Five questions settle whether AI belongs on an assignment, and all five have to come back yes. They run before any decision about how much AI a student gets, because a use that hands the thinking to the tool does not become sound when you shrink it. Thirty-five common classroom tasks are sorted below by the condition each one needs before AI touches it.

Whether comes before how much

You make three decisions about AI on an assignment, and running them in the wrong order wastes the work. Each has its own guide in our AI and K-12 library. The five tests decide whether AI belongs at all. The Four Modes of AI Assignments decide how much access students get; Attempt → Coach → Verify → Transfer decides when AI enters. Under that last sequence, students produce evidence of their own thinking first, AI takes a narrow role second, the teacher verifies, and the student performs the skill unaided.

Run the tests before you design anything. All five yes, and you go pick a mode. Any no, and you redesign the task or take the tool off it.

Test 1: What AI adds that a text, a partner, or a teacher would not

Fail signal

The strongest answer available is that students find it engaging, that it saves time, or that AI is the future.

UNESCO’s guidance on generative AI in education asks that using the tool make learning more effective than a no-tech or other alternative approach, which puts the burden on a comparison. Name the thing you are beating. Sometimes that is a text, a partner, or the teacher; sometimes a graphic organizer, a worked example, or nothing at all. Answers that survive the comparison describe a mechanism: every student gets rapid rubric-aligned questions after a first draft, or the class tests counterarguments against a source packet. Converting an inaccessible format without dropping the conceptual demand qualifies, and so does repeated oral rehearsal before a live discussion. UNESCO’s is one of many external frameworks a school ends up stacking, and the frameworks glossary sorts the rest.

One answer passes on different grounds. Where the learning goal is to understand and critique AI systems, the tool is the object of study, and the comparison never arises.

Test 2: Who performs the thinking the objective names

Fail signal

AI performs most of the verbs in the objective.

Write out the verbs your objective actually contains — identify, retrieve, compare, infer, contextualize, argue, organize, compose, revise, calculate, model, interpret — and then say who performs each one on the assignment as designed. If AI performs most of those verbs, the student is probably supervising rather than learning. The finished artifact is not where that shows.

We like this one because it costs a minute and needs no new material: you already wrote the objective, and the verbs are sitting in it waiting to be counted.

Test 3: Whether the sources and reasoning can be inspected

Fail signal

A consequential claim arrives with nothing behind it a student could check.

Students and teachers both need to see where a claim came from and how the tool got there. A fluent unsourced sentence looks exactly like a sourced one at the moment a student decides whether to believe it, and in social studies, history, science, and research writing, untraceable claims should normally be excluded.

Two designs satisfy the test: the AI operates from a source packet you supply, or its claims get independently verified before anyone builds on them. Keep the verification as a short written record. A student checks “consider economic motives” against source 4, paragraphs 2–3, and modifies it, because the source supports a specific trade incentive and not a general economic motive.

Test 4: Whether the student can use the tool critically yet

Fail signal

The student cannot reliably tell fluent language from evidence.

A system that is productive for an AP history student may be inappropriate for a fourth grader. Developmental fit covers reading ability, background knowledge, impulse control, understanding of privacy, the ability to recognize uncertainty, and the capacity to distinguish fluent language from evidence. Any one of those missing turns a critical-use activity into an acceptance activity. Check the fit against the students you have.

Test 5: How you will see the skill when AI is absent

Fail signal

Nothing in the unit shows the student performing the skill without the tool.

Decide, while designing the task, what the student will produce unaided that demonstrates the same skill. A new but analogous source, an in-class paragraph without AI, an oral defense, delayed retrieval a week or two later: the form matters less than where it sits. It has to land after the AI-assisted work. Without transfer evidence there is no secure inference about learning: you know the work got done, and nothing about who can do it next week.

Which tasks AI fits, and which it doesn’t

AI is best suited to work that is low-stakes, reversible, frequent, and independently checkable. It is least suited to work that is high-stakes, authoritative, identity-sensitive, irreversible, or itself the intended learning process. Find the task you had in mind, read its condition, and go back to the five tests. A row appearing here does not exempt it from them.

Generally strong or promising uses

Task	Design condition
Rubric-aligned feedback	Only after an initial student attempt; prohibit rewriting; require a revision memo
Counterargument generation	Student constructs the original claim first; AI provides objections; student verifies and responds
Question generation	Teacher or student reviews the questions for accuracy and alignment
Retrieval-practice variants	AI generates questions; students answer without AI; teacher checks key items
Worked-example comparison	Student compares an AI solution or paragraph against a teacher exemplar
Error analysis	AI provides intentionally flawed or naturally flawed responses for diagnosis
Oral rehearsal	Low-stakes practice before a live presentation, seminar, or language exchange
Vocabulary and language scaffolding	Preserve disciplinary terminology and conceptual demand
Formatting and accessibility	Convert structure, text presentation, or directions while preserving the learning goal
Feedback translation	Translate teacher-approved feedback, with attention to meaning and tone
Idea expansion	Student supplies the initial ideas; AI proposes alternatives; student evaluates them
Source-checking exercises	Treat the AI answer as a claim set to be investigated, never as a source
Perspective comparison	Use abstract or documented positions; avoid fabricated personal testimony
Metacognitive reflection	Student identifies where AI helped, misled, or changed the approach

Conditional uses

Task	Conditions
Summarization	After the student has read and summarized, or where summarization is not the learning objective
Brainstorming	Require an initial human list and a rationale for what gets selected
Translation	Verify important nuances, names, quotations, and culturally specific language
Reading-level adaptation	Compare against the original; do not remove essential concepts
Practice conversations	Bound the topic; review for cultural and factual problems
Study guides	Build from teacher-approved curriculum; students should contribute retrieval questions
Research planning	AI may suggest search terms and categories; students locate and evaluate the actual sources
Coding or data analysis	Require explanation, debugging, and an independent application
Exemplar generation	Label as synthetic; teacher verifies quality; avoid turning one example into a hidden template
Peer-feedback preparation	AI may help a student formulate feedback; it does not replace peer response

Generally poor or high-risk uses

AI as the first and authoritative explanation of unfamiliar content.
Autonomous generation of final essays, reports, claims, or reflections.
Generation of citations or quotations without independent database verification.
AI-written reading responses where reading comprehension is being assessed.
AI-generated personal reflection, identity writing, or social-emotional disclosure.
Autonomous high-stakes grading.
Discipline, counseling, special-education, safeguarding, or placement decisions.
Fabricated historical testimony or impersonation of victims.
Automated evaluation of political beliefs, personality, motivation, or character.
Foundational practice in reading, handwriting, number sense, or sentence construction where the practice itself is what the student needs.
Continuous chatbot access through every phase of a lesson.

Once a use clears the tests, the next question is what evidence you will collect, and designing assessment for an AI-present classroom takes it from there.

Where the Five Tests come from

We wrote them. The Five Tests are Kindred K-12’s own synthesis from our research on instruction in an AI-present world, so treat them as a decision procedure that has to earn its keep in your room. Nobody else has validated them as a model. The recommendation inside the first test belongs to UNESCO and is theirs to stand behind.

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