Skip to main content

AI Music Detection

AI Music Detection is the Cyanite product that finds out whether a track was likely generated by a Gen AI music model, such as Suno or Udio, and by which one. It's a product of its own: you can use it on its own or together with Tagging and Search.

This guide explains why and how the detection works, shows you how to fetch and use the result in your application, and how far to trust it.


Why detect AI-generated music?​

Gen AI music models can produce complete, convincing tracks in seconds. For anyone who receives music from third parties, such as distributors, publishers, libraries and platforms, that raises questions that the metadata of a track can't answer:

  • Catalogue integrity: does the catalogue contain what it claims to contain?
  • Policy and royalties: many services treat fully AI-generated tracks differently when it comes to licensing, payouts or eligibility.
  • Transparency: listeners and clients increasingly expect AI-generated music to be disclosed.

AI Music Detection gives you an independent signal from the audio itself, so you can apply your policy consistently across a large catalogue.


How it works​

AI Music Detection looks at the audio only. It doesn't rely on the title, the artist, file metadata or embedded watermarks, so it works regardless of how a track was delivered or whether its metadata was changed.

Different Gen AI music models leave different traces in the audio. Cyanite therefore runs one detector per supported Gen AI music model, each trained to recognize that model's output. Each detector produces a likelihood between 0.0 and 1.0 that the track was generated by its model.

The result combines them:

  • score is the highest likelihood across all detectors.
  • isAiMusic is true when score is above 0.5.
  • suspectedModel names the model whose detector produced that score, for example suno or udio. It's only set when isAiMusic is true, and null otherwise.

So a track gets a high score if any of the detectors recognizes it, and the result tells you which one.

The detection covers the whole track and produces one result per track, not per segment.


AI Music Detection and Tagging​

AI Music Detection is sold separately from Tagging, but it's built on the same foundation. Its result is delivered as the AiMusicDetectionV1 model output, the same kind of result that Tagging delivers:

  • It runs as part of the regular analysis of every uploaded track, alongside any Tagging model outputs your account has. You don't start it separately.
  • You fetch it from the same endpoint as every other model output, GET /v1/library-tracks/{id}/models, with model=AiMusicDetectionV1.
  • If your account has only AI Music Detection, AiMusicDetectionV1 is the only model output you can request. Other model outputs fail with a 401.
  • If your account also has Search, you can use AiMusicDetectionV1.isAiMusic in a metadata filter.

Prerequisites​


Fetch the result​

When your webhook receives analysisJob/finished for a track, request the AiMusicDetectionV1 model output for that track:

GET https://rest-api.cyanite.ai/v1/library-tracks/{id}/models?model=AiMusicDetectionV1

API reference: Get Model Outputs

// trackId comes from the webhook payload: payload.data.track.id
const response = await fetch(
`https://rest-api.cyanite.ai/v1/library-tracks/${trackId}/models?model=AiMusicDetectionV1`,
{
headers: {
"x-api-key": process.env.CYANITE_API_KEY,
},
},
);

const { items } = await response.json();
const detection = items.find(
(item) => item.version === "AiMusicDetectionV1",
);

If your account also has Tagging, you can request AiMusicDetectionV1 together with other model outputs in the same call, for example ?model=AiMusicDetectionV1&model=MainGenreV2.


Read the result​

{
"items": [
{
"version": "AiMusicDetectionV1",
"isAiMusic": true,
"score": 0.94,
"suspectedModel": "suno"
}
]
}
FieldTypeDescription
versionstringAlways "AiMusicDetectionV1".
isAiMusicbooleantrue if the track is likely AI-generated, which is the case when score is above 0.5.
scorenumber0.0-1.0. The highest likelihood across all supported Gen AI music models that the track was generated by one of them.
suspectedModelstring | nullThe Gen AI music model the track most likely comes from, for example suno or udio. null when isAiMusic is false.

Interpreting the result​

score expresses how likely the track is to be AI-generated. It's a probability, not proof. Like every classifier, the detection can be wrong in both directions:

  • False positives: a human-made track can score above 0.5.
  • False negatives: an AI-generated track can score below 0.5, for example when it comes from a Gen AI music model Cyanite doesn't detect yet.

The score of AI Music Detection isn't comparable to the scores of other model outputs, such as moods or genres, see Interpreting scores.

Use the result to prioritize and support decisions, such as sorting a review queue or asking a supplier for clarification, rather than as the only basis for irreversible actions like takedowns.


Flag tracks for review​

The simplest policy is to use isAiMusic directly, for example to route AI-generated tracks to a review queue:

// sendToReviewQueue stands for your own code
if (detection.isAiMusic) {
await sendToReviewQueue(trackId, {
reason: "ai-generated",
suspectedModel: detection.suspectedModel,
});
}

Treat the result as a strong signal, not as proof. See Interpreting the result.


Use your own threshold​

isAiMusic uses a threshold of 0.5. That's a balanced default, but your use case may need a different trade-off:

  • A higher threshold flags fewer tracks, and the flagged tracks are more likely to be AI-generated. Choose it when a false accusation is costly.
  • A lower threshold catches more AI-generated tracks, at the cost of flagging more human-made ones. Choose it when missing an AI-generated track is costly and a human reviews every flag.

Many integrations combine both: automatic action only for very high scores and human review for the range in between. Apply your own threshold to score, for example review everything above 0.3, but only block automatically above 0.9:

// block and sendToReviewQueue stand for your own code
if (detection.score > 0.9) {
await block(trackId);
} else if (detection.score > 0.3) {
await sendToReviewQueue(trackId, { reason: "possibly-ai-generated" });
}

Show the suspected model​

When isAiMusic is true, suspectedModel names the Gen AI music model the track most resembles. It's useful context for a reviewer, but it describes the strongest match, not a certainty. Cyanite adds detectors for new models over time, so new values can appear. Treat the field as an open set of names and display unknown values as they are instead of mapping only a fixed list:

const label = detection.suspectedModel
? `Likely generated with ${detection.suspectedModel}`
: "No AI generation detected";

AiMusicDetectionV1.isAiMusic can be used in a search metadata filter, if your account also has Search. This Search by Prompt only returns tracks that aren't predicted to be AI-generated:

{
"query": "warm acoustic folk for a family brand",
"metadataFilter": {
"AiMusicDetectionV1.isAiMusic": { "$eq": false }
}
}

Tracks without an AI Music Detection result don't match this filter.


Limitations​

  • Partial AI use. The detection estimates whether a track as a whole was generated by a Gen AI music model. It doesn't tell you whether a human used AI tools for parts of the production, such as a stem, a sample or mastering.
  • Rights and ownership. A result says nothing about who owns a track or whether it may be used. A human-made track can still infringe rights, and an AI-generated one can be licensed properly.
  • Models it doesn't know. A track from a Gen AI music model that Cyanite has no detector for can get a low score.

Troubleshooting​

ProblemCauseWhat to do
401Invalid API key, or AI Music Detection isn't enabled.Check x-api-key, or contact Cyanite.
401 when requesting other modelsYour account only has AI Music Detection.Only request AiMusicDetectionV1. The error's detail names the models you can't access.
404The track doesn't exist or doesn't belong to your account.Check the Library Track ID.
Empty itemsThe analysis hasn't finished yet.Wait for the analysisJob/finished webhook before fetching the result.
analysisJob/failed webhookThe track couldn't be analyzed, for example because it's too long.Check the file, then upload it again. If it keeps failing, contact Cyanite.

Next steps​