SubmitHub checked a million song submissions and found 23.2% were fully AI-generated

SubmitHub ran its own catalogue of promo submissions through AI detection and published the result this week. Out of more than 1 million releases that went through the pipeline, 23.2% were fully AI-generated, and another 15.3% showed evidence of AI-generated audio a human had gone back and modified or processed afterwards. Music Ally surfaced the numbers on August 18.

Why a submission number hits differently than an upload number

We already had figures for how much AI music reaches streaming catalogues, and they are large. Deezer crossed 50% of daily uploads earlier this year and now detects roughly 60,000 AI tracks a day. Those numbers describe what gets distributed, and distribution is close to free, so a flood there tells you mostly about the cost of uploading.

SubmitHub measures something else. Every track in this sample was pitched by someone who wanted a Spotify playlister, a blogger or an influencer to listen to it, and on SubmitHub that pitch costs money. The people behind these 23.2% were not dumping files into a catalogue and walking away. They were spending on promotion.

That is the part worth sitting with. A quarter of the songs being actively worked, by people paying for a shot at coverage, came out of a generator with no human in the audio at all.

What the 15.3% bucket says about how creators are actually working

The hybrid number is the more interesting one, and it is the number a binary detector handles worst. Those tracks fired the AI signal and then showed human hands afterwards, which describes an enormous range of real practice: a producer who generated a bed and rebuilt it, someone who took an AI vocal and cut a new arrangement underneath, a writer who used a generated demo as a sketch and replaced most of it.

None of those look alike creatively, and a single AI flag says the same thing about all of them.

SubmitHub built the detection in-house and is now licensing it to the industry, which is the exact path Deezer took after solving the same problem for itself. That makes at least four companies selling audio-side AI detection into a market where Spotify is about to start labelling AI personas, and where Cyanite and RTM Audio are already selling detection APIs.

For an independent artist pitching through SubmitHub, the practical read is short. Curators now open submissions knowing roughly 4 in 10 have AI in them somewhere, so the burden of proof on a human record has quietly gone up, and the fastest way to answer it is the stuff detection cannot read from a file: your face, your process, your live playing, your story.