By: Dr. Kim Abel // Edited By: Dr. Reed Randall

Twenty-eight juniors in Mrs. Maggie Polanco’s high school English class had a one-sentence shortcut sitting right in front of them.

They were studying Chapter 4 of The Great Gatsby. Any of them could have typed one line into a chatbot: explain the most important relationship in this chapter. Ten seconds later, they would have had a finished answer. Most students would have turned it in word for word.

The class did not take that shortcut due to the way Polanco structured the assignment.

Instead, students worked through a three-step sequence designed by Polanco using an AI tool. First, they identified the most significant relationship in the chapter. Then they explained how that relationship affected the plot. Then came a follow-up step built specifically to push their thinking one level deeper than their first answer.

Each step required them to build on the one before it. A student could not skip to a finished take. The tool would not let the sequence resolve until the student worked through it.

Building that sequence took Polanco planning time up front that a single “explain this chapter” prompt would not have needed. The cost of her time is real, but it is a one-time cost. The sequence works the same way every time she reuses it, this time and in every class period after.

Twenty-seven of twenty-eight students completed the activity. Most correctly landed on the Gatsby-Daisy relationship as central and backed it up with specifics, like Gatsby buying his house across the bay from her. But the tool also surfaced the class’s gaps in real time: some students had latched onto secondary relationships, like Nick and Jordan, instead of the one driving the plot while other students summarized what happened without explaining why it mattered.

Polanco did not have to read all twenty-eight responses individually to catch this. She saw data from the AI tool at a class-level and used it, on the spot, to decide the next lesson needed to model the difference between naming a relationship and explaining how it causes what happens next.

The students did not type in a question and receive a single answer. The tool fed students enough to keep them thinking on their own, and helped their teacher reach a faster reteaching decision than she would have gotten individually.

This is not new pedagogy. Socrates never handed a student a finished answer. He asked the next question until the students did the work themselves. AI lets a teacher run that method with twenty-eight students at once, in real time, instead of questioning one student at a time.

The Claim This Piece Is Making

AI helps students grow when it sends them back into their own thinking. It does the opposite when it hands them a finished answer.

AI is not inherently cheating and it is not inherently safe. What decides the outcome is narrower, and more testable: whether AI hands a student an answer, or opens a pathway to learning.

The “AI is just cheating” critique treats every use of the technology as the ten-second chatbot answer. It is not wrong to notice the shortcut. It is wrong to dismiss the possibility of deeper student engagement, because Polanco’s classroom ran the same technology toward the opposite outcome, and the research on what separates the two is no longer thin.

Most claims about AI in schools cannot be tested. This one can. Researchers checked whether feedback timing changes revision behavior. The test found that timely feedback motivates students to try again.

What Research Shows

Meyer and colleagues ran a randomized controlled trial with 459 tenth graders. One group received AI-generated feedback on their essays. The other group received none.

The feedback group wrote better revisions. They also reported more motivation and more positive emotion about the task itself. This is a controlled study with a no-feedback comparison group, not a survey of people who already liked the tool.

Education Perfect’s data adds the number that makes the mechanism visible:

Among students whose first attempt was rated low quality, 83% tried again when AI feedback was available in the moment. Without it, 7% tried again.

Same grade of students. Same weak first drafts. Wildly different odds of ever revising.

Put those findings together and the mechanism is not subtle. Feedback that shows up while a student can still act on it changes whether they try again. This mechanism is not limited to essay drafts. It explains why Polanco’s three-step sequence worked: the tool never let a student stop at their first answer, and it forced each new response to build on the last one instead of replacing it.

What This Means for How Policy Gets Written

Most AI policies I’ve read are detection policies wearing an education costume. They define what counts as cheating, list the tools to watch for, and leave teachers to enforce a line that moves every six months as the tools improve.

What follows is not a philosophy. It is what a policy looks like when it is built around the timing mechanism instead of getting caught.

At Optima Academy Online, we made a different choice, and it shows up in specific details:

 

  • We do not try to catch AI use. We built the policy around whether a student used it well.
  • Voice is not a vibe a teacher senses. We defined it across three measurable dimensions: whether writing matches how a student talks and writes elsewhere, whether they take and defend a real position instead of hedging, and whether they can explain their own work out loud when asked an unexpected follow-up.
  • Access to AI scales with developmental readiness, not one blanket rule K through 12. Kindergartners watch a teacher think aloud with AI and nothing more. Third graders get bots that ask guiding questions and refuse to give direct answers. Eleventh and twelfth graders build their own agents against assignments with defined learning objectives.
  • Process is gradable evidence. Prompt logs, drafts, verbal defenses, recorded revisions count toward a grade the same way a finished paragraph does, sometimes more.

 

The process for using AI runs in both directions, for teachers as well as students. Mrs. Thalia Trussell, also in our high school English department, uses AI to run her own instructions and rubrics through an accessibility check for students with dyslexia and other exceptionalities. She tells students to do the same when directions are not landing.

On feedback, she writes her own notes first, usually a couple of sentences per student, then has AI expand them so the reasoning behind a mistake is clear while the tone stays right. She still reads every expanded note before it goes out. After a full round of feedback, she asks the AI what she flagged most often across the class and builds the next lesson around that pattern.

Trussell and Polanco arrived at the same practice from two different assignments. Neither used AI to produce the final answer. Both used it to see across their own students’ work faster than they could alone, then made the next teaching decision themselves.

The Question This Leaves for District Leaders

Every AI policy has a wager built into it about what it rewards. It is time for policies to say that wager out loud and be rewritten for the changes AI actually requires. Grade only the finished draft, and the policy has room for nothing but AI misuse. Grade the distance between attempt one and attempt three, and the policy drives the correct use of AI instead. Either way, a policy has to say so explicitly or teachers will keep grading the old way out of habit.

So here is the question I would ask any district leader reading this: when a student in your building gets something wrong on attempt one, does your policy get them to attempt two, or does it stop there?