Method & limits
What this can do, and what it cannot
The honest headline
No tool reliably tells you whether a person or a machine wrote a piece of text. Detectors are probabilistic classifiers with measurable false positives (real human writing flagged) and false negatives (machine text missed). Anyone who promises certainty, or a "zero-error" result, is not telling the truth. Sift never does that.
Two very different things happen here
The local signature (always available)
Five cheap, on-device statistical readings: sentence rhythm, vocabulary range, repetition, stock phrase density, and character predictability. These describe how the text feels. They are an experimental texture profile, not a detector, and they carry no authority to say who wrote it. This runs entirely in your browser and never leaves your device.
The connected detector (optional)
If an owner runs the server version with a real detection API configured, Sift calls that provider and shows its verdict with provenance: which provider, which model, how long it took. Sift sends your text to that third party when you choose to run it, and reports what the provider returned without dressing it up as proof.
What the local signature really measures
Privacy, in plain terms
Sift stores nothing you paste and runs the local analysis without sending text anywhere. The only thing that leaves your device is an explicit, deliberate action: pressing "Run connected detector", which submits the text to the provider an administrator configured. That external hand-off is disclosed in the interface before it happens.
Performance, honestly
Local readings are single-pass over a capped window (the first 4,000 words), so they stay quick even on modest machines. No benchmark figures are claimed because none have been measured. Real detector speed depends entirely on the provider you configure.