from an AI tutor is not fixed by the tool; it is shaped by the goal you bring to it. Approach the conversation wanting to understand, and the questions you ask will pull the model toward genuine explanation. Approach it wanting to look impressive or to beat someone, and you may collect a lot of shiny facts while the actual learning slips through your fingers. For more on how mindset and machine interact, browse our wider artificial-intelligence coverage.

Sources

1. Schmidt, L., Obergassel, N., & Roelle, J. (2025). AIming High: Do Goal Structures Matter in Learning With ChatGPT? Applied Cognitive Psychology. https://onlinelibrary.wiley.com/doi/10.1002/acp.70148
2. Schmidt, L., Obergassel, N., & Roelle, J. (2025). AIming High: Do Goal Structures Matter in Learning With ChatGPT? [Record and PDF]. ResearchGate. https://www.researchgate.net/publication/397784629_AIming_High_Do_Goal_Structures_Matter_in_Learning_With_ChatGPT

This article summarizes published research for general informational purposes only and does not constitute professional advice.