Known limits
Where this gets it wrong.
This page exists so that nobody has to say “we did not know”
after something goes wrong.
| Situation | What happens |
|---|
| Edited generated tracks (pitch, reverb, re-encode, excerpt) |
We miss roughly 29.0% |
| A generator that was not in training | Detection drops to 84.0% |
| Human-made music | 1.5% are wrongly flagged as generated |
| Which tool produced it | Unknown unless the metadata says so |
| Proving a human made it | Cannot be done — by us or anyone |
| Video, speech | Not covered |
Tightening the threshold to catch more edited material
raises false positives against human artists. When royalties depend on the call, we consider
that the worse error — which is why the default operating point sits at 1.5%
false positives.
We could catch more — at a price
89.7% is the setting we chose, not a ceiling. Loosen it and unedited detection
reaches 96.0%. But false positives on human music rise from 1.5% to
5.0%.
| False positives | Unedited | Edited | Unseen generator |
|---|
| 1.0%% | 84.7%% | 65.2%% | 74.7%% |
| 1.5%% | 89.7%% | 71.0%% | 84.0%% |
| 2.0%% | 91.0%% | 72.4%% | 86.7%% |
| 5.0%% | 96.0%% | 83.3%% | 94.7%% |
On a 12,480‑track human catalog, 1.5% means 187 tracks wrongly
accused; 5.0% means 624. That is why the strict setting is the default.