Can Teachers Use AI to Grade? District Policy Language and What Utah and DC Actually Require

Current status: Last reviewed September 7, 2026. This resource summarizes primary statutory text, state model policies, and association guidance as read on that date. Utah's operative model policy was not yet published at the time of review. This is not legal advice.

“Can teachers use AI to grade?” is being answered in the press as a yes or no. Every authority that has actually written a rule answers it as a ladder. No state, model policy, or union guidance reviewed here lets an AI system set a final grade. Almost all of them permit AI to help draft formative feedback. The disagreements are in the middle, where a tool proposes a score and a busy teacher accepts it, and in the disclosure question nobody expected: whether students and families have a right to know.

Two widely repeated claims are wrong, and a district that repeats them in a board memo will be corrected in public. Washington DC did not ban AI grading. Utah did not adopt a permissive standard; its 2026 law is the most restrictive one on the list.

If you want the working documents that support this policy, the acceptable-use checklist and the tool vetting template are in the free district AI governance starter kit.

What the authorities actually say

Ohio, Tennessee, Maryland, North Carolina, and Illinois require or direct AI policy work without adding a grading-specific rule in the sources reviewed. That absence is not permission; it means the decision is local.

The DC tiers, in detail

The DC model policy is the most useful document on this list because it sorts staff uses into three tiers rather than issuing a single verdict, and districts in other states can borrow the structure.

Prohibited outright in the DC model: physical surveillance, educational trajectory decisions such as grade promotion, placement, and advanced-course readiness, student discipline, teacher evaluation, health and mental health service decisions, and eligibility determinations for an individualized education program or a Section 504 plan.

Permitted with conditions, including reviewing and grading student work, generating rubrics, and providing formative feedback generated outside class time. The four conditions are worth quoting as a model because they translate directly into district language: educators are responsible for all summative grading decisions; the text of the prompt provided to the AI tool must be included when returning AI-graded work; student identifying data including race and ethnicity must not be disclosed to the tool; and students must be allowed to appeal an AI grade and may ask for a human to provide additional feedback.

Note the seam between the tiers. Drafting IEP language is limited use; determining IEP eligibility is prohibited. Grading an assignment is limited use; deciding promotion is prohibited. That seam is where districts get into trouble, because AI-assisted grades roll into a grade point average that drives exactly the promotion and placement decisions the same policy forbids AI from making.

The Utah rule, in detail

Utah HB 273 took effect July 1, 2026. It defines a high-stakes determination as a decision on a student’s placement, discipline, academic progression, or eligibility for a program or service, and it directs the state board’s model policy to prohibit educators from using generative AI to independently grade student work or issue those determinations.

Two things follow. First, the operative word is “independently.” A teacher who reads the work, forms a judgment, and uses a tool to draft the comment is in a different position from one who accepts a generated score unread, and a district policy should say where that line sits rather than leaving each teacher to guess. Second, Utah pairs the restriction with notice: where generative AI is used for instruction, assessment, or classroom activities, the school must notify the student’s parent in writing.

One caveat for anyone drafting from this: the Utah state board model policy that carries the operative language was not yet published at the time of this review. Districts should track its release before finalizing local text.

The FERPA line teachers cross without noticing

Graded student work maintained by the school is an education record. Disclosing personally identifiable information from an education record to an outside party without consent is permitted only under an exception, and the one that applies here is the school official exception. It requires that the outside party perform an institutional service the district would otherwise use employees for, remain under the direct control of the district with respect to the use and maintenance of education records, and be subject to redisclosure limits. District practice adds a fourth element: designation in the annual FERPA notification.

Apply that to two teachers doing the same task.

A teacher who pastes a student essay, with the student’s name at the top, into a personal consumer account has disclosed personally identifiable information from an education record to a vendor with no district agreement. There is no direct control, no redisclosure limit, and no designation. The vendor’s consumer terms, not the district’s, govern retention and training use.

A teacher who does the same work in a district-contracted deployment covered by a data agreement that forbids training use and permits deletion is inside the exception, provided the district actually exercises the control the agreement describes.

De-identified work falls outside FERPA entirely, which is why the DC model bars sending student identifying data to the tool. The practical trap is that student writing is frequently self-identifying: a personal narrative about a family member, a hometown, an injury, or a diagnosis re-identifies the author even with the name stripped. Any policy that relies on de-identification needs to say so, and needs to say who checks.

The FERPA and AI guide covers the analysis in depth and the vendor and DPA review questions turn it into contract language.

What the reliability research supports

Published work on large language model scoring of student writing reports a consistent pattern: the models are highly self-consistent, producing similar scores for the same work across runs, while agreement with human raters is materially lower. One 2026 study of LLM writing assessment reported strong intra-rater consistency alongside notably weaker agreement between the model and human scorers.

The policy implication is narrower than either camp usually claims. High self-consistency is not evidence of accuracy; it means the tool reliably reproduces its own judgment, including its own errors. That supports using these tools to draft, structure, and check, and it argues against using them to decide. It also gives a concrete reason for the appeal right the DC model requires, since a student contesting a machine score is contesting something reproducible rather than random.

Six questions a district policy has to answer

Most district AI policies stop at “teachers remain responsible.” That sentence does not settle a single argument in a department meeting. These do.

1. May AI draft feedback on student work? If yes, on which assignment types, and must the teacher read the student work first?
2. May AI propose a score? This is the real fight. A proposed score a teacher adopts without independent judgment is the practice Utah’s “independently grade” language targets.
3. May AI set a final grade? No authority reviewed permits this. Say so plainly.
4. What must be disclosed, to whom? Utah requires written parent notice for classroom use. California’s model informs students. The DC model returns the prompt and grants an appeal. Choose, and write the mechanism.
5. Which tool, on which account? A consumer account and a district-contracted deployment are different legal situations for identical work. The AI product age and consent table covers which products offer a school path at all.
6. What about students with an IEP or 504 plan? Eligibility decisions are prohibited in the DC model and are high-stakes determinations under Utah and Oklahoma law. If AI-assisted grades feed a placement or promotion decision for a student with a disability, the district needs to know where that chain runs. The AI in IEP and 504 writing guide covers the adjacent question of drafting.

Sample policy language by tier

District counsel should write the final text. This is a starting structure that matches what the authorities above actually require.

Tier 1, permitted with review. Staff may use district-approved AI tools to generate rubrics, draft formative feedback, suggest revision questions, and check their own comments for clarity and tone. The educator must review and may edit all output before it reaches a student. Student identifying information may not be entered into any tool that is not district approved for that purpose.

Tier 2, limited and disclosed. Staff may use district-approved AI tools to assist in reviewing summative work only where the educator has independently reviewed the student’s work, the educator determines the score, the use is disclosed to students and families in the manner the district specifies, and the student may request human re-review. The tool may not be the sole reader of the work.

Tier 3, prohibited. AI systems may not determine or independently generate a final grade, promotion, retention, placement, course-eligibility, discipline, or special education eligibility decision, nor may AI output be used as the primary basis for such a decision. AI may not be used to evaluate staff performance.

Applies to every tier. Only approved tools on approved accounts. Records of the approval and the rule that applied. Periodic review of feedback quality across student groups.

Making the rule reach the classroom

A grading policy is a promise about behavior in a hundred classrooms at once, which is where these policies usually fail. The teacher who most needs the rule is the one grading eighty essays on a Sunday night, and a PDF in a staff handbook does not reach that moment.

Tenet by TrueMadeAI is K-12 AI governance software. Tenet Edge applies district and classroom policy on supported direct-use AI surfaces on managed Chrome. For a grading policy specifically, a district implementation could use Tenet to help:

• present the district’s rule on supported AI surfaces at the moment a staff member opens one, with Tenet District resolving rules by class, subject, teacher, and schedule;
• apply supported on-device data-loss-prevention checks so student names and other identifiers are caught before text reaches an approved tool, which is the FERPA exposure described above;
• block supported unapproved AI chat and writing interfaces when the district enables the control, so the policy is not quietly defeated by a personal account on a district device; and
• keep bounded evidence that the rule was applied, without treating student work or staff prompts as an analytics archive.

Tenet does not decide which tier a district adopts, does not review grades, and does not cover every product surface. Coverage depends on the specific web surface; review the dated capability matrix before relying on a particular behavior.

Tenet Basic is free and applies one district-wide baseline. Try Tenet Basic, or request a Tenet District conversation if you need the classroom and subject context a grading rule depends on.

What this guide does not establish

This guide does not determine:

• whether a particular grading practice complies with a specific state’s law;
• what Utah’s state board model policy will say when published;
• that a tool’s accuracy is sufficient for any particular assignment;
• that de-identification is achievable for a given student work product; or
• that disclosure to students satisfies a state notice requirement.

Those conclusions require the current official sources, the district’s facts, and qualified review.

Sources

• Utah HB 273 (2026), enrolled copy
• DC OSSE, LEA AI Model Policy for Staff Use, September 2026
• DC OSSE release on the AI model policy
• California State Model Policy on the Use of AI in Education
• Code of Virginia section 22.1-20.2:1
• NEA policy statement on artificial intelligence in education
• AFT, Commonsense Guardrails for Using Advanced Technology in Schools
• U.S. Department of Education, What is an education record?
• Privacy Technical Assistance Center, Responsibilities of Third-Party Service Providers

Frequently asked questions

Is it legal for a teacher to use AI to grade student work?

No federal law prohibits it, and most state guidance permits AI-assisted feedback with human review. Utah law requires the state model policy to prohibit educators from using generative AI to independently grade student work or issue high-stakes determinations. The consistent rule across every authority reviewed is that a human educator remains responsible for the final grade.

Did Washington DC ban AI grading?

No. The DC model policy places reviewing and grading student work in its limited-use tier, not its prohibited tier, and permits it with conditions: educators own all summative grading decisions, the prompt text is returned with the work, student identifying data including race and ethnicity are not disclosed to the tool, and students may appeal an AI grade and request human feedback. What DC does prohibit outright includes promotion and placement decisions, discipline, teacher evaluation, and IEP or Section 504 eligibility.

What does Utah require about AI and grading?

Utah HB 273, effective July 1, 2026, requires the state board’s model policy to prohibit educators from using generative AI to independently grade student work or issue high-stakes determinations, defined as decisions about placement, discipline, academic progression, or eligibility for a program or service. It also requires written parent notice when generative AI is used for instruction, assessment, or classroom activities.

Does a teacher pasting student work into ChatGPT violate FERPA?

It can. Graded student work maintained by the school is an education record. Disclosing personally identifiable information from it to an outside party without consent is permitted only if that party qualifies as a school official, which requires direct district control over the use and maintenance of the records and limits on redisclosure. A personal consumer account with no district agreement fails that test. Properly de-identified work falls outside FERPA, but student writing is often self-identifying.

Should students be told when AI helped grade their work?

Three of the authorities reviewed say yes in some form. Utah requires written parent notice for classroom use of generative AI, California’s model policy says students should be informed, and the DC model policy requires the prompt text to be returned with AI-graded work and gives students an appeal right. Districts adopting AI-assisted feedback should decide the disclosure question deliberately rather than by default.

Is AI grading accurate enough to rely on?

Published work on large language model scoring of student writing reports high self-consistency but noticeably lower agreement with human raters. That pattern matters for policy: a tool can produce the same score twice and still diverge from the teacher’s judgment, which is an argument for review rather than for replacement.