COSMIC

Classification of Synthetic Music in Content

Industry leading AI Music Detection

Analyse a recording

Analyse a recording

Drop a file or paste a Spotify, YouTube, SoundCloud or TikTok link.

Needs at least 20 seconds of audio. Spotify, YouTube and SoundCloud links analyse a 30-second section.

Idle

By analysing, you agree to the Terms of Service and Privacy Policy.

Audio is deleted when the analysis ends.

Why COSMIC

  • Still catches processed tracks

    Re-encoded, slowed down or pitch-shifted, COSMIC kept at least 89% of what it caught on the original audio. Some other detectors lost nearly all of theirs.

  • Catches most AI tracks

    94.0% of AI tracks caught at the same 1% false-flag rate. The best of the other detectors we tested caught 63.2%.

  • Every check reported

    Every result says which checks found a signal and which found nothing, so a person can judge it instead of trusting a single score.

  • Quick results

    Most analyses finish in about 4 seconds once the audio has uploaded.

What COSMIC checks

COSMIC combines its model with evidence in the recording. Each result shows what was checked and what was found.

COSMIC model
Looks for signals of generated music in the audio.
Content Credentials
Checks signed provenance that still matches the file.
Watermarks
Looks for supported markers embedded in the audio.
File metadata
Reads tags and hidden fields. Editable, so never enough on their own.
How to read a result
AI-generated audio detected
COSMIC found signals consistent with AI-generated audio. The percentage reflects the model’s assessment, not proof of how the recording was created. Results can be incorrect and do not establish whether the recording was used with permission.
No supported signal detected
Every requested check ran and found nothing. That is not proof a person made it.
Inconclusive
A required check was unavailable, or only editable metadata points to AI.

A service is named only when a signed credential identifies one.

Who COSMIC is for

  • Labels and distributors

    Check a catalogue before release, and flag recordings that were re-uploaded after being changed.

    Get in touch

  • Rights holders and royalty bodies

    A consistent, documented result to base registration and payout decisions on.

  • Platforms and marketplaces

    A result with its evidence attached, so moderation decisions can be explained.

Detection rate by generator

How many AI tracks COSMIC catches, by generator. For a generator it was never trained on, the average drops to 75.3%.

Version 1.0 · Sep 2026 · versions of the same generator are combinedFull table

These are the share caught in our tests. They are not a promise about every track a generator can make.

About the detection evaluation

The 95% ranges resample whole artists and source recordings together. The Suno range is the spread across its eight versions, not a confidence interval.

Against other detectors

In October 2026 we ran COSMIC, four open-source detectors and one commercial detector on the same tracks, each set to wrongly flag the same 1% of human music. Those detectors are third-party tools, used only for this comparison; see sources and notices.

Pick an edit to compare each detector.Full table

How often human music is flagged as AI

Human-made tracks that COSMIC called AI by mistake. The worst group measured was 1.47%.

Same cutoff as above.Full table

COSMIC after common processing

The same tracks before and after everyday processing. COSMIC keeps most of what it catches after each change.

Measured on 14,919 AI tracks (catches) and 13,881 human tracks (false flags).

Common production edits

The same AI tracks with each edit applied. Shorter bars are better.Full table
Test coverage and the final evaluation

The wider test covered 70 manipulations in 17 families. Those results are used to improve COSMIC and are not published.

A later held-out evaluation found the four largest generator populations within 0.4 points of development, with a false-positive rate about 1.5× higher. See the final evaluation.

Ways to use COSMIC

  • In the browser

    Check your tracks here with our easy-to-use web interface.

    Analyse a file

  • API

    Send audio from your own service and get the same result and evidence.

    Create a key API documentation

  • In AudioAuditor

    COSMIC’s service is also built into our free, open-source audio analysis app, AudioAuditor.

    Visit AudioAuditor

  • A pilot on your catalogue

    COSMIC measured on your own music.

    Plan a pilot

Why it matters

AI tracks are uploaded in bulk and played to collect royalties that would otherwise go to real artists. Deliveries of generated music have grown quickly at Deezer. The chart follows its published daily counts over time.

Listeners and artists want to know what they are hearing. COSMIC reports what it finds in a recording and shows the evidence behind it, so people can decide for themselves.

Fully AI-generated tracks delivered to Deezer per day, as detected by Deezer.Jul 2026

What a result does not tell you

  • Generators it has not seen

    A generator released after training is harder to catch: about 75% on average when a whole family was held out.

  • False positives depend on the music

    Between 0.8% and 1.5% per artist on the collections measured. No single rate fits every catalogue.

  • Music, not speech

    Evaluated on music only. Speech, podcasts and sound effects are outside what the results describe.

  • Your catalogue may differ

    Every rate was measured on the populations described. A pilot measures COSMIC on your own music.

  • Human and AI mixed together

    COSMIC is less reliable on recordings that mix human-made and AI-generated material, and a result may not reflect the AI part.

  • Short recordings

    COSMIC needs at least 20 seconds of audio. Below that it gives no AI verdict either way.

Get COSMIC updates

Get notified about new benchmarks, models, API features and research.

No spam. Unsubscribe any time.

Contact

Bugs, ideas, or something that could work better. You can also email us.