A demo lands in your inbox. A beat seller sends you "an exclusive". A track blows up overnight with an artist page that didn't exist last week. The question didn't exist three years ago, and now it's everywhere: was this made by a person — or generated?
Since Suno, Udio and AI voice models went mainstream, fully generated tracks have been flooding streaming platforms, sample packs, sync libraries and demo submissions. Some of it is harmless fun. But if you're a label listening to demos, a producer buying loops, a curator building playlists, or an artist protecting your own name from voice clones — you need a way to check.
Signing or playlisting a generated track can mean rights problems you can't untangle later. Who owns a song no one wrote?
That "royalty-free" sample pack might be AI output trained on copyrighted catalogs — a legal grey zone you don't want in your masters.
Voice cloning means your vocal identity can be borrowed without you. Knowing how detection works is the first line of defense.
The platforms are moving. Deezer already tags AI tracks, Spotify has purged AI slop farms, and distributors increasingly ask you to declare AI use. Getting caught misdeclaring can pull your whole catalog.
Before any tool, your ears can spot a lot:
No drift, no push-pull between musicians, every hit machine-centered (and not in the intentional, quantized-trap way).
AI voices still struggle with hard T's, K's and S's; words melt into each other under scrutiny.
Generated lyrics love safe clichés and abstract emotion with no concrete detail.
Sections repeat with cosmetic variation instead of intent.
Reverb tails and fade-outs often carry a subtle metallic shimmer, the fingerprint of the vocoder stage.
AudioKit's AI Detector analyzes the audio itself — not the metadata, not the artist name — and estimates the probability that a track was AI-generated:
Like every AudioKit analysis tool, the whole file is analyzed, and the result comes back in seconds.
Honesty matters here: no detector is a court of law. Detection is a probability, not a proof — and it's an arms race, with generators improving every month. A high AI score on a heavily auto-tuned, quantized human production is possible; a clean score on a carefully post-processed generation is too.
Use the score the way a doctor uses a test: as one strong signal, combined with your ears, the context (who sent it, where it came from, whether the artist exists anywhere else), and — when the stakes are high — a direct conversation with the creator, stems included. Real sessions have stems. Generations don't.
Mixed tracks are the hardest case — the score reflects the blend. Isolate the vocal first (the Voice Remover does this) and analyze it alone for a clearer read.
The analysis runs on our servers over an encrypted connection and the file isn't kept — same policy as AudioKit's other server-side tools.
That's exactly the use case. Ask for a preview file, run it through, and factor the result into the deal.