What an AI music detection API returns
An AI music detection API takes an audio file and returns an automated estimate that it was machine generated. For each track, the ArtifactNet API returns a verdict (ai or real), the probability behind it, the decision threshold that produced the verdict, the model version, the file's SHA-256 hash, and a probability for every 4-second segment.
| Field | What it holds |
|---|---|
verdict | ai when the probability reaches decision_threshold, real otherwise. |
probability | An automated estimate, from 0 to 1, that the audio was machine generated. |
decision_threshold | The cut-off used for this verdict, recorded with the analysis that produced it. |
model_version | The model that produced the verdict, so a result can be traced to the model that ran then. |
audio_sha256 | The SHA-256 hash of the submitted file, to match a verdict back to your own copy. |
segments | For each consecutive 4-second window: segment_index, start_seconds, end_seconds, probability and verdict. |
duration_seconds | The measured length of the analysed audio. billable_seconds is the length billed. |
Limits that travel with every export
- A probability is an automated estimate, not a determination that any person used AI. The detector is wrong in both directions.
- The threshold is a product choice, not a property of the audio. Re-read the probabilities against your own cut-off if it differs from ours.
- This model generation does not identify which generator produced a track.
- A tail shorter than two seconds is not analysed. Per-segment detail is kept for 30 days; the verdict stays after that.
API, dashboard batch or web checker?
- Web checker (demo.intrect.io): one song at a time, no signup, 2 minutes per track and 6 minutes in total.
- Dashboard batch (app.intrect.io): upload many files in the browser and export XLSX, CSV or JSONL. The free tier includes 10 minutes of audio in total, with no credit card.
- REST API (
api.intrect.io): the same results and exports for an ingestion pipeline. It needs a Pro API key; see the quick start below.
How it works
ArtifactNet reframes AI music detection as forensic residual physics. Instead of learning generator-specific stylistic fingerprints, it isolates the physical residue that neural audio codecs imprint on generated audio. A bounded-mask UNet extracts those codec residuals from the magnitude spectrogram, HPSS decomposes them into multi-channel forensic features, and a compact CNN classifies real vs. AI-generated.
spectrogram
codec residual extractor
7-channel forensic features
residual structure scoring
+ calibrated confidence
The mechanism: commercial AI music generators (Suno, Udio, Stable Audio, MusicGen, Riffusion) all rely on neural audio codecs that use Residual Vector Quantization (RVQ), mapping continuous audio onto discrete codebook vectors. When a separation model trained only on human music meets AI-generated audio, the reconstruction residuals have measurably different structure — a phenomenon the paper terms forensic residual amplification. ArtifactNet detects that structural difference, so it generalizes to generators never seen in training. Read the paper (arXiv:2604.16254) →
Who uses it
Evaluation versions and limitations
September 2026 evaluation paper. ArtifactBench: Lineage-Aware Evaluation (v1) reports ArtifactNet AUROC 0.982 and balanced accuracy 0.918 on the 562-track common-success test intersection. The public Deezer detector scores 0.761 and 0.776 respectively on that intersection. These are not F1 scores, universal accuracy claims, or measurements of every file submitted to this service.
The protocol groups related recordings, separates calibration from testing, and reports inference coverage separately from classification errors. Scores depend on the cohort, threshold and successful-inference subset. This is developer-authored research: ArtifactNet's developer also authored the evaluation. Read the preprint and evaluation code for methods and limitations. A dated, measured comparison of eight public detectors on one test set is in the Artifact Observatory field report.
Historical detector-paper results — not directly comparable
The figures below come from ArtifactNet v9.4, unseen test: 2,263 tracks in the earlier detector paper (v2). Its full collection contains 6,183 tracks. Do not interpret differences from the September evaluation as a model improvement or regression: the evaluation populations and metrics differ.
The historical full collection spans 22 AI music generators and 6 real music sources. The full collection and the held-out test subset are different populations. Source: Oh, H. (2026). ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics, arXiv:2604.16254v2.
(identical eval conditions)
Common questions
Quick start
Every call takes a Pro API key as $ARTIFACTNET_API_KEY: create one in the dashboard under Settings → API. The API reference also shows a one-call upload (POST /v1/batch) for a single file.
# 1. Upload tracks and start a batch job
curl -X POST https://api.intrect.io/v1/batch/create \
-H "Authorization: Bearer $ARTIFACTNET_API_KEY"
# 2. Append audio files
curl -X POST https://api.intrect.io/v1/batch/$BATCH_ID/files \
-H "Authorization: Bearer $ARTIFACTNET_API_KEY" \
-F "files=@track.wav"
# 3. Commit to queue
curl -X POST https://api.intrect.io/v1/batch/$BATCH_ID/commit \
-H "Authorization: Bearer $ARTIFACTNET_API_KEY"
# 4. Poll for results (xlsx | csv | segments-csv | jsonl)
curl https://api.intrect.io/v1/batch/$BATCH_ID/results?format=jsonl \
-H "Authorization: Bearer $ARTIFACTNET_API_KEY"Simple, usage-based plans
- Everything in Free
- Priority support
- Everything in Artist
- One credit for detection and cleanup
- Everything in Artist
- Per-segment forensic export
- Batch history 90 days
- Everything in Creator
- API key access
- Multiple API keys
- SLA support