Draw the boundary clearly
An offline redaction workflow can reduce the amount of information disclosed during preparation because the source file does not need to be sent to another scanning service. It can inspect and transform a local copy before any AI upload occurs.
The word offline ends at the upload button. When you send the prepared copy to ChatGPT, Claude, or Gemini, the selected data leaves the device. The AI provider's current terms, workspace controls, retention, and any connected third party then apply.
This distinction matters in product claims. A tool may accurately say that its preflight processing is local while it would be inaccurate to suggest that the subsequent cloud AI interaction remains local.
How to evaluate a local redaction tool
| Question | Evidence to look for |
|---|---|
| What stays local? | A specific statement covering file bytes, extracted text, OCR, detected secrets, and derived data |
| Does it require an account? | Whether sign-in or cloud activation is needed for file processing |
| Can the claim be tested? | A zero-network test, firewall test, reproducible documentation, or independent review |
| What formats are supported? | A matrix of detected, removable, unsupported, and manual-review items by format |
| What happens to the original? | Copy-first processing, clear output path, and failure behaviour |
| Is output verified? | Reopen, readability check, second scan, and a report of remaining findings |
Do not infer offline behaviour from an installer, a desktop window, or the absence of a browser. Licensing, telemetry, OCR, threat scanning, or AI-assisted detection can all create network traffic. The documentation should separate content processing from optional update or licensing traffic.
A six-step local-to-cloud workflow
- Duplicate. Preserve the original and prepare a working copy under the same access controls.
- Minimise. Remove pages, rows, fields, attachments, and precision not needed for the AI task.
- Inspect locally. Scan supported visible content, hidden structures, metadata, and credentials.
- Transform. Apply true redaction, consistent placeholders, or generalisation appropriate to the format.
- Reopen and review. Verify the exact clean copy and document unresolved findings.
- Approve the destination. Check the AI service, plan, settings, retention, administrator access, and organisation policy.
Where possible, test the prompt with synthetic data first. A synthetic example can reveal whether the task design works before any real company information is considered.
ChatGPT, Claude, and Gemini are not one policy
Each provider offers multiple consumer, business, enterprise, and API products. Do not write a policy for the brand as a whole. OpenAI's API documentation, for example, states that API data is not used for training by default but also describes abuse-monitoring retention and endpoint-specific application state. That does not establish the policy for a personal ChatGPT workspace.
Anthropic documents separate rules for consumer Claude and commercial products, including training and retention. Google likewise distinguishes Gemini Apps from Gemini features in Workspace. Confirm the current documentation for the exact account and feature on the day of use.
Questions to record include: Is content used for model improvement? How is it retained? Can an administrator access it? Does a connector send it to another service? Is deletion supported? Is the use covered by the organisation's contract and policy?
When local preprocessing is not enough
Some documents should not be uploaded even after redaction. Examples can include material whose context remains identifying, information covered by strict contractual restrictions, active credentials, highly sensitive legal or security records, or data whose transformation would destroy the task's meaning.
Alternatives include using a synthetic dataset, asking a general question without the document, running an approved model in a controlled environment, or conducting the work manually. Preflight is a decision point, not an instruction to upload.
Official sources
This guide uses primary sources available on August 19, 2026. Product policies and software features can change, so confirm current terms before handling sensitive material.
- Your data: default usage policies by API endpoint — OpenAI Developers
- How long do you store my organization's data? — Anthropic Privacy Center
- Is my organization's data used for model training? — Anthropic Privacy Center
- Gemini Apps Privacy Hub — Google Help
- Artificial intelligence for small business — Australian Cyber Security Centre