Should I upload a whole file to AI or just an excerpt?
File upload feels convenient: drop a PDF, attach a DOCX, add a screenshot, and ask the AI to summarize or rewrite it. The privacy problem is that the full file brings everything with it, not only the part you care about.
A Word document may include comments and author properties. A PDF may include extra pages and document metadata. A screenshot may show browser tabs and account names. A spreadsheet export may include hidden columns or unrelated rows. If the task only needs one excerpt, the full file is unnecessary exposure. That exposure is routine: Cyberhaven found that source code and client data are among the most common confidential data types employees paste into AI tools — exactly the material a narrow excerpt can leave behind.
Start with the task
Ask what the AI is actually doing:
- Rewriting one paragraph.
- Explaining one contract clause.
- Summarizing one section.
- Turning one error log into troubleshooting steps.
- Drafting a reply from a few ticket facts.
- Describing a single chart or screenshot crop.
If the task is narrow, create a narrow input. Copy only the relevant section, then sanitize it before pasting.
Build a sanitized excerpt
A useful excerpt should keep the structure and remove the identity layer:
- Copy the smallest relevant section.
- Replace names, emails, IDs, addresses, and internal links with placeholders.
- Add only the context the AI needs: audience, goal, tone, constraints, and output format.
- Avoid unrelated history, side comments, and attachments.
- Review the AI output before reuse.
For text excerpts, AI Prompt Privacy Checker automatically detects common sensitive strings, lets you review or restore replacements, and lets you manually label anything missed before the AI sees it.
Use screenshots and PDFs carefully
Sometimes an excerpt is visual. If the AI needs to describe a dashboard chart, crop the chart area and remove surrounding private UI. If it needs to review a PDF clause, extract that page or section rather than uploading the whole document. If the PDF page still includes names, signatures, or account numbers, use PDF Redactor first.
For screenshots, use Screenshot Redactor to black out visible identifiers. If the source image may carry hidden metadata, inspect it with Metadata Inspector.
When a full file is justified
A full file may be necessary when page layout, cross-document context, table structure, or visible formatting is the task. For example, asking an AI to identify inconsistencies across a multi-page policy may require more than an excerpt.
When the full file is justified, clean it first. Remove unnecessary pages, strip metadata, redact visible identifiers, and verify the output — the check files before uploading to AI guide walks through that review. The rule is not “never upload files.” The rule is “do not upload the parts the task does not need.”
Excerpts improve both privacy and prompt quality
Narrow inputs are often better prompts. They reduce noise, focus the model, and make the answer easier to review. They also create a cleaner record of what was actually shared.
Before using a file upload, try asking: can I do this with a sanitized excerpt instead?