A chatbot can improve the work a student hands in without improving anything that student can do the next morning on their own. A field experiment in Turkey with nearly 1,000 high-school students measured both halves of that. Students working with a general GPT interface performed about 48% better during practice, then about 17% worse than the control group once the tool was taken away. A safeguarded GPT tutor produced a gain of roughly 127% during practice and largely avoided the decline. That study ran in mathematics and outside the United States, so its numbers do not transfer to a U.S. social studies room, and a 2026 Stanford review that screened 818 studies found only 20 with strong causal evidence, none of them student-facing studies in American K-12 classrooms. What survives the caveats is the shape. Each of the seven uses below hands the tool the thinking the lesson was built to produce; the AI in K-12 hub holds the other half, on running the good uses well.

1. Generating answers before students have the knowledge to judge them

What it looks like in class

Students meet a new topic by asking a chatbot to explain it, before any reading, source, demonstration, or discussion.

A student cannot evaluate an answer in a subject they have not studied yet. They have no way to see what the answer omitted, flattened, or got wrong, because seeing that is the knowledge the unit was supposed to build. Open chatbot use during first exposure closes a circle: the student needs the knowledge to judge the system and is using the system in place of acquiring it.

Take a class that opens a unit on the Dust Bowl by asking a chatbot what caused it. Nobody in the room can yet tell whether the answer treats the Dust Bowl as a drought that happened to farmers or as the consequence of how that land had been broken and farmed. Telling those two apart is the unit. Whichever account comes back sits underneath every later task, and you meet it again in their essays without ever having seen where it came from.

Put the chatbot after the first four steps, not before them. Teach the content, from instruction, a text, a source, or a discussion. Have students retrieve and explain it in their own words, then apply it once. Bring AI in at that point to critique or extend what they produced, and finish with a transfer task they do without it.

2. Assigning AI the work the assignment exists to teach

What it looks like in class

The objective is for students to write a thesis, and the tool writes the thesis.

An assignment turns unstable the moment the capability being automated is the capability being taught. If the objective is to build a thesis, that is precisely the thesis the tool should not build. The same holds for choosing the evidence, summarizing a hard source, writing at the sentence level, and judging whether a source is credible. What arrives on your desk still looks like the assignment, while the grade has quietly moved onto how well each student directed a model.

The tool still has real jobs in that assignment. It can ask questions about a draft, point out what the student left unaddressed, build the strongest opposing case, or hold the student’s own writing against the rubric. Every one of those puts the student back into the capability you meant to develop. Deciding which of those a given task can bear is its own judgment, and the five tests for any classroom AI use are how we work through it.

Write down the one capability the task is meant to develop and grade, then hand over everything that is not it. Formatting, translation, mechanical cleanup, and locating a passage the student has already identified are all fair to automate. The capability you wrote down is not.

3. Calling a reading-level change personalization

What it looks like in class

A worksheet is regenerated three grade levels down and handed out as differentiated instruction.

Ask a model to simplify a passage on the Fourteenth Amendment and it will do it well, and “due process of law” may come back as “fair treatment.” The lesson was about that phrase. Simplification of this kind removes the disciplinary vocabulary a standard names, drops the conceptual difficulty that made a reading worth assigning, or supplies a friendly analogy that quietly misleads.

You can change a passage’s reading level in seconds. That adaptation is worth doing, and instructional personalization needs much more: what this student already understands, which misconception is in the way, their language and cultural context, and the standard the work answers to. You also have to know what evidence would show improvement. A general-purpose system holds none of that. A 2026 systematic review of 28 studies of teachers designing with AI kept finding the same two weaknesses in the output: generic content, and poor alignment to the curriculum and the local context.

Say out loud which difficulty you are removing. Confusing wording, an inaccessible layout, an unfamiliar term the standard does not test, directions in a language the student is still learning — remove those freely. Retrieving what they know and tracing evidence to a claim is the assignment itself, and it stays where it is.

4. Delegating a whole lesson

What it looks like in class

A complete lesson plan arrives in twenty seconds and goes straight into tomorrow’s second period.

A generated plan reads well at every individual point and may still not hold together as an hour. Coherence lives in the alignment among what students already know, the goal, the misconceptions you expect, the checks that would catch them, and the lesson this one has to set up. A model assembles those parts from the shape of plans it has seen. You find out whether they cohere in front of the class.

The evidence on teacher design supports AI as an ideation and drafting partner much more strongly than as an autonomous designer — a large EEF/NFER randomized trial found the teachers who actually saved planning time were the ones using AI for one component at a time, a quiz item or an activity idea, never a whole lesson (more on that trial).

Hold the goals, the sequence, and the evidence of learning; delegate components. Ask for three options rather than an answer, then verify, edit, and localize what comes back. The saved time only counts as a gain if it goes to feedback, conferencing, or curriculum study rather than being absorbed by other paperwork.

This is the part of the job Kindred K-12 is built for: you decide what the lesson is for, and it drafts the materials around that decision. It does not decide the objective.

5. Grading only the finished product

What it looks like in class

A unit grade rests almost entirely on one polished essay written at home.

Grade a take-home artifact and you may be measuring several things at once, most of which you did not intend. Access to tools moves that score. So does knowledge of prompting, adult help at the kitchen table, editing skill, a read on how firmly the syllabus rule was meant, and the sophistication of whichever model the student could reach. The content knowledge you meant to assess may move it only weakly.

Better proctoring does not repair this, because the defect is in what a single artifact can tell you. One product is one observation, and you need more than one to say a student understands the work.

Collect affirmative evidence in more than one form. A short in-class response written without tools, an oral explanation of a choice the student made, and the process artifacts behind a draft each show you something the polished document cannot. Two of them agreeing is worth more than any one alone. Designing assessment for an AI-present classroom takes that further, including how to build a secure writing sample worth comparing against.

6. Acting on an AI detector’s score

What it looks like in class

A detector returns a percentage, and the percentage becomes the accusation.

Turnitin states that its own AI-writing indicator can misidentify human writing, AI-generated writing, and AI-paraphrased writing, and that it should not be the sole basis for adverse action. Vanderbilt University disabled Turnitin’s AI detector after considering false positives, the opacity of the score, privacy, and the potential effect on students who are not native English writers. Research has also documented systematic detector bias against non-native English writing.

The cost lands unevenly. A false positive falls hardest on the student with the least standing to argue back, and an accusation you cannot substantiate costs you the working relationship the rest of your teaching depends on.

Treat the score as a reason to look, never as the finding. Talk with the student about what they wrote. Compare the work against a secure sample you watched them produce, read the version history and their source notes, and ask for an oral explanation or a short follow-up task. Each of those gives you something you can act on.

7. Treating a simulated identity as testimony

What it looks like in class

A chatbot is asked to “be” an enslaved person or a Holocaust survivor and answer students’ questions in character.

A model asked to speak as a person who lived through something invents experiences it has no record of, and collapses many different lives into one composite. It delivers all of that in the register of a first-hand account. Students then reason from the performance the way they would reason from a document.

The damage is done at that moment. The class has practiced treating generated language as evidence, in the one discipline whose central skill is telling evidence apart from everything that resembles it. Simulation is not testimony, and synthetic language is not a primary source.

None of that makes AI-generated perspective writing off-limits. Used as an object of critique, with the learning objective stated openly, it is one of the strongest things you can do with a model. Students audit the output, and the audit is what you grade. That version is also good fun to teach. Put a fabricated diary entry next to a real one and see how fast someone notices that the fake narrator states his own age, trade, and county in the opening line. That is not how anyone writes for an audience of themselves.

Start from the documentary record, then give the tool a job beside it. Use authentic testimony and scholarship as the account of what happened. AI can generate the questions students should investigate, surface the tensions among the sources you supplied, or produce an explicitly fictional claim for students to take apart. Where a position needs a voice, have it argue the abstract policy rather than impersonate someone who was harmed.

What to carry into next week’s planning

The move that prevents most of the seven takes about a minute: before you decide what AI may touch in an assignment, write down the one capability that assignment exists to build. Everything else on the page is negotiable.

Most schools already have an AI policy. Actually getting it into classrooms is the hard part. We help with that. Talk to us.