Before Uploading a Document to AI, Check These Privacy Settings
Before you upload a tax return, medical report or workplace spreadsheet to an AI assistant, pause at the attachment button. A document can reveal much more than the question you want answered: account numbers, signatures, diagnoses, customer records and details about other people. Even a seemingly routine PDF may contain comments, hidden text or identifying information you cannot see on the first page.
In 2026, the useful question is not simply whether an AI service has a privacy setting. It is whether the exact product, account and feature you are using are appropriate for that particular file. Training controls, storage rules and sharing permissions address different risks. Check each separately, then decide whether a smaller excerpt, a properly redacted copy or no upload at all is the safer choice.
1. Decide what the AI actually needs to see
Start with the task, not the file. If you want an explanation of a mortgage term, you probably do not need to upload the entire loan package. If you want help drafting questions for a doctor, a short description may be enough. Share the minimum information needed, and remember that typing or pasting sensitive details into a prompt still discloses them to the service.
Read every page and worksheet before choosing what to send. Consider both direct identifiers and details that become identifying in combination. A rare diagnosis, a small employer and an appointment date may point to one person without naming them. Documents involving children, legal disputes, identity verification or someone else's private information deserve extra caution. If you do not have authority to share the information, stop rather than trying to solve that problem with a setting.
- Remove Social Security numbers, full account numbers, passwords, access codes and identity-document images from anything you plan to submit.
- Look for names, addresses, signatures, birth dates, patient identifiers and details about relatives, customers or coworkers.
- Ask whether invented sample data or a brief, nonidentifying excerpt would let the assistant perform the same task.
2. Verify the product and account, not just the brand
A familiar AI brand may offer consumer chats, business workspaces, developer services and integrations inside other software. Those offerings can have different contractual protections, defaults and administrative controls. Paying for an individual subscription does not, by itself, mean you have an enterprise agreement or permission to upload confidential records. Confirm which account is active, especially if personal and work profiles share a browser.
Open the settings and the policy pages linked from that exact service. Look for terms covering uploaded files, prompts, generated answers and connected data sources. If you access a model through another company's app, that app's handling of your document matters too. A model provider's policy alone does not explain what the app stores, logs or shares.
Check whether an employer or school manages the workspace and what administrators can access or retain. For work material, obtain approval through the organization's process rather than assuming the availability of an upload button means permission. If the policy is ambiguous, ask the administrator or provider a specific question about your document type and feature before proceeding. Save the relevant policy link and check its effective date; an older help article may describe a different product or plan.
3. Separate training controls from confidentiality
Look for controls labeled along the lines of model improvement, training or data use. Read the explanation rather than relying on the switch's name. Determine whether the control covers future conversations only, whether file attachments are included and whether it applies across your account or just one workspace. Do not assume changing it later reverses earlier use.
Opting out of training is meaningful, but it is not a promise of zero storage or zero access. Depending on the service and agreement, content may still be processed and retained for operations, security, abuse prevention or legal obligations. Limited human review may also be allowed. These are separate questions that need separate answers in the documentation.
Check exceptions before sending feedback on a sensitive conversation. Some services explain that submitting ratings or feedback can authorize additional use of associated content. Temporary or private chat modes may offer different treatment, but their labels are not guarantees that nothing persists. Verify the stated retention period and feature limitations each time the stakes are high. Also distinguish training on your content from processing it to generate the answer you requested; turning off the former does not eliminate the latter.
| Control or claim | What it addresses | What it does not establish |
|---|---|---|
| Training opt-out | Use for model improvement within its stated scope | No storage or human access |
| Temporary chat | History and retention under that mode's rules | Instant deletion everywhere |
| Paid account | Features and terms of the purchased plan | Approval for regulated or work data |
| Encryption | Protection in the specified storage or transfer context | A ban on provider processing |
4. Find out what deletion actually removes
Retention is about how long information remains; deletion is about what action removes it and from where. Locate the rules for the conversation, the uploaded file and any separate file library, project or knowledge collection. Removing a chat may not remove a document stored elsewhere. Likewise, deleting a source file from your computer does not delete the copy already sent to a service.
Look for a stated deletion timeline and any exceptions for backups, security investigations or legal requirements. Avoid treating a disappearing item in the interface as proof that all copies are gone. If the provider does not clearly explain attachment retention, that uncertainty belongs in your decision about whether to upload, not on a to-do list afterward.
Connected storage adds another layer. An assistant may retrieve files from a cloud drive or maintain an index that helps it search them. Disconnecting the drive may stop future access without necessarily erasing material already copied or indexed. Before connecting an account, check the permission scope, synchronization behavior and documented process for deleting imported content. If access can be limited to one folder or selected files, prefer that over granting access to an entire personal or company drive.
| Action | What to verify |
|---|---|
| Delete a conversation | Whether its attachments are also scheduled for deletion |
| Remove a project file | Whether other conversations or collections retain copies |
| Disconnect cloud storage | Whether previously imported data or indexes remain |
| Close the account | Deletion timing, exceptions and export options |
5. Check who can open, copy or reshare the result
Uploading privately and sharing privately are different operations. Before creating a conversation link, inspect its audience setting. Depending on the product, a link may be restricted to invited people, available to a workspace or accessible to anyone who obtains it. Do not assume a link is protected merely because it is difficult to guess or does not appear in search results.
Examine what the recipient will actually receive. A shared answer might reproduce sensitive passages even when the original attachment is not downloadable. A shared project, folder or custom assistant may expose files or information drawn from them to additional users. Preview the shared view when possible, and verify whether the link shows a snapshot or later updates. Revoking access can stop future opening through that link, but it cannot retrieve screenshots, downloads or copied text.
Review invitations and links again when a project ends or someone's role changes. A sharing choice that was appropriate for one meeting should not quietly become permanent access.
- Prefer named recipients and the narrowest available permission over an unrestricted link.
- Check whether collaborators can invite others, edit the project or access its source files.
- Remove unnecessary links and collaborators, and avoid including sensitive information in shared titles.
6. Redact the file, then test the copy
Redaction must remove information, not simply cover it visually. Drawing a black rectangle over text, changing the font color or cropping an image inside a document can leave the underlying content recoverable. PDFs can contain searchable text behind scanned pages. Office files can retain comments, tracked changes, hidden sheets, speaker notes and metadata that reveal information outside the visible page.
Work on a duplicate using a tool with an actual redaction feature, apply the redactions and save the sanitized version. Follow the tool's instructions for removing hidden information. Then reopen the output and try searching for and copying the details you intended to remove. Inspect all pages, attachments and file properties. Those checks help catch mistakes, but they are not a guarantee that every hidden element is gone.
For a simple question, a newly created plain-text excerpt containing only necessary facts may be easier to inspect than a complex original. Use neutral placeholders consistently. Remember that removing names does not necessarily anonymize a narrative: dates, locations and unusual circumstances can still identify someone. For highly sensitive material, have a qualified person check the redaction rather than trusting its appearance or asking an online AI service to sanitize the original.
7. Set a higher bar for financial, medical and work records
A privacy toggle cannot make an unsuitable service suitable. A tool that promises not to train on uploads may still lack the contract, access restrictions, retention controls or organizational approval required for a particular document. Financial records can enable fraud; medical records can expose deeply personal information; workplace files may be governed by confidentiality agreements and security policies.
For US health information, do not assume a consumer AI app is covered by HIPAA simply because you upload a medical record. HIPAA obligations depend on the entities and relationships involved. A healthcare organization's use may require a business associate agreement and other safeguards; an individual's consumer app use is not automatically equivalent. Ask the healthcare organization about approved options when unsure.
For work documents, follow employer rules even when you personally pay for the tool. For sensitive financial or medical tasks, consider an approved portal, a qualified professional or a genuinely local tool whose network access, telemetry and storage you have verified. The safest final check is straightforward: if you cannot explain what leaves your device, who can access it and when it is deleted, do not upload the original. Use a nonsensitive example or ask the question without the file.
Check the service's current terms and your organization's requirements before uploading sensitive information; seek qualified privacy, legal or security advice when necessary.
AI Assistance Used — AI assistance was used in researching and drafting this article. It was reviewed, edited, and fact-checked by Jordan Singh before publication.