← Back to CAPrivAI
AI Transparency Notice
A plain-language, honest account of exactly where your text goes when you use CAPrivAI, and what has to happen for any of it to reach a third party.
Default: detection runs on our servers, nothing goes to a third party
When you submit text or a document, detection (regex pattern matching, checksum validation, document structure analysis, and entity resolution) runs on CAPrivAI's own servers over an encrypted (HTTPS/TLS) connection. This is not on-device processing — your text is transmitted to our backend to be analyzed — but it is not sent to any third-party AI provider. The raw content is processed in memory for the duration of the request and is not written to disk, logged, or retained afterward. No external network call is made for a standard analysis request unless you explicitly opt in to one of the layers below.
Optional layer: on-server AI model (no third party)
To catch free-text entities (mainly names) across the languages CAPrivAI supports, we run a small open-source AI model directly on our own servers, in the same process as the rest of detection. This is still not a call to an outside company — it's our infrastructure running the model — but it is more server compute than the regex pass alone.
Optional layers that do leave our servers
Two features send data to third-party AI providers, and both are off by default. They only run if your consent settings explicitly allow external AI processing (Settings → Privacy & Consent):
- Hugging Face Inference API — used for language/category coverage our on-server model does not handle. Before anything is sent, everything already detected is replaced with category labels (e.g., a found email address becomes
[EMAIL]) — only the leftover, not-yet-classified text is sent. - OpenAI (GPT-4o-mini) — used for (a) a secondary pass that flags entities regex/NER may have missed, using the same sanitize-before-send approach, and (b) generating natural-language risk summaries and "safe prompt" rewrites, which operate on your already-masked text, not the original.
If your consent setting is off — the default for every new account — neither call fires. If we cannot confirm your consent for any reason, we treat that as "no" and skip the external call rather than sending your data anyway.
Immediately before an external AI call, CAPrivAI runs an independent AI Validation Gate over the exact outbound payload. If the gate detects remaining sensitive content or the consent state is not valid, the request is blocked and only metadata about the block is audited.
Maximum privacy / local-only mode
Where the product is configured for maximum privacy or local-only behavior, external AI calls are blocked regardless of any lower-level provider setting. This mode takes priority over optional enhancement features.
What is never sent anywhere, under any setting
- Raw document/text content is never stored in our correction or learning-rule data — only a derived pattern shape, never the literal value.
- Detected sensitive values themselves (card numbers, passwords, government IDs, etc.) are never included in analytics, history, or admin audit logs — those store category/action metadata only.
- Masked output is not retained on our servers beyond the request unless you have opted into the History feature, which stores metadata about the analysis, not the content.
- Raw originals are not used to train CAPrivAI models or third-party AI provider models by CAPrivAI.
In short
With no opt-ins, your text is analyzed on CAPrivAI's own servers and never reaches a third party. If you turn on external AI features, a sanitized residual sample — already-known sensitive values replaced with labels — may be sent to Hugging Face and/or OpenAI to catch anything regex and our on-server model missed, or to generate a summary from your already-masked text. Third-party retention terms for those providers vary by provider and are summarized per platform in the AI Platforms directory inside the app.
This notice describes current system behavior as implemented. For how we handle your account data, see our Privacy Policy.