The Four Modes of AI Assignments
Mode 0 to Mode 3: a permission scale set per assignment and per phase of the work, in place of a schoolwide line about using AI responsibly in the classroom.
How much AI should students use on this assignment?
Setting the policy per assignment, not per school
A schoolwide line like “AI may be used responsibly” gives a ninth-grader nothing to act on, and it gives you nothing to point at when a submitted essay reads like nobody wrote it. Responsibility is not a behavior. The student who pasted the prompt in and the student who argued with the chatbot for an hour can both say they were responsible, and both are telling the truth as they understood the rule.
The unit of an AI policy is the assignment, and inside the assignment it is the phase. What a student may do while reading a source is a different question from what they may do while drafting a thesis. A policy that cannot tell those apart leaves the student to decide, which is not their job. Four modes cover the range.
The Four Modes of AI Assignments answer how much. The Five Tests ask the prior question of whether AI belongs in the task at all, starting from UNESCO’s guidance on generative AI in education. Attempt → Coach → Verify → Transfer sets the order of events inside the lesson once you have said yes. We built all three out of the research; none of them is a validated model, and they sit together on the AI and K-12 hub.
What each mode permits, and when to reach for it
The four differ on who may touch AI and at which phase of the work.
The same four, with the full conditions on each:
Modes attach to phases, so one unit can carry several. A research project might run Mode 1 while you build the source packet the night before, then Mode 0 for the background-knowledge quiz, then Mode 2 through drafting. Label each phase on the assignment sheet. A course-level rule cannot tell a student which phase they are in.
Mode 0 needs one more line than most teachers expect. A bare “no AI” never says whether spellcheck, text-to-speech, translation, and approved accommodations are still allowed, so say which of them are. A student with a read-aloud tool in their IEP should not have to guess whether the rule took it away.
Writing the mode into the assignment itself
Students meet a mode as a paragraph at the top of the handout, which means that paragraph carries the whole policy. A Mode-2 policy for a document-based argument runs about like this:
mode-2-policy.txt
AI stays closed until your claim-and-evidence chart and working thesis are done. After that you may use the approved chatbot to challenge your reasoning and to apply the class rubric. It may not give you evidence, citations, historical facts, sentences, or a replacement thesis. Turn in the part of the exchange that mattered, plus a short memo on what you accepted, changed, or rejected. Every factual claim comes from the assigned source packet. You will write a related in-class response with no AI.
Every clause there is load-bearing. The opening condition is the difference between Mode 2 and unrestricted use from the first minute: a student who has already committed to a thesis reads a counterargument as something to answer. The role limits name what the tool may and may not produce, so a dispute gets settled by looking at the assignment sheet. The revision memo is the cheapest look you get at what a student did with the suggestions. The in-class response tells you whether any of it stuck.
The memo gets easier to write when students keep a verification record as they draft. Four columns: the AI suggestion, the evidence the student checked it against, whether they accepted, changed, or rejected it, and why. A student goes to the library database, cannot find the citation the chatbot offered, and writes “reject: source could not be found.” That is epistemic vigilance, which is checking a claim instead of complying with it.
Modes govern what students may do while working. How you gather evidence of learning once students have access to AI is assessment design.
Redesigning an assignment you already teach
Teachers rarely design an assignment from nothing. The mode gets retrofitted onto a unit you have taught for six years, which takes five moves.
Name the non-negotiable learning. Write down what students in this course must be able to do with nobody helping. In social studies that list usually holds corroborating accounts, telling evidence from assertion, explaining causation, and constructing an argument in writing. It is the protected core, and every later move on this list is a decision about how to protect it.
Audit the assignment against a chatbot. Ask whether a general chatbot could produce a satisfactory product, whether a student could finish without reading the assigned sources, and whether the rubric rewards polish more than reasoning. Then ask the harder one: what would become invisible to you if a student used AI here? An assignment a chatbot can complete easily is not necessarily obsolete. It may need a different sequence and a different evidence structure instead.
Classify the mode, then rebuild the evidence. Label every phase with one of the four. Add the kinds of evidence the mode leaves you short of — an initial attempt, source annotations, a revision memo, an unseen transfer task. Two or three is plenty. You end up with more than the finished product to look at.
Define the AI role in verbs, and grade the verification. Never write “AI allowed.” Write what it may do: ask questions, generate a counterargument, apply the rubric without assigning a grade, format a bibliography once the sources have been checked. Write what it may not: supply prose, evidence, or citations. Then put points on the verification, so catching an AI error and explaining a rejected suggestion are worth something on the rubric.
Protect the human parts, then check locally. Specify where students have to defend a position aloud, teach a peer the causal chain, or work with no devices on the table. Judge the redesign on independent no-AI transfer, delayed retention, and source-verification accuracy. Students enjoying it, finishing more work, or handing in cleaner prose is not grounds to approve it.
Building the materials a Mode-2 assignment needs
A Mode-2 sequence asks you to make three things before class: the attempt task students complete first, a source packet they can actually read, and a no-AI transfer task for the end. That is three things to build before the first bell, on top of the lesson itself.
Kindred K-12 is built for that part of the work. Bring a source in as a link, a PDF, or a photograph of a page. Cut it to the passage you want, adjust the reading level, or translate it for the students who need that, then build the worksheet around it. Kindred K-12 has no student accounts, keeps no personal information about students, and its AI never chats with a student. If you are looking at this for a department, what Kindred K-12 offers instructional leaders is the place to start.
Labeling the mode before the work starts
Next time an assignment comes back suspiciously polished, ask which mode it was in, and whether it ever said so on the page students read.
Most schools already have an AI policy. Actually getting it into classrooms is the hard part. We help with that. Talk to us.