Using AI in high school
A high school student's AI permissions belong to the task, not to the grade level. Six levels of autonomy, and how to set them one assignment at a time.
By ninth grade the question stops being whether a student may use a chatbot and becomes which task this is. The same student who pressure-tests an argument with one on Tuesday sits down Thursday to a ten-minute in-class paragraph with nothing but a pen. That paragraph is the unaided sample everything else she hands in gets read against. Grades 9-12 are where the distinction becomes teachable: older students can carry the whole discipline and be assessed on it.
Graduating autonomy task by task
Ninth through twelfth graders should be able to tell six levels of autonomy apart. Four sit between the extremes: mechanical assistance, feedback and coaching, research support, and AI-integrated disciplinary inquiry. At one end is an AI-free assessment. At the other is the substitution that stays prohibited whatever else an assignment allows.
A teacher’s job is to decide which one a given assignment is and to say so in the instructions. Teaching students to make the same sort for themselves is work a blanket rule cannot do.
Blanket rules fail in both directions, and they fail quietly. A department that permits AI across the board gets students using it on the one task that was measuring independent proficiency. Nobody notices until a secure sample comes back short of the take-home work. Ban it outright and the same students use it at home, unsupervised, where nobody at school ever sees it.
Fit depends on what the student already knows. A source-bounded argument coach that sharpens an AP history student’s reasoning may be a poor fit for a fourth grader, whose reading and background knowledge are still being built. So is the ability to tell fluent language from evidence. That is the developmental test inside the Five Tests, and it runs task by task in every grade. The elementary and middle school guidance answers it with tighter defaults, which are design recommendations rather than findings; no experiment has established an age threshold behind them.
Five obligations make AI-present work assessable
A finished essay cannot show who did the thinking, and paper does not settle it. A blue book gives you a time-bounded independent sample and makes it harder to reach a chatbot during the period, but a student can prepare material beforehand, memorize it, and reproduce AI-written sentences by hand. A handwritten final answer reveals almost nothing about how sources were chosen or how an argument developed.
The finished essay is the one part nobody doubts. What the teacher is missing are the other two sides of the evidence triangle, and five obligations supply them.
Three produce process evidence, the record of how the work was made:
- a process record: notes, annotations, and drafts;
- verified claims, with a source log showing where each one was checked;
- disclosure of material assistance, in a revision memo and the relevant part of the AI exchange.
Two produce performance evidence, the student doing it in front of you:
- a defense of the work, unrehearsed and out loud;
- a demonstration of independent transfer: a similar prompt, argued with no AI.
Ask for documentation only where AI materially affected the intellectual product. Requiring a full transcript for every trivial use buys compliance work and no thinking. A disclosure worth reading is short. Name the tool, the purpose, and the stage of the work; then one suggestion the student accepted, one they rejected or corrected, and how they checked accuracy. Written that way, the disclosure is itself an exercise in judgment.
A policy that never reaches the assignment guides nobody
A schoolwide statement that AI may be used responsibly is too vague to guide behavior. It never reaches the assignment, which is where the decision is actually made, and three frameworks divide that decision among them.
The Five Tests ask whether to authorize the use at all. The Four Modes of AI Assignments then sets how much AI the assignment allows, phase by phase. Attempt, Coach, Verify, Transfer fixes when in the work the tool appears, opening it only after an independent attempt. All three are our own synthesis of the research rather than published standards, and designing assessment for an AI-present classroom covers the evidence the resulting judgment rests on.
Social studies puts all of that into one assignment. Knowledge and an initial argument come first, with no AI. A coach bounded to the rubric and the source packet then questions the argument without writing any of it, and the student files a revision memo recording what they took and what they refused. A defense and an independent transfer task close it out. The six phases are worked end to end on the social studies page.
The National Council for the Social Studies has connected classroom ChatGPT use to lateral reading and fact-checking rather than to prohibition. The same page carries the routine: extract a claim, leave the chatbot, verify independently. Teaching of that kind, embedded in the discipline the student is already working in, is usually stronger than a separate lesson about AI. For what students should understand about the systems themselves, see teaching AI literacy to students.
Naming the level before you set the assignment
The next time an assignment lands where AI could plausibly help, the useful question is which of the six levels that particular task belongs to. The Kindred K-12 newsletter goes out with more guidance of this kind.