Most guides to AI writing feedback assume students submit their work as Google Docs or typed text. A large number of teachers don’t work that way. Students scan handwritten drafts, download assignment templates as PDFs, or submit through an LMS that converts everything into a PDF on receipt. In practice, a significant portion of writing teachers are sitting with a folder of PDF files and wondering how to get AI feedback working when the content is locked inside a formatted file.
This guide covers that specific workflow: how to handle PDF submissions, which tools actually accept PDF uploads, what to do when the PDF is scanned rather than text-based, and the prompts that produce useful feedback from whatever text you extract. It also covers the privacy considerations that apply specifically when you’re uploading student work files to an AI service because PDF submissions almost always contain student names and identifying information that text-only prompts don’t.
The Core Challenge: Getting Text Out of a PDF
AI tools don’t read PDFs the way you do. They read text. When a PDF is text-based, meaning the student typed in a Word document and saved it as a PDF, the text is extractable, and most tools handle it cleanly. When the PDF is scanned, a photo of a handwritten page, or a document that was printed and re-scanned, the text is stored as image data, and most AI tools either can’t read it or read it with significant errors.
Before anything else, identify which type of PDF you have:
- Text-based PDFs: Open the PDF and try to select a word. If you can highlight and copy it, the text is extractable.
- Scanned/image PDFs: Click on the text, and nothing selects. The page is an image.
For text-based PDFs, the workflows below work cleanly. For scanned PDFs, you need an OCR (optical character recognition) step first or a tool with built-in visual PDF reading capability.
Tool 1: Claude: The Best Free Option for PDF Upload
Claude handles PDF uploads better than most general-purpose AI tools, and its free tier is genuinely usable for this task. Claude can visually scan and extract text from PDFs, giving teachers a powerful tool for quickly reviewing student submissions. The free tier includes a 200,000-token context window large enough for most student essays, even long ones, and the ability to upload PDF files directly.
How it works in practice: Open Claude, click the paperclip or attachment icon, upload the student’s PDF essay, then type your feedback prompt. Claude reads the content and responds based on what’s in the file.
One important limitation to flag immediately: Claude has trouble with complex layouts in PDFs; multi-column academic formatting, dense footnotes, or scanned handwriting can reduce reading accuracy. Test with a sample PDF first before committing to a workflow for a whole class set.
The key prompt for PDF essay feedback:
“I’ve uploaded a student’s [ASSIGNMENT TYPE, e.g., ‘argumentative essay’] for my [SUBJECT] [GRADE LEVEL] class. The assignment was assessed on: Read the essay and draft written feedback of 120–150 words structured as one to two sentences on specific strengths, two to three sentences on the main area for improvement with a reference to a specific passage, and two actionable next steps. Do not grade the work; I will assign the grade separately. Do not use the student’s name in the feedback.”
That last instruction matters: even if the PDF contains the student’s name in a header, the feedback comment should not. This keeps the draft appropriate for your edit-and-send workflow without needing to manually strip the reference.
A variation for scanned PDFs: Claude’s visual reading capability can sometimes extract text from scanned documents if the scan quality is decent. Upload the PDF as you normally would and add this instruction before your feedback prompt: “First, confirm whether you can read the handwritten or scanned text in this document. If you can, proceed with the feedback prompt below. If not, tell me what you can and can’t read.”
That diagnostic step saves you from finding out partway through a response that Claude has been hallucinating text it couldn’t see clearly.
Pricing: Claude’s free tier requires no credit card, includes PDF uploads, and works with the 200K context window. Claude Pro is $20/month or $17/month billed annually and increases the usage limit for heavier weekly marking. The free tier is sufficient for reviewing individual essays or a small class set per sitting.
Tool 2: ChatGPT: PDF Upload With Free-Tier Limits
ChatGPT accepts PDF uploads on its free tier, but with a documented constraint that matters for classroom use: 3 file uploads per day on the free plan. If you’re working through a class set of thirty essays in one evening, that daily cap is a real problem.
For occasional PDF feedback, a few drafts at a time, or working through a class set across multiple days, the free tier works. For batch marking in a single session, Plus ($20/month) removes that restriction.
The upload workflow is identical to Claude: paperclip icon, upload PDF, type your prompt. ChatGPT reads text-based PDFs reliably. Scanned PDFs are hit-and-miss; the free-tier model’s visual document reading is less consistent than Claude’s in independent comparisons.
The prompt structure: the same rubric-aligned template above works in ChatGPT. One small modification worth adding for ChatGPT specifically, since it has a tendency toward over-positive feedback generation:
“Be honest about weaknesses in the essay rather than diplomatically hedging. Weak arguments should be called out specifically, not softened into vague ‘areas for growth.’ Students learn more from clear identification of problems than from gentle framing of everything as opportunity.”
That instruction shifts the output toward the direct, criterion-specific comments students actually act on, rather than the encouraging-but-vague style ChatGPT defaults to.
Tool 3: Gemini Notebook: Multi-PDF Research and Feedback
Gemini Notebook (formerly NotebookLM) takes a different approach that’s particularly useful when you want to process multiple PDF submissions through the same rubric in a single session. You create a notebook, upload multiple student PDFs as sources, then run feedback queries across the whole collection.
The core feature is source-grounding: every answer Gemini Notebook produces can be traced back to a specific passage in the uploaded document, which means its feedback references the actual writing rather than generalities.
A workflow for a small group:
- Create a new Gemini Notebook
- Upload three to five student PDFs as sources
- For each student’s document, ask: “Based only on [Student A]’s essay source 1 draft feedback against these criteria: [paste rubric]. Flag specific passages that support or undermine each criterion.”
Important limitation: Gemini Notebook is not designed to process student names or identifying information. Upload PDFs with identifiers removed where possible, or ask the tool to ignore the name in its output.
Free plan limits: the standard free tier allows 100 notebooks, 50 sources per notebook, and 50 chat queries per day. For a class set of thirty essays across a marking period, that capacity is genuinely sufficient.
Tool 4: Adobe Acrobat AI: Feedback Directly on the PDF
For teachers who want to annotate the PDF itself rather than produce a separate written comment, Adobe Acrobat’s AI assistant can interact with PDF content directly inside the document.
Adobe Acrobat AI Assistant answers questions about PDF content, generates summaries, and can help prepare feedback within the document environment. The practical difference: instead of copy-pasting feedback into an LMS comment box, you can add AI-informed annotations directly to the student’s PDF file before returning it.
Pricing: Adobe Acrobat AI Assistant requires a paid Adobe subscription, Acrobat Standard ($12.99/month) or Acrobat Pro ($19.99/month). A free trial is available. This is the most expensive tool in this guide, and the AI features are more focused on document interaction and summarization than on generating criterion-based writing feedback in the way that Claude or ChatGPT does. It suits teachers already working within Adobe’s ecosystem; for teachers who don’t already subscribe, the other tools in this guide produce better feedback outputs at lower cost.
Handling Scanned PDFs: The OCR Step
If a significant portion of your PDF submissions are scanned images of handwritten work, you need to convert them to readable text before any AI feedback tool can process them reliably.
Free OCR options:
- Adobe Acrobat (free tier): can OCR scanned PDFs; accuracy is good for typed text scanned at high quality, less reliable for handwriting
- Google Drive: upload a scanned PDF, right-click, and open with Google Docs. Google automatically OCRs the content and opens it as editable text; you can then copy it
- Smallpdf or ILovePDF: browser-based OCR tools, free for occasional use
The Google Drive route is the most accessible for most teachers: the conversion takes thirty seconds and produces text you can immediately paste into any AI feedback prompt. The accuracy is high for legibly handwritten or printed work; difficult handwriting will still produce errors that need checking.
After OCR, treat the text as you would any typed submission, paste it into your Claude or ChatGPT prompt alongside the rubric, and generate feedback from there.
Privacy: What You Must Check Before Uploading Student PDFs
This is the section where PDF feedback differs most significantly from text-prompt-based feedback, and it’s the section most guides skip over.
When you type a paragraph of anonymized notes about a student’s essay into a chatbot, you control exactly what information the AI sees. When you upload a PDF, the tool sees everything in that file: the student’s name in the header, the submission date, any identifying metadata, grading rubric annotations you may have started, and potentially the institutional letterhead or assignment cover sheet.
Three checks before uploading any student PDF:
1. Check your institution’s policy on cloud document upload. Many schools and universities have data governance policies that restrict what files can be uploaded to third-party cloud services. Student submissions may be covered by these policies. If you’re unsure, contact your data protection officer or IT administrator before building this into a regular workflow.
2. Use tools with documented educational data protection. Claude’s free tier, like most consumer AI tools, processes inputs through Anthropic’s servers and may use them to improve models unless you opt out — the setting is in account data controls. For student writing that contains names or identifying information, either opt out of training data use or use an educator-specific plan where data handling is contractually defined. ChatGPT for Teachers (free for verified US K-12 educators) provides a FERPA-aligned workspace with stronger data protections than the consumer ChatGPT tier.
3. When in doubt, extract and anonymize first. Before uploading, open the PDF, copy the text, paste into a blank document, remove the student’s name and any other identifying information, and use that anonymized text in your prompt instead of uploading the original file. This takes sixty seconds and removes most of the data protection risk.
A Complete Workflow From PDF Submission to Returned Feedback
Here is a practical, end-to-end sequence that a teacher can follow for a class set of PDF essays:
Step 1: Test your setup. Take one PDF from your set and run it through Claude or ChatGPT with your rubric prompt. Read what comes back. Is it criterion-aligned? Is it specific enough to the writing? Does it avoid grading? Edit the prompt until you’re satisfied with the output on that test case.
Step 2: Process in batches, not all at once. For the free tier of any tool, work through five to ten PDFs per session rather than all thirty at once. This avoids hitting daily upload limits and gives you natural review pauses.
Step 3: Read and edit each draft before saving it. Never save AI-generated feedback without reading it first. The draft may have miscited a passage, produced a generic comment that doesn’t reference the specific essay, or missed the rubric criterion most relevant to this student’s main problem. Each edit takes thirty to ninety seconds; that editing is where your professional judgment enters the feedback.
Step 4: Personalize the opening sentence. The most effective addition you can make to any AI draft: write the first sentence yourself, referencing something specific about the student’s choice of argument or their particular strengths on this piece. The AI draft handles the criterion coverage; your sentence handles the relational element that students remember.
Step 5: Return as typed comment, not as annotated PDF. Unless your LMS specifically supports PDF annotation return, paste the edited feedback into the LMS comment field as plain text. This makes it searchable, screen-reader accessible, and easier for students to copy into their own revision notes.
What to Do When the AI Gets It Wrong
Claude and ChatGPT will occasionally misread a PDF citing a passage that isn’t where they placed it in the text, summarizing an argument slightly differently from what the student wrote, or generating a generic comment that doesn’t actually address the specific essay. This happens more often with complex PDF layouts and scanned documents.
When you notice the feedback misrepresents the work, don’t edit the AI draft start over with a more explicit prompt:
“The previous feedback was too generic. Read the essay again and this time, quote the exact sentence from the student’s writing that most clearly illustrates their main weakness. Use that direct quote in the feedback comment, then explain what the weakness is and what the student should do differently.”
Requiring a direct quote forces the tool to actually engage with the specific text rather than producing a comment that could apply to any essay on the topic.
The Takeaway
Using AI to give writing feedback on PDF submissions adds one practical layer compared to text-based feedback: getting the content out of the file and into a format the AI can read. For text-based PDFs, Claude’s free tier handles this directly and cleanly. For scanned PDFs, the Google Drive OCR route converts the file in under a minute. Once the text is accessible, the same rubric-aligned prompts that produce useful feedback in any context work here too. Check what data your chosen tool sees before uploading student files, opt out of training data use where available, and edit every draft before it reaches a student. The PDF format changes the extraction step; everything else about producing good feedback stays the same.
Suggested Internal Links:
- Article: How to Use AI to Give Students Writing Feedback “the broader guide to AI writing feedback for teachers”
- Article: How to Use AI to Give Students Writing Feedback Online “AI feedback workflows for online course submissions”
- Article: Best AI Tools for Teachers in 2026 “the best AI tools for teachers this year”
- Article: How to Write Better Prompts for AI “writing better prompts to get more accurate feedback drafts”
External Sources:
- Anthropic — Claude file upload documentation and free tier details
- OpenAI — ChatGPT free plan file upload limits
- Google — NotebookLM (Gemini Notebook) source limits and free tier documentation
- Adobe — Acrobat AI Assistant
- blog.aieducator.tools — “Faster Feedback on Student Essays: 6 AI Tools for Teachers” (October 2026) — Claude PDF handling comparison
- frizzle.com — “AI Feedback Generator for Teachers: Practical Guide” — anonymization and privacy workflow
