AI music detector API

Send a song, get a score from 0 (no AI fingerprint found) to 100 (AI), the verdict, a per-section breakdown and, optionally, a verifiable PDF report. Access is by personal API key; each key has its own request allowance.

Getting access

API keys are issued individually by the operator of this service, each with its own request allowance. Ask the operator for a key.

Authentication

Send your key as Authorization: Bearer aim_... (or X-API-Key). Keep it secret; it is shown once when it is created.

Analyse a file

curl -X POST https://api.youmadeit.org/api/v1/analyze \
  -H "Authorization: Bearer aim_YOUR_KEY" \
  -F "file=@song.mp3"

# or the raw bytes:
curl -X POST https://api.youmadeit.org/api/v1/analyze -H "Authorization: Bearer aim_YOUR_KEY" \
  -H "X-Filename: song.mp3" --data-binary @song.mp3

# or a YouTube link:
curl -X POST https://api.youmadeit.org/api/v1/analyze -H "Authorization: Bearer aim_YOUR_KEY" \
  -H "Content-Type: application/json" -d '{"url": "https://www.youtube.com/watch?v=..."}'

By default the request waits for the result (usually 10-40 s, at most about 80 s). Add ?wait=0 to get a job id immediately and poll GET /api/v1/jobs/<id>. If a waiting request times out you get 202 with the job id as well.

{
  "id": "5b1f...", "state": "done",
  "result": {
    "file": "song.mp3", "score": 0.9731, "score100": 97,
    "verdict": "AI",                       // "AI" | "unsure" | "real" (= no AI fingerprint found)
    "verdict_label": "Likely AI-generated",
    "band": {"low": 0.5, "high": 0.77},    // AI >= high, unsure between, no AI found <= low
    "duration_s": 173.8, "n_clips": 8,
    "clips": [{"start_s": 0.0, "end_s": 20.1, "p": 0.98}, ...],   // AI probability of each ~20 s section
    "per_model": {"cnn_chain_v2": 0.97, ...}, "notes": ["..."], "analysed_at": "2026-10-12T16:39:56Z",
    "model_version": "v2"
  },
  "report_pdf": "/api/v1/jobs/5b1f.../report.pdf",
  "quota": {"limit": 100, "used": 3, "remaining": 97, "period": "total", "resets_at": null}
}

PDF report

GET /api/v1/jobs/<id>/report.pdf (same key) returns the report as PDF. It carries a report ID and a link; anyone can check it on the site's "Verify a report" form. Generating it stores a small record of the result (file name, verdict, score, section scores, time) for 365 days, like the web page's PDF export.

Allowance and limits

Each accepted analysis uses one request from the key's allowance (a fixed total, or a monthly allowance). Failed analyses are not charged (refunded automatically), and neither are rejected files. GET /api/v1/usage shows what is left; every response carries X-Quota-Limit / X-Quota-Remaining.
Per key: a number of requests per minute and of analyses in progress at the same time (see /api/v1/usage).
Files up to 40 MB; the first 6 minutes of a track are analysed. Formats: MP3, WAV, FLAC, M4A/AAC, OGG/Opus, WebM, WMA, AIFF, MP4.
Results are kept for 60 minutes and are visible only to the key that requested them. Audio is deleted right after analysis.

Errors

HTTP	code	meaning
401	invalid_api_key	missing or wrong key
403	key_revoked / key_expired / links_not_enabled	the key may not be used (for this)
413 / 415 / 400	too_large / not_audio / bad_request	the input was rejected; nothing is charged
422	analysis_failed	the file could not be analysed; the request is refunded
429	quota_exceeded / rate_limited / too_many_concurrent	allowance used up, or slow down (see `Retry-After`)
503	busy	the service is at capacity; retry shortly

Python

import requests
KEY = "aim_YOUR_KEY"
with open("song.mp3", "rb") as f:
    r = requests.post("https://api.youmadeit.org/api/v1/analyze", headers={"Authorization": f"Bearer {KEY}"},
                      files={"file": f}, timeout=120)
r.raise_for_status()
res = r.json()["result"]
print(res["score100"], res["verdict_label"])
pdf = requests.get("https://api.youmadeit.org" + r.json()["report_pdf"], headers={"Authorization": f"Bearer {KEY}"})
open("report.pdf", "wb").write(pdf.content)

What the score means

The detector recognises songs from two AI generators (Suno and Treblo) against real music. A low score means no known AI fingerprint was found, which is not proof that a human made the track; other generators are not recognised. Treat a result as evidence, not proof.