How Teachers Are Using Claude to Grade Papers in 2026 (Without Crossing the FERPA Line)

Grading is the part of teaching that eats your weekends. A stack of thirty essays at twelve minutes each is six hours, and the last ten papers always get less attention than the first ten because you are tired. AI tools, used well, can give you back two or three of those hours and improve the quality of feedback on the last papers in the stack.

Used badly, the same tools can put you on the wrong side of FERPA, your district's data policy, or your students' trust. I want to walk through the workflow I have built and tested with classroom teachers — a workflow that saves real time without crossing the lines that matter.

Why teachers are reaching for AI grading

The honest answer is exhaustion. The average U.S. teacher spends about nine hours a week on grading and feedback outside the school day. Most of that time is not the part teachers care about. The part teachers care about is the substantive comments — "your thesis is strong but paragraph three loses focus" — not the mechanics — "comma splice line 6, run-on line 14."

Claude is excellent at the mechanics and quite good at first-draft substantive feedback. The right division of labor is: Claude drafts, you decide. That division gives you most of the time savings while keeping the teacher's professional judgment as the source of truth.

The FERPA basics, in plain language

FERPA is the Family Educational Rights and Privacy Act. It is the federal law that protects student education records. The short version: you may not share a student's identifiable education record with a third party without parental consent (or, for college students over 18, the student's consent), with a few narrow exceptions.

Anthropic, the company that makes Claude, is a third party. When you upload a student paper to Claude.ai, you are sharing that record with Anthropic. Whether that sharing crosses a FERPA line depends on two things: whether the record is identifiable, and what data agreement you have with Anthropic.

If you remove the student's name and any other identifiers, the record is no longer identifiable, and FERPA's restrictions do not apply in the same way. If you also use a Claude plan with a written no-training-on-your-data agreement, you have closed the second risk. Together, those two steps make the workflow defensible.

Four rules I never break

Here is the safety floor. If you cannot meet all four, stop and rethink.

Rule one: never upload identifiable student work to a personal-tier AI account. Personal accounts (Claude Pro at $20/month, ChatGPT Plus at $20/month, Gemini AI Pro at $20/month) do not give you the written data agreements your district expects. Use Claude Team ($30/seat/month) or Claude Enterprise. Get your principal or IT director to sign the data processing agreement. Keep a copy.

Rule two: anonymize before upload, always. Strip the student's name from the document header, the file name, and the body of the essay. Strip ID numbers. Strip any "About me" or "My family" details that could identify the writer. Replace with "Student A," "Student B," etc.

Rule three: the teacher is the grader of record. Claude proposes a score and feedback. You read both, edit, and decide the final grade. If a parent challenges the grade, your professional rationale is the answer — not "the AI said so."

Rule four: be transparent with students and parents. Tell your class, in writing, that you use AI to draft feedback on their work. Most students appreciate the honesty and the better feedback. The few who object can have their work graded the old-fashioned way. This transparency is the difference between a tool and a betrayal.

Step-by-step Claude Project setup

The whole setup takes about twenty minutes and you do it once per class.

1. Sign up for Claude Team. Go to claude.ai/team. Pick the smallest seat count you need. Get your district's data processing agreement signed before you upload anything.
2. Create a new Project. In the Claude.ai sidebar, click "Projects," then "Create Project." Name it after your class — "English 9 Period 3, Spring 2026." This becomes the workspace that holds your rubric and exemplars.
3. Write the Project instructions. Paste in: "You are a thoughtful 9th-grade English teacher. You give specific, kind, rubric-anchored feedback. You never invent quotations or claim a student wrote something they did not write. You always identify three strengths before any area for improvement. You write at an 8th-grade reading level so students can understand you."
4. Upload the rubric. Add your full rubric as a PDF or text file. The more concrete the rubric (with examples of what a 4 looks like vs a 3), the better Claude's feedback will be.
5. Upload three anonymized exemplars. Pick a strong, average, and weak paper from a previous semester. Strip names. These calibrate Claude to your standards.

That is the Project. Every chat you start inside it now knows your rubric, your exemplars, and your voice. You only set this up once per class per term.

The grading workflow, paper by paper

Here is how I run the actual grading session. Block out ninety minutes for a stack of thirty papers — about a third of what the same stack used to take.

1. Anonymize the stack. If the papers are in Google Classroom or Canvas, download them as a folder. Use a quick rename script (or a built-in classroom tool) to replace student names with "Student A, B, C..." Keep a private legend of which letter is which student.
2. Open a new chat in your Claude Project. Paste or upload one essay. Use the prompt below.
3. Read Claude's draft feedback. Don't just accept. Read it. Edit anything that does not match your judgment. Add a personal sentence at the end that only you can write — it makes the difference between feedback that feels generic and feedback that feels like a teacher.
4. Assign the final grade yourself. Claude can suggest "this looks like a 3 on your rubric." You decide if it is a 3, a 3 with a plus, or a 3 with a minus. The professional judgment stays with you.
5. Paste the final feedback into your gradebook. Re-attach the student's name only when you put the comment in the LMS. Keep AI and identity separate at every step.

Three rubric prompts that work

These are the three prompts I use most often. Adapt the language to your class. The structure is what matters.

Prompt 1: First-draft feedback on an essay. "I am about to paste an anonymous student essay. Using the rubric in this Project, give me: (a) three specific strengths with line references, (b) two areas for improvement with concrete suggestions, (c) the rubric score you think this paper earns and why, and (d) one sentence I could read aloud to the student that would encourage them to revise. Keep the whole response under 250 words."

Prompt 2: Comparing two drafts. "I am pasting Draft 1 and Draft 2 of the same anonymous student. Tell me: (a) which two areas the student improved between drafts, (b) one area that did not improve, (c) whether the rubric score changed and by how much, and (d) a one-sentence note acknowledging the student's effort."

Prompt 3: Spotting AI use without accusing. "I am pasting an anonymous student essay. Without making a definitive accusation, tell me which paragraphs sound substantively different in voice, vocabulary, or sentence structure from the rest of the essay. List your evidence in plain terms. I will make my own judgment."

Common questions

What about ChatGPT or Gemini for the same workflow? Both can do this. ChatGPT Plus is on similar ground; for FERPA you would want ChatGPT Team or Enterprise. Gemini in Google Workspace, with the right business plan, is also defensible. I default to Claude because the writing feedback quality is, in my testing, the most teacher-like. See my full three-way workspace comparison for the trade-offs.

Can I use Claude to draft my own assignments and rubrics? Yes. That is FERPA-clean because no student data is involved. It is also one of the highest-leverage uses of AI for a teacher. A well-designed rubric pays back across an entire term.

Does Claude flag AI-generated student writing reliably? No. Neither Claude nor any AI detector is reliable enough to be the sole basis of an academic-integrity decision. Use the third prompt above as a hint to investigate, never as a verdict. The conversation with the student is the verdict.

What happens if my district bans AI use for grading? Comply with the ban while you advocate for a clearer policy. Most district bans were written before tools like Claude Team existed. A policy that allows AI use with anonymization, business-tier accounts, and teacher-of-record requirements is reasonable and defensible. Bring this article to the policy meeting.