Almost No Student Can Talk to AI Effectively

We scored 4.9 million student messages on StudyFetch across 144,544 students over 50 days. Every message was graded on two dimensions: how well the student communicated with AI (prompting), and how responsibly they used it (responsibility).

The responsibility numbers were reassuring. 92.9% of messages scored A or B. Students come to StudyFetch to learn, and they use it that way.

The prompting numbers were not.

0.5% of students with five or more messages consistently wrote prompts good enough to be rated A.

What C-Level Prompting Looks Like

55% of students average a C on prompting. A C means the student has a clear goal but writes something like "help with chapter 3" or "explain this." The AI has to guess what they need. It usually guesses wrong, or guesses too broadly, and the student gets a generic response instead of the specific help they were looking for.

A B just requires being clear enough for the AI to give a useful response. Most students don't reach that bar consistently.

A looks like this: "I'm working on a calculus problem where I need to find the derivative of f(x) = x³sin(x). I tried using the product rule and got 3x²sin(x) + x³cos(x). Can you check if that's right and explain why the product rule applies here?"

Almost nobody writes prompts like that. The students who do get dramatically better results.

Education Level Doesn't Fix This

Education Level Avg Prompting (out of 4) % A-level % A or B

High School 2.38 6.8% 42.1%

College 2.48 8.6% 48.1%

Med School 2.50 9.2% 49.2%

Grad School 2.52 9.6% 50.5%

Grad students write A-level prompts at 1.4x the rate of high schoolers. That's the entire education pipeline, from freshman year through a PhD, producing a 40% improvement. Even at the graduate level, only half of messages score A or B.

Meanwhile, responsibility scores show no education gap at all. Every level is above 92% A/B. Students at every stage know how to use AI ethically. They just can't tell it what they need.

But Students Can Learn

Students who start with the lowest prompting scores do improve with practice on the platform. In a cohort of students who began at D-level and sent 50+ messages, 87% moved up to at least C-level. The progression shows a clear inflection point:

Messages Sent Avg Prompting % Scoring C or Better

1 to 5 1.28 25%

6 to 10 1.27 26%

11 to 15 1.75 60%

31 to 50 1.98 73%

Something changes around messages 11 to 15. Students go from 26% of their messages reaching C-level to 60% in one jump, and the improvement holds through message 100.

Across all students with 20+ messages, 30% improve their prompting score over time, 40% stay the same, and 29% decline. More students improve than decline at every engagement level.

The Good News: Students Aren't Cheating

The responsibility data is the other half of the story, and it's encouraging. 92.9% of messages scored A or B on responsibility. Only 0.08% of messages, 3,967 out of 4.9 million, were flagged in dishonesty-related categories.

StudyFetch is a learning platform, and students come here to study. Our AI tutor doesn't give direct answers, which discourages shortcut behavior from repeating. The high responsibility rate reflects both student intent and product design. We can't claim the same rate would hold on a general-purpose AI tool. But it tells us that when the tool is built for learning, the vast majority of students use it that way.

In a separate study of 10,000 AI-tutored quiz questions, students who engaged responsibly with the tutor were 2.3x more likely to answer correctly on their first attempt and 2x more likely to master the question. The students who engaged the least scored the worst on every metric. 

Why This Matters Beyond School

AI is becoming the interface for knowledge work. The students using StudyFetch today are the workforce of tomorrow, and the data says most of them cannot communicate clearly with AI systems.

This is a skill, not a character trait. Students have the ethical instincts (92.9% responsible). What they lack is the communication technique. The good news is that it can be learned. The question is whether anyone is teaching it.

What We're Doing About It

Students who attach study materials to their chats score +0.15 higher on prompting. Materials give the AI built-in context, which means even a short prompt like "quiz me on this" becomes effective. That's a product-level fix, and it works.

But the deeper fix is teaching students to provide that context themselves. That's why prompt training is moving to the center of our roadmap. We're building features that go beyond scoring messages after the fact. The goal is active coaching: showing students what a better version of their prompt looks like, explaining why specificity matters, and giving them structured practice at communicating with AI before they enter the workforce.

Our AI literacy grading system already shows students real-time feedback on their prompting before they hit send. Every message is a chance to practice. We're expanding that into dedicated prompt training tools, guided exercises, and classroom-level prompting analytics for teachers.

If prompting is the skill gap, and the data says it is, then the platform that teaches it will prepare students for more than just their next exam.

Privacy: No student message text, quiz questions, quiz answers, or student responses are included in any published dataset. All user identifiers have been salted and hashed. The published data contains only anonymized IDs, numerical scores, timestamps, and metadata. StudyFetch does not use student data to train AI models. This analysis was conducted on structured score data only.