A few years ago, the school where I teach made a thoughtful decision: rather than treat artificial intelligence only as a threat, it encouraged interested educators to invest time and imagination in exploring what the technology might make possible for faculty and students. Like many schools, we started small, building custom chatbots for specific needs. Over time, those experiments coalesced into two school-supported tools: one designed to offer writing feedback using the Self-Regulated Strategy Development model already used in our middle school, and another designed to support student research.
The work has been promising. It has also made many of us more uneasy.
Learning Isn’t Always Efficient
As an English teacher, I began wondering what might disappear each time AI made a process faster or more polished. Some of the most important moments in my classroom are not efficient: a student sitting with a blank page during a write-to-think; a peer reading a classmate’s draft closely enough to offer an original observation; a room falling silent during a discussion because no one yet knows what to say.
That silence is not evidence that learning has stopped. Often, it is where learning begins.
I noticed the concern most sharply among some of my strongest students. They could produce polished work, but when asked to enter a collaborative, classwide discussion without a prepared answer, some seemed increasingly uncomfortable with the uncertainty required to form an idea in public. I cannot prove AI caused that discomfort. But it clarified the question I could not stop asking: Are we using AI to support thinking, or training students to experience not knowing as a problem that should be eliminated immediately?
That silence is not evidence that learning has stopped. Often, it is where learning begins.
Guardrails for Educational AI
Educational AI should be evaluated not only by what it can do, but by what it is designed not to do for the student. That conviction led me to begin developing an experimental “Private Harkness” coach, named after the student-led discussion method that anchors much of my teaching. Its design premise is simple: Instead of supplying an answer, it should press the learner to make a claim, identify evidence, confront a counterargument, and decide what they think. I was not trying to build an AI that knew more. I was trying to build one that knew when not to answer.
Research increasingly suggests that this distinction matters. In a randomized field experiment involving nearly 1,000 high school mathematics students, researchers compared unrestricted access to a GPT-4-based assistant with a tutor designed to provide hints without giving away full solutions. Both tools improved performance while students could use them. But when AI was removed, students who had used the unrestricted version scored 17 percent lower than students who had never used it. The tutor with guardrails largely eliminated that penalty, although it did not create a positive exam gain. The study was limited to mathematics in one high school, but its central lesson is worth carrying across disciplines. Assisted performance and learning are not the same thing, and design determines whether a tool supports one while undermining the other.
Guardrails should not solely be pedagogical. Just as important, students need ethical frameworks for deciding when and how AI should enter their work. UNESCO’s AI competency framework for students places a human-centered mindset and ethics alongside technical knowledge and system design.
Assisted performance and learning are not the same thing, and design determines whether a tool supports one while undermining the other.
Advice for AI Developers, Schools, And Teachers
For educational AI developers, schools, and teachers, I propose four rules:
- Do not supply the student’s claim. Ask the student to state an initial position, even if it is incomplete. AI can challenge, clarify, or test that claim, but it should not originate the idea the student is meant to develop.
- Do not choose the evidence. AI may help a learner evaluate whether evidence is relevant or sufficient. It should not remove the work of selecting, interpreting, and defending that evidence.
- Do not resolve productive ambiguity. In literature, history, ethics, and many real-world problems, uncertainty is not a defect. A good tool can surface tensions and alternative interpretations without collapsing them into a single fluent conclusion.
- Do not replace final judgment. AI can expose consequences, raise objections, and identify what a student has overlooked. The final decision, and the responsibility for it, must return to the learner.
These rules are not an argument for making students struggle needlessly. AI can remove barriers that have little educational value, improve access, offer timely feedback, and help teachers manage work that otherwise limits their attention to students. Productive struggle also requires support; withholding an answer should never mean abandoning a learner.
Creating Space For Deep Thinking
The practical question is whether the tool removes an obstacle to learning or removes the intellectual work that constitutes learning. Before adopting an AI tool or redesigning an assignment around one, educators should ask, ‘What thinking is the student supposed to practice here? Does the system require an attempt before offering help? Does it provide the least assistance necessary? Could the student explain and defend the final decision without the tool?’
The most effective educational AI may occasionally feel less impressive because it does not immediately comply. It leaves room for the unfinished sentence, the uncertain pause, the handwritten first thought, the peer’s question, and the student’s own judgment. In education, knowing when to remain silent may be one of the most intelligent things AI can do.
In education, knowing when to remain silent may be one of the most intelligent things AI can do.
Disclosure: The author developed the experimental AI coaching prototype referenced in this essay. The prototype is discussed solely as the origin of the proposed design principles and is not being promoted or linked here. The views expressed are the author’s own and do not represent her employer.
