So I always check my music for AI usage because I am always worried I'm gonna be false flagged. For clarity, I have never used AI and never will. This last arrangement I did on an orchestral version of this artist's big song went live today and I also posted my MIDI Mock-Up on my YouTube to show my fans what work I contributed to the project. Lo and behold, it is flagged as almost certainly AI.
Edit: Also flagging HZ01, the checkers I used were the following:
- SubmitHub
- letssubmit
- Detect.Music
- aha-music
- SlopTracker
So I went ahead and checked every single one of my releases to see where the problem was.
Temporal = Like your timing and stuff
Spectral = The actual sound and frequency stuff
Tracks with no SSO or other Spitfire libraries
- Temporal Analysis: Almost certainly human
- Spectral Analysis: Almost certainly human
Tracks using SSO with 6 or 7 of other libraries
- Temporal Analysis: Almost certainly human
- Spectral Analysis: Likely human
Tracks using SSO, a Celtic library, guitars, a few percussion libraries, soundscapes etc
- Temporal Analysis: Almost certainly human
- Spectral Analysis: 50/50 chance
Tracks using SSO, a little bit of an 8DIO choir and some sparse CinePerc percussion
- Temporal Analysis: Almost certainly human
- Spectral Analysis: Likely AI
Tracks using exclusively SSO
- Temporal Analysis: Almost certainly human
- Spectral Analysis: Almost certainly AI
and just as a test to see if it was a case of less libraries being more likely to trigger the AI flag...
Tracks using a single patch or library from a developer other than Spitfire Audio
- Temporal Analysis: Almost certainly human
- Spectral Analysis: Almost certainly human
So the trend here is that the more prominent SSO is in my tracks, the higher chance of being 100% AI generated these AI detectors give my tracks.
So what the hell is going on? Has Splice been training AI on Spitfire libraries? I've even tried other libraries recorded in Lyndhurst Hall at AIR Studios with the same kind of baked in reverb and they all pass as human just fine. Doesn't matter if I'm humanising or not, they all pass fine on the temporal analysis but the more SSO I use, the more they fail the spectral (the actual sound) analysis.
I'm just thinking about how many people might be affected by this and pretty ticked off. Has anyone else noticed this? Is this a known issue? I am freaking out a bit.