Redact a PDF Before Uploading It to ChatGPT or Claude: A Mac Checklist
The useful question to ask an AI assistant rarely needs your client’s name, your bank account number, or your date of birth. The model needs the clause, the numbers in the table, or the structure of the argument. It does not need the identifiers wrapped around them, yet a PDF carries all of it in more places than the visible page.
This guide covers the pre-upload pass on a Mac: what to check before you send anything, which fields to remove or replace, why a black rectangle is not redaction, and how to verify the redacted copy locally before uploading it.
First, decide whether the document may go at all
Check three quick gates before opening any redaction tool:
- Permission to share. A client file, an employer’s internal record, or an agreement under NDA may not be yours to upload to a consumer AI service at any level of redaction. Redaction reduces exposure; it does not create permission.
- Account tier. Business and enterprise plans generally exclude your content from model training by default, while consumer plans generally require you to manage a privacy toggle or use a temporary session. The same file carries a different risk on a personal account than on a company workspace.
- Minimum data for the task. Summarizing a lease needs the termination clause, not the tenant’s home address. Explaining a bank statement needs the transactions, not the full account number printed in every header. Write down the one job the AI has to do; everything outside it is a removal candidate.
Check the service’s data settings before uploading
Data settings are the second line of defense, not the first. They govern what the provider may do with what it receives, and they cannot un-send a file. As of November 2026 (verify on the provider’s own pages on the day you upload):
| Service | What its documentation says |
|---|---|
| ChatGPT personal plans | Settings, then Data Controls, then “Improve the model for everyone” governs whether new chats and uploaded files may be used for model training. Turning it off applies going forward and does not delete existing chats or files. |
| ChatGPT Temporary Chat | Temporary chats stay out of your chat history and are not used to improve models while temporary. OpenAI notes it may keep a copy for up to 30 days for safety purposes. |
| ChatGPT Library files | Deleting a chat removes it from your view immediately and schedules deletion within 30 days (subject to legal and security exceptions), but it does not delete a file saved separately in Library or Space. Remove the file there too. |
| Claude Free, Pro, and Max | Settings, then Privacy governs whether chats and coding sessions may be used for model improvement. Leaving it on extends retention of de-identified training data to up to 5 years; turning it off stops future training, and a deleted chat leaves backend storage within 30 days. |
| Claude Incognito chats | Incognito chats are excluded from model improvement even if the account toggle is enabled. Commercial tiers such as Team, Enterprise, and API data are not used for training by default. |
Settings only control what happens after upload. The rest of this checklist controls what the upload actually contains.
Checklist: what to remove and what to keep
Work through the PDF in this order. “Remove” means deleting the underlying text from the file you upload, not covering it with a shape drawn on top.
| Field | Default action | Why and exceptions |
|---|---|---|
| Full names of private individuals | Remove or replace with a role (“Tenant”, “Consultant”) | Keep only if the task requires that name and you are permitted to share it |
| Email addresses and phone numbers | Remove | Almost never needed for analysis or summarization |
| Postal addresses | Remove; keep city or state if jurisdiction matters | Governing-law city is usually enough for a contract review |
| Government IDs (SSN, passport, national ID) | Remove, always | No consumer AI task needs the real identifier |
| Bank account, IBAN, card numbers | Remove; keep last 4 digits only for reconciliation | Full numbers often repeat in statement headers and footers on every page |
| Dates of birth | Remove; keep age or birth year only if relevant | Date of birth paired with a name is a classic identity pair |
| Signatures (images and typed) | Remove | A signature image is reusable far beyond the current task |
| Account, customer, or case reference numbers | Replace with a consistent placeholder | Keep structure (“Invoice 1042”) so the model’s answer stays easy to map back |
| Salaries, balances, line amounts | Keep if the task is financial; otherwise generalize | The model often needs the numbers, not whose they are |
| Document metadata (Title, Author, company) | Strip | PDFs store an Info dictionary and XMP block that never appear on the page |
| Comments, annotations, form values, attachments | Strip or review one by one | Review comments often hold the most candid internal notes in the file |
| Hidden text, off-page text, covered layers | Remove | Text under a drawn “redaction” box or outside a cropped margin stays in the file |
| Company letterhead or generic boilerplate | Usually keep | Low sensitivity and often useful context, unless the client name itself is confidential |
Two habits speed up review: use consistent placeholders so the model can still follow who owes what to whom, and keep the placeholder key offline on your Mac rather than uploading it alongside the PDF. For a deeper look at non-visible PDF layers, see how to remove metadata and hidden data from a PDF on Mac.
A black rectangle is not redaction
Drawing a black box over text in most PDF editors adds a shape on top of the page. The original text stays in the PDF content stream: selectable, searchable, and copyable. Real redaction deletes the underlying text and then covers the space.
macOS Preview includes both tools in its Markup toolbar: a rectangle shape (visual only, not redaction) and a dedicated Redact tool (Tools > Redact, which draws a black box with gray diagonal cross-lines and permanently deletes the covered text when you close the file). Preview’s Redact tool works well for a few marks on a short PDF, though it overwrites the file you opened in place unless you duplicate it first. See how to redact a PDF on Mac without Adobe for a side-by-side comparison.
Whatever tool you use, run the same test after saving: open the exported copy, press Cmd+A and copy, or press Cmd+F and search for a value you removed. If it matches, it was never removed.
A verified offline workflow on Mac with Inkfish
Inkfish processes PDFs locally on macOS 14 or later, with no network access and no account, and writes a separate -redacted.pdf copy while leaving your original untouched.

- Open the PDF and run detection. Inkfish scans each page with regional rule packs (United States, EU, UK, and mainland China, plus finance and healthcare reference packs) and an on-device model that catches names, addresses, and organizations the rules miss. Each finding is labeled with its source so you can judge it in context.
- Review and add what is missing. Tick what should go, untick anything the AI task needs, and use Mark area or Add rule for repeated project codes or client names across pages. For image-only scans, Inkfish runs Apple’s on-device text recognition in English, Chinese, and German; check recognized words against the page image, since OCR can misread small or skewed text (see redacting a scanned PDF on Mac).
- Strip invisible layers in the same pass. Enable cleanup options for document metadata, hidden text and layers, comments, form values, and embedded attachments so nothing leaks outside the visible page.
- Save and verify the copy. Inkfish saves the
-redacted.pdfcopy and reopens it to confirm no extractable text remains under any cover. If a page fails verification, you can redo that page as a flattened image. Run your own search for one removed name and one removed number before uploading. - Upload the copy and clean up afterward. Ask your question against the safe copy, then delete the chat when you finish, including any copy saved in ChatGPT’s Library.
The free download covers single-file redaction end to end, including the 1.2 on-device detection, verification, and metadata cleanup. Inkfish Pro ($9.99 USD one-time) adds batch processing and batch summary reports.
What redaction does not do
Removing identifiers is not a guarantee of anonymity. A document can still identify someone through a rare job title, a unique project name, or a combination of dates and locations. Read the saved copy once with a stranger’s eyes before uploading. And if gate 1 rules out sharing the file externally, no amount of redaction makes the upload permitted.
Frequently asked questions
Is it safe to upload a PDF to ChatGPT or Claude after redacting it?
Redaction removes the identifiers you select from the file itself, which is the part under your control. Overall safety still depends on your permission to share the document, your account tier, and the provider’s current data settings. Redact a copy locally first, verify the output, and use a business plan for client work.
What should I remove from a PDF before uploading it to an AI?
Remove names, email addresses, phone numbers, street addresses, government IDs, bank and card numbers, dates of birth, signatures, and case reference numbers, along with invisible layers such as metadata, comments, form values, and attachments. Keep only the clauses, figures, or role labels needed for your prompt.
Does drawing a black box over text redact a PDF?
No. A drawn shape covers text visually while leaving it selectable and searchable in the PDF stream. Real redaction deletes the underlying text. Always test the saved file with Cmd+F for a removed value before uploading.
Do ChatGPT and Claude train on files I upload?
Commercial workspace plans (such as ChatGPT Business/Enterprise and Claude Team/Enterprise) exclude your data from training by default. On consumer plans, training depends on your account settings (“Improve the model for everyone” and Temporary Chat in ChatGPT; the model improvement toggle and Incognito chats in Claude). Check the provider’s current policy on the day you upload.
Can I redact a scanned PDF before uploading it to AI?
Yes, using on-device OCR. Inkfish reads scanned pages locally in English, Chinese, and German so you can review and remove detected text or mark areas by hand. Because OCR can misread faint or skewed print, always visually inspect scanned pages after saving.
Why redact on my Mac instead of using an online PDF redactor?
An online redactor requires uploading the unredacted original, the exact file you want to keep private, to a remote server before redaction happens. An offline Mac app creates the safe copy locally so only the reviewed file ever leaves your machine.
If you need to share a physical printout where a recipient still needs to verify a number, a print watermark is a separate workflow.
