the most accurate AI song detector

your AI songs bore me.

↓ upload any song to check How Human it is ↓

that doesn't look like a spotify or apple music song link
or
song links only analyze a short clip of the song. for the most accurate results, upload a file of the full song.song links only analyze a short clip. upload the full song for the best results.

listening closely…

AI-generatedhuman-made
human-sounding AI tells

how we decided

    report card preview

    let people know...

    a square image sized for instagram and facebook.

    recent report cards

    ✗ detention AI slop detected
      ✓ honor roll certified human

        the most accurate AI song detector

        HowHuman.fm is a free AI music detector that checks whether a song was made by AI. paste a spotify or apple music link, or upload an audio or video file (mp3, wav, flac, m4a, ogg, mp4, mov), and seven independent forensic tests listen for the fingerprints AI music generators leave in the audio. you get one honest verdict: human-made, AI-generated, or a human/AI hybrid, plus a HowHuman score from 0 to 100 and a colored map of where the AI tells appear in the track.

        how accurate is this AI song detector?

        we benchmarked this exact pipeline on 1,136 tracks: 440 songs from 22 AI generators and 696 real human songs across every major genre. it caught 93% of AI songs, cleared 94% of human songs, and falsely accused under 3% of humans. every number was measured with the tested source held out of training, and the full breakdown is under "how it works".

        which AI music generators can it detect?

        suno (including recent versions), udio, riffusion, stable audio, musicgen, yue and more: 22 generators in the benchmark, plus checks for AI watermarks and signed C2PA content credentials. brand-new generators may leave marks we haven't learned yet, so a clean result means "no known tells", not a guarantee.

        can it check tiktok sounds and instagram reels?

        yes: save the video to your phone and upload it here. the audio is pulled out of mp4 and mov files automatically. direct tiktok and instagram links aren't supported because those platforms block automated downloads.

        is it really free?

        yes. HowHuman.fm is a free, non-commercial research tool: no account, no signup, no ads. everyone gets 2 checks a day, which keeps the free server alive.

        how we catch the slop

        AI music generators are impressive liars. they're also sloppy. every generator leaves fingerprints all over the audio, and we know where to dust. every song you check runs a gauntlet of seven independent tests:

        the receipts. we benchmarked this exact pipeline on 1,136 tracks: 440 from 22 AI generators (suno, udio, riffusion, stable audio, yue and friends) and 696 real human songs spanning every major genre, from indie archives to famous hits. it caught 93% of AI songs and cleared 94% of human songs; it falsely accused under 3% of humans, and honestly said "too close to call" on the rest. every number was measured with the tested source held out of training, so we can't grade our own homework. unlike some people, we don't publish numbers we didn't earn.

        a clean result means "no known tells", not a papal blessing: brand-new generators may leave marks we haven't learned yet.

        about HowHuman.fm

        a free, non-commercial research tool that estimates whether a song was made by an AI music generator (Suno, Udio, MusicGen …). it listens for the subtle traces those systems leave in the audio with seven independent tests, checks for inaudible AI watermarks, signed content credentials (C2PA) and file-metadata tells, and sums it all up as a single verdict with notes, a timeline of where the traces appear, and a per-instrument view.

        scores are probabilistic evidence, not proof. do not use this as sole grounds to accuse anyone of anything. in our own benchmark (1,136 tracks across 22 generators and every major human genre, details under "how it works") we caught 93% of AI songs and falsely accused under 3% of human ones; that risk is why a bad score alone proves nothing. the timeline and per-instrument views are experimental and noisier than the overall verdict.

        what it can't catch

        credits

        built on published research and open models: lofcz's fakeprint detector (MIT; method by Deezer Research, ISMIR 2025, patent-pending) · robust fakeprint by Barand, Dugelay et al. (ISMIR 2026, CC BY-NC 4.0) · ArtifactNet by Intrect (CC BY-NC 4.0, patents pending) · SpecTTTra from SONICS (ICLR 2025, MIT) · codec-forensics detector by Deezer Research (ICASSP 2025, CC BY-NC 4.0) · AudioSeal by Meta AI (MIT) · C2PA content credentials (CAI) · Nes2Net by Liu et al. (Apache 2.0), with WavLM by Microsoft · stem separation by Demucs (Meta, MIT) · yt-dlp, FFmpeg, ONNX Runtime, PyTorch.

        your audio is analyzed on this server. the playable copy on the results page is kept for up to an hour so you can listen back, then deleted. songs tested by link appear with their score and artwork in the public honor roll and detention lists; uploaded files never do. everyone gets 2 free checks a day; it keeps the free server alive.