ManavAI vs. human writing is unsolvable. Attribution is.
Previously framed as "detect AI-generated work via keystroke entropy." Honest reframe: detecting AI-generated content reliably is unsolved (OpenAI withdrew their own classifier). Manav replaces detection with attribution. The worker signs an origin declaration of what tools were used. The signature carries the legal weight of a sworn statement.
- Worker chooses what to disclose. No content analysis. No keyboard or screen monitoring.
- Optional process receipts cover hashes of in-progress state, not content. Privacy-clean.
- AI-content watermarks (SynthID and successors) are verified at submission time, not by surveilling creation.
The bad framing, AI-output detection
"We detect AI-generated work by measuring keystroke entropy and writing cadence." Requires continuous typing surveillance. Doesn't work reliably, OpenAI withdrew their own classifier as inaccurate. Any vendor who claims to detect AI-generated text reliably is either lying or about to be lying.
The privacy-preserving framing, required attestation
Replace detection with accountability. At submission, the worker signs a declaration of what tools were used and what's original. The signature carries the legal weight of a sworn statement (perjury, fraud). The worker chooses what to disclose; nothing about their keyboard or screen is monitored.
An example attestation
Board memo, Q2 strategy review
memo_q2_2026Tool attestation (signed by author):
{
"declaredTools": [
{ "name": "Claude 4.5", "use": "copy-edit only" },
{ "name": "Notion AI", "use": "none" }
],
"originalAnalysis": true,
"watermarksDetected": [],
"sworn": true
}
✓ Signed and attached to memo
Optionally, the worker can opt into process receipts, signed checkpoints covering only hashes of in-progress state. The chain proves "I made N checkpoints over D hours" without revealing the content or surveilling the writing.
Want to ship this in your own app?
Three lines of JavaScript. Demo key mnav_test_demo works on localhost.