K-12 AI Assistance Ladder for Teaching and Learning

Current status: This ladder is an instructional planning framework, not a universal rule, proof of learning, legal advice, or a representation that every AI product can reliably stay within a selected level.

Use the lowest level of AI assistance that preserves the learning objective and the evidence of mastery an educator needs. Move higher only when additional help is part of the objective, removes an access barrier without replacing the target skill, or supports staff work under appropriate human review.

The K-12 AI Assistance Ladder is a TrueMadeAI planning framework that gives teachers and districts a shared vocabulary for doing that. Its seven-level taxonomy is an original synthesis informed by the research and public guidance cited on this page, not an external standard or certification. It has seven levels:

0. AI off
1. Explanation
2. Hints and guided practice
3. Feedback and error diagnosis
4. Collaboration and critique
5. Full generation
6. Agentic execution

The levels describe what AI may do for a particular task. They do not grade a product, certify an output, or prove that a student learned. The same student might appropriately use Level 0 during an independent check, Level 2 during practice, and Level 4 while revising a project.

The seven levels at a glance

Higher is not automatically better, more advanced, or less ethical. It simply delegates more of the task. A Level 5 activity can be instructionally sound when students must audit a generated essay against primary sources. A Level 2 activity can undermine an assessment if the hints disclose the reasoning the teacher is trying to measure.

Level 0: AI off

AI is not used during the defined work. The teacher may still have used AI earlier to prepare materials under district rules, but the student task itself is completed without it.

Use Level 0 when the objective requires an independent sample of recall, fluency, reasoning, writing, speaking, calculation, observation, or performance. Examples include:

• a grade 2 student reads a passage aloud so the teacher can observe decoding;
• a grade 5 student completes a short fraction check after guided practice;
• a grade 8 science student records direct lab observations before receiving feedback;
• a grade 11 student writes an in-class source analysis that establishes independent mastery.

State the boundary precisely. “AI off for the ten-minute mastery check” is clearer than “never use AI in math.”

An AI-off rule should not silently remove an approved accommodation. Determine whether the access support can remain available without performing the target skill. For example, text-to-speech may change the construct in a decoding assessment but may be an appropriate access support when the objective is scientific reasoning. See AI, accommodations, and student privacy.

Level 1: explanation

AI may explain a concept or direction, but it should not complete the assigned problem or produce the response.

Examples include:

• a teacher-mediated tool explains the difference between a fact and an opinion to a grade 3 class using two new examples;
• a grade 6 learner asks for a visual explanation of equivalent ratios before attempting the assigned questions;
• a grade 9 biology student asks for a definition of homeostasis without providing the worksheet item;
• a grade 12 government student asks for neutral background on judicial review, then works from assigned primary sources.

The distinction between explanation and completion depends on the assignment. Explaining the Pythagorean theorem may be Level 1 before practice. Explaining exactly how to solve the numbered item on an independent assessment may perform the assessed work.

Level 2: hints and guided practice

AI may support the next step while keeping the learner responsible for solving the problem. Useful behavior can include asking a diagnostic question, pointing to a relevant concept, offering a parallel example, or releasing a hint only after an attempt.

Examples include:

• a grade 4 student who has attempted a multi-step word problem receives one question about which quantity is unknown;
• a grade 7 pre-algebra student receives a hint to check the inverse operation rather than a completed solution;
• a grade 10 chemistry student is asked to identify which values are known before selecting an equation;
• a beginning coding student receives a description of the failing condition without generated replacement code.

A published high school mathematics study compared an unrestricted GPT-style interface with a tutor configured to provide learning-oriented safeguards. It found that interface and guardrail design changed both practice behavior and later performance. One study in one subject and setting is not a universal effect estimate, but it is strong reason to distinguish access to answers from guided help rather than treating all AI assistance as equivalent. Read the PNAS study.

Level 3: feedback and error diagnosis

AI reviews work the student has already produced. It may identify a possible misconception, point to a rubric criterion, ask the student to verify a source, or explain why a test case fails. The student decides what to change and performs the revision.

Examples include:

• a grade 5 writer submits one paragraph and receives feedback about whether each reason supports the claim;
• a grade 8 science student compares a lab conclusion with a teacher-provided rubric and checks whether the evidence supports the claim;
• a grade 10 Spanish student receives feedback on verb agreement without having the response rewritten;
• an AP Computer Science student receives a failing test case and a description of the likely logic error.

Require a visible first attempt when it matters. Without one, “feedback” can quietly become full generation. Educators should also verify that the AI’s diagnosis is correct and aligned to the assigned rubric.

Level 4: collaboration and critique

AI becomes a structured thinking partner. It may propose competing explanations, challenge an argument, role-play a stakeholder, compare strategies, or help organize ideas. The student remains the decision-maker and must verify factual claims.

Examples include:

• a grade 6 social studies class compares two AI-generated explanations with the textbook and identifies missing context;
• a grade 8 engineering team asks for three design tradeoffs, then tests the options against its own constraints;
• a grade 10 English student asks AI to challenge a thesis after completing an evidence outline;
• a grade 12 civics student questions an AI-generated stakeholder perspective using primary-source evidence.

At this level, the learning evidence often shifts from the first artifact to the student’s evaluation, source checking, decision record, oral defense, or revision history.

Level 5: full generation

AI produces a complete artifact. That can be educationally appropriate only when the teacher is not using the generated artifact as proof that the student independently performed the generation skill.

Appropriate examples can include:

• students audit an AI-generated historical summary for unsupported claims and omissions;
• a media-literacy class compares three generated public-service announcements for audience, evidence, and bias;
• a coding class tests generated functions against teacher-written cases and documents each correction;
• an art class critiques composition choices in generated reference images before creating an original work under the assignment rules.

If the objective is “write a coherent evidence-based essay,” submitting a generated essay does not demonstrate independent writing. If the objective is “evaluate and improve a generated argument using verified evidence,” generation may be part of the task.

Level 6: agentic execution

AI plans and performs a sequence of actions, potentially using tools or changing an artifact. This level requires the clearest permissions, stop conditions, logs appropriate to the use, and human review.

A student example might be a high school computer science sandbox where an agent can modify a disposable project, run teacher-provided tests, and propose a patch while the student explains and approves each change. It should not have access to production systems, unrelated files, student records, or external publishing.

Staff examples might include assembling a first draft of a public-information packet from approved documents or transforming a non-sensitive dataset into several proposed charts. The responsible employee must still verify facts, privacy, accessibility, tone, and authority before use.

Do not treat consequential decisions as ordinary automation. Grades, placement, discipline, disability determinations, safety decisions, personnel actions, and communications made on behalf of the district require defined human authority and additional legal, policy, privacy, and validity review.

Choose the level from the objective, not the product

Before selecting a level, answer these questions in order:

1. What is the learning objective? Name the knowledge, reasoning, process, expression, or AI-literacy skill the activity is intended to develop.
2. What evidence of independent mastery is needed? Decide what the student must do without AI and whether that evidence occurs before, during, or after assisted practice.
3. What has the teacher directed? State the permitted level for this assignment, not only the general course policy.
4. What access or accommodation is required? Preserve approved supports and distinguish barrier removal from substitution for the skill being assessed.
5. Is the use age-appropriate and account-eligible? Review provider terms, district approval, student readiness, adult supervision, and an equivalent non-AI path where appropriate.
6. What data is involved? Use only an approved product and account for the intended data. Minimize student, family, employee, confidential, and protected information.
7. What does district policy allow? The selected level cannot override district privacy, safety, acceptable-use, procurement, security, records, or academic-integrity rules.

UNESCO recommends a human-centered, age-appropriate approach to validation and pedagogical design. The U.S. Department of Education’s education-leader toolkit likewise places AI adoption inside instructional, privacy, security, civil-rights, equity, and community processes. The ladder turns those broad responsibilities into an assignment-level question, but it does not replace them.

Preserve evidence of independent mastery

Assisted performance and independent learning are different measurements. A student may solve more practice items while help is available without retaining the skill when the help is removed.

For a consequential learning claim, use a sequence such as:

independent baseline
-> AI-assisted practice at the selected level
-> teacher or peer feedback
-> independent check without the assistance
-> transfer task in a new context when appropriate

The independent check does not have to be a high-stakes test. It can be a short explanation, worked example, oral conference, new problem, source defense, or code walkthrough. The key is that the evidence matches the learning objective.

This is also why model accuracy alone is not a teaching score. K-12EduBench separates subject knowledge, problem-solving, and educational-goal cognition, while the LearnLM report treats pedagogical behavior as instruction following that can be evaluated. See how TrueMadeAI evaluates AI for K-12 learning for the proposed scorecard methodology and evidence boundaries.

Apply accommodations without erasing the target skill

Accessibility can justify a different interface, representation, response mode, pacing, or support. It does not mean every student needs the same tool or that the highest assistance level is automatically appropriate.

CAST’s Universal Design for Learning Guidelines encourage multiple means of representation, engagement, and action or expression, along with graduated support for practice and performance. A district team can use those principles while asking two concrete questions:

1. Which barrier does the support remove?
2. Does it also perform the knowledge or skill this task is intended to measure?

For example, reading directions aloud may remove an access barrier in a math-reasoning activity. It may change what is measured in an oral-reading fluency check. The responsible educator or district team must make that distinction for the learner and context, not the model.

Do not copy a full IEP, Section 504 plan, diagnosis, or unrelated learner record into general AI context merely to configure assistance. Use the least information necessary through an approved workflow and provide a fallback if the AI surface cannot deliver the support reliably.

Student learning and staff operational work are different

For students, the central question is often: What must the learner think, make, or perform for this activity to produce valid evidence of learning?

For staff operational work, the central question is different: Can AI reduce effort while a qualified person retains authority, protects data, and verifies the result?

That means higher assistance can be reasonable for some staff tasks even when it would be inappropriate for a student mastery task.

Do not reuse a classroom assistance level as a blanket staff permission. Put staff uses in the district’s AI application register and approval workflow.

Write an assignment AI card

A teacher can publish the selected level in a compact assignment card:

Learning objective: Build and defend an evidence-based claim from the assigned sources.
AI assistance: Level 4, collaboration and critique, after completing your own claim and evidence outline.
Allowed: Ask the district-approved AI account for two counterarguments and questions that challenge your reasoning.
Not allowed: Generate the final response, invent sources, or upload another student’s work.
Evidence: Submit the original outline, the counterargument you selected, source verification, and the final response.
Disclosure: Include a two-sentence note describing how AI affected your revision.
Alternative: Ask the teacher or a peer reviewer for the same counterargument exercise.

Use the K-12 AI acceptable use policy checklist to connect these assignment cards to the district baseline. The K-12 AI governance guide covers the wider operating model, and the AI governance readiness assessment helps identify missing ownership and implementation controls.

How the ladder relates to Tenet District

The ladder is a planning framework. It is not a claim that Tenet reads an assignment, recognizes one of seven labels, or perfectly determines whether an AI answer stayed within a level.

Tenet Basic applies one district-wide baseline on supported managed Chrome paths. Tenet District adds grade, roster, class, teacher, subject, period, and schedule context where the selected capability supports it. A district and teacher can translate an assignment’s assistance level into concrete rules, such as “offer guiding questions but do not provide a completed solution,” on supported direct-use AI surfaces.

During a scheduled class, Tenet District can use available class and schedule context to select a configured teacher rule on a supported path. For homework or off-period work, a teacher can enable a take-home rule. Local subject detection can then try to match the prompt to an enabled class rule. Subject detection is probabilistic. If there is no reliable match, the district baseline remains.

Available controls vary by product surface, content type, account, configuration, tier, and rollout scope. Native applications, files, images, voice, connectors, embedded assistants, unmanaged devices, and newly changed interfaces require separate validation. See Tenet Basic compared with Tenet District and the supported-products capability matrix for current boundaries.

A district implementation sequence

1. Adopt the seven labels as instructional language, not as a technology approval by themselves.
2. Map common assignment types to a default starting level while preserving teacher judgment inside district policy.
3. Require teachers to state exceptions, independent evidence, disclosure, approved accounts, data rules, and alternatives.
4. Review age, accessibility, privacy, security, and account eligibility before student use.
5. Configure supported technical controls from the concrete rule, not only the level number.
6. Test representative products, prompts, roles, schedules, content types, and important near-misses.
7. Compare assisted work with an independent check when making a learning claim.
8. Review the framework after incidents, product changes, curriculum changes, and evidence from educators and students.

NIST’s AI Risk Management Framework emphasizes that governance, context mapping, measurement, and risk management continue across the system lifecycle. A ladder helps communicate instructional intent, but districts still need the ownership, product review, data boundaries, testing, incident response, and change management described in the broader governance program.

Frequently asked questions

What is the K-12 AI Assistance Ladder?

It is a seven-level planning framework that moves from AI off through explanation, guided hints, feedback, collaboration, full generation, and agentic execution. Educators use it to match permitted AI help to a learning objective and the evidence students must produce independently.

Should students always use the lowest AI assistance level?

Use the lowest level that serves the objective and preserves needed evidence of mastery. A higher level may be appropriate when evaluating AI output is the objective, when the teacher intentionally teaches collaboration with AI, or when an approved accessibility support removes a barrier without replacing the target skill.

Which AI assistance level is appropriate for homework?

There is no universal homework level. The teacher should state the objective, permitted level, required process evidence, disclosure expectation, approved account, data rule, and an equivalent path when needed. Independent practice often calls for AI off, explanation, or hints rather than answer generation.

How should accommodations affect the assistance level?

An accommodation can change how a student accesses directions, content, or a response mode without changing the skill being assessed. The responsible district team should distinguish barrier removal from substitution for the target skill, minimize learner-specific data, and provide a tested fallback.

Is full AI generation ever appropriate for students?

Yes, when generating or evaluating an AI artifact is part of the stated objective, such as critiquing alternative drafts or testing code. It is usually not appropriate when the submitted artifact is supposed to demonstrate the student’s unaided writing, reasoning, calculation, recall, or performance.

Can Tenet District enforce all seven assistance levels automatically?

No. The ladder is an instructional framework, not a universal detector or a claim of complete technical enforcement. Tenet District can apply configured district and classroom rules on supported direct-use AI paths, with context that can include grade, roster, class, teacher, subject, period, schedule, and supported take-home behavior. Coverage varies by product surface, content type, configuration, and rollout scope.

Sources

• Empowering Education Leaders: A Toolkit for Safe, Ethical, and Equitable AI Integration, U.S. Department of Education
• Guidance for Generative AI in Education and Research, UNESCO
• Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics, PNAS
• K-12EduBench, Proceedings of the AAAI Conference on Artificial Intelligence
• LearnLM: Improving Gemini for Learning, LearnLM Team
• CAST Universal Design for Learning Guidelines Version 3.0, CAST
• Artificial Intelligence Risk Management Framework Core, NIST

This resource is educational information, not legal advice, an individualized accommodation decision, or proof that a particular AI system improves learning. Districts should use qualified instructional, accessibility, privacy, legal, security, procurement, and community review for their circumstances.