How do I detect if a song was made by AI?

Upload the audio file (MP3, WAV, FLAC, etc.) at demo.intrect.io. ArtifactNet separates the track into harmonic and percussive components, computes a spectral residual with a U-Net, and scores every 4-second segment with a 7-channel CNN. The page shows a verdict (AI / Human-Made / Partial AI) with a probability, per-segment scores, and a forensic feature radar — all in a few seconds. Segment scores and the forensic feature radar are shown in full on your first analysis of the day; later analyses that day show the verdict and probability, and sign-in (free) keeps every analysis unlocked.

Which AI music generators can ArtifactNet detect?

ArtifactNet is trained on tracks from Suno, Udio, Stable Audio, Riffusion, MusicGen, and a residual 'other' class, and it generalises to AI models it has not seen because it targets spectral-residual artefacts left by STFT-domain generators rather than model-specific fingerprints.

Is ArtifactNet free? Do I need to sign up?

The demo is free and requires no signup: it analyses the first 2 minutes of each submission, up to 6 minutes of audio in total per visitor, with short-term limits of 3 analyses per minute and 20 per hour. Your first analysis of the day returns the full forensic detail without an account; later analyses that day show the verdict and probability. A signed-in account starts a 7-day trial of the Artist plan (card required, cancel anytime) for full-length analysis up to 10 minutes per file and forensic detail on every analysis. A paid API (api.intrect.io) is available for batch processing and integrations.

Does ArtifactNet store my uploaded audio?

No. The demo operates on a zero-retention policy. Your audio is streamed to a GPU worker, analysed in memory, and discarded. Two buttons are the only things that save anything: press Report to tell us a verdict is wrong and we archive the spectral analysis together with the original audio, which a report requires because the verdict cannot be reproduced without it; press Share and that one verdict is published at a link until you revoke it. Sharing stores the verdict and its scores, never the audio and never your filename.

How accurate is ArtifactNet?

On ArtifactBench (v9.4, unseen test, 2,263 tracks; full benchmark 6,183 tracks, 22 AI generators) ArtifactNet reports F1 0.9829 and a 1.49% false-positive rate on real music. Rare genres remain the focus of ongoing retraining. The demo lets you report suspected misclassifications so they can improve the detector. See arXiv:2604.16254v2.

What does the forensic feature radar show?

The radar plots six summary signals from the analysis: mean and median AI probability across segments, consistency (how uniform the verdict is), high-frequency residual energy, U-Net mask strength, and harmonic share. Together they reveal why the model reached a given verdict and expose edge cases. The radar is shown in full on your first analysis of the day; after that, sign in to keep it on every analysis.

Why did ArtifactNet flag my own song as AI?

Heavy spectral processing (strong multi-band limiting, resampling through AI-era codecs, aggressive stereo widening) can produce residuals that resemble AI-generated artefacts. Hit the Report button and choose 'Human-Made' with a short note — a human reviewer will look at the spectral residual and the case can be used to improve the classifier.

Free AI Music Detector — ArtifactNet Demo by Intrect

About ArtifactNet

ArtifactNet is a free online tool that detects AI-generated music. Upload an audio file and get an AI-vs-Human verdict in seconds, with per-segment probabilities and forensic feature readings. F1 0.9829 on ArtifactBench (v9.4, unseen test, 2,263 tracks). Zero-retention — audio is not stored unless you submit a Report.

What it detects

ArtifactNet is trained on tracks from Suno, Udio, Stable Audio, Riffusion, and MusicGen, with a residual other class for unseen generators. It targets STFT-domain residual artefacts rather than model-specific fingerprints, so it generalises to AI models it has not seen during training.

How it works

Decode the track to 44.1 kHz mono and split into 4-second overlapping chunks.
A codec-aware U-Net predicts a spectral residual mask that isolates AI artefacts.
HPSS (median-filter harmonic/percussive separation) runs on the residual.
A 7-channel CNN scores every chunk — 0 means clearly human, 1 means clearly AI.
The median across chunks becomes the track verdict. Outlier distributions are labelled Partial AI.

FAQ

How do I detect if a song was made by AI?: Upload the audio file at the top of this page. The demo returns a verdict with probability, a per-segment timeline, and a forensic feature radar in a few seconds. No signup — your first analysis of the day shows the segment scores and the radar in full, and signing in (free) keeps them on every analysis.
Is this free?: Yes. The demo is free and does not require an account. The demo reads the first 2 minutes of each submission, and anonymous use is capped per IP at 6 minutes of audio in total (plus short-term limits of 3/minute and 20/hour). Signing in with Google starts a 7-day trial of the Artist plan (card required, cancel anytime). For batch processing or integrations, see the paid ArtifactNet API.
Do you store my audio?: Not for the analysis itself — that is zero-retention. Your file is streamed to a GPU worker, analysed in memory, and discarded. Two buttons are the only things that save anything, and they differ: Report tells us a verdict is wrong and requires the original audio, because a wrong verdict cannot be reproduced without the file that produced it — pressing it uploads and keeps that audio together with the spectral analysis. Share publishes that one verdict at a link until you revoke it, and stores neither your audio nor your filename. Full details in the Privacy Policy.
Which AI generators does it detect?: Trained explicitly on Suno, Udio, Stable Audio, Riffusion, MusicGen. Generalises to unseen generators via residual-artefact targeting.
Can I analyse a song from a streaming site?: Not from a link. The demo used to fetch audio from a URL and that was withdrawn — download or export the track yourself and upload the file. Nothing you upload is stored unless you submit a Report.
How accurate is it?: F1 0.9829 / FPR 1.49% on ArtifactBench (v9.4, unseen test, 2,263 tracks; full benchmark 6,183 tracks, 22 AI generators). Rare human-made genres remain the main retraining target — please report any miss you find.
Why did it flag my own song as AI?: Heavy spectral processing (aggressive limiting, AI-era codec resampling, extreme stereo widening) can mimic AI residuals. Use the Report button with Human-Made and a short note so we can investigate and improve the detector.
Is there an API?: Yes. api.intrect.io exposes the same detector as a REST API with batch processing, designed for distributors, labels, and streaming platforms.

ArtifactNet

1. Add your track

2. Result

Think this is wrong?

About ArtifactNet

What it detects

How it works

FAQ