A student opens a chatbot, pastes in the essay prompt, and has a finished five-paragraph response in under two minutes. The paper reads fine, but nothing about producing it taught the student anything about argument, evidence, or revision. The tool did the work. The student just supplied the prompt.

That is the dividing line that matters in classroom AI, and it has nothing to do with which product a school adopts. The line runs between AI that produces the output and AI that builds the capacity to produce it. A model can generate a paragraph, a proof, or a lab report faster than any student could write one. None of that generation, on its own, moves a student's ability forward. What moves a student's ability forward is friction: making a claim, defending it, getting evaluated against a standard, and revising with that standard in view. Remove the friction and the tool has replaced the learning event, not supported it.

Where the Optimism Gets Ahead of the Evidence

Sal Khan makes the strongest public case for AI tutoring in Brave New Words (2024), and the case leans on Benjamin Bloom's 1984 finding that one-on-one tutoring can move students roughly two standard deviations above ordinary classroom instruction, the "2 sigma problem." Khan's argument is that AI can deliver tutoring-level attention at a scale no school could staff. The argument holds only if the AI tutoring actually replicates what made Bloom's tutors effective: withheld answers, structured questioning, and required student reasoning before help arrives. A chatbot that answers on request replicates none of that. It replicates a vending machine with a friendlier interface.

Ethan and Lilach Mollick's 2023 research brief, Using AI to Implement Effective Teaching Strategies in Classrooms, makes the same distinction from the opposite direction. The same underlying model produces very different instructional value depending on the prompt structure wrapped around it. A model instructed to give answers gives answers. A model instructed to ask a question, wait for a response, and withhold the answer until the student has attempted one, functions closer to a tutor. The technology is identical in both cases, but the instructional design is not.

What This Means for Adoption Decisions

Neil Selwyn's long-standing critique of education technology, argued fully in Should Robots Replace Teachers? (2019), is that schools tend to adopt tools that automate the visible task, grading, drafting, generating, without examining whether that task was ever the actual point. Grading a paper fast is not a goal that moves the student forward—it just saves a teacher's time. Instead, building a student's capacity to produce work that would earn a strong grade unassisted should be the goal. A tool that automates the former while leaving the latter untouched has optimized the wrong variable, and the adoption decision needs to name that difference before signing a contract, not after.

Guided Scholar's three modes, Teach Me, Coach Me, and Evaluate (ACT prep), are built to keep that friction in place rather than remove it. Teach Me uses scaffolding to assist the student in developing an answer to the teacher's prompt. While feedback is available at any point, it does not provide any answers, just feedback so the student can continue to hone his/her work with the next revision. Coach Me responds to a draft with feedback, not rewrites. Evaluate applies the ACT rubric transparently with feedback and suggestions, before the actual test, so the student sees the standard at work instead of just the score. None of the three modes generates finished work on a student's behalf. That is a design constraint, not a limitation waiting to be engineered away.

The right question to ask about any classroom AI tool is not whether it is accurate, fast, or well-reviewed. It is whether a student who used it correctly for a full semester would be a better writer or would simply have produced more finished essays. Those are not the same outcome, and most of what is currently on the market is optimized for the second one.

Further Reading