Skip to main content

Generator hub

What each AI music generator means for a detection reading

No two generators build audio the same way, and those differences decide which measurements are actually worth reading. Each page below explains how a given platform works and how much confidence its output deserves — including the cases where the honest answer is “not much”.

One thing holds across every platform here: this tool does not identify which generator made a track. It estimates whether a recording looks machine-generated, nothing more specific than that.

Check a track now

Generators covered on this site

A side-by-side view: output type and verification context

Read the right-hand column carefully — it names which measurement is actually informative for that platform’s output. It is never a claim that a file can be traced back to where it came from.

Swipe the table sideways to see all columns.

GeneratorWhat it producesVerification contextAttribution
SunoNo
UdioNo
ElevenLabs MusicNo
Stable AudioNo
RiffusionNo
MubertNo
Seed MusicNo
MiniMaxNo
Mureka (Sonauto)No
SoundrawNo

Shorter write-ups

These five platforms don’t need a full standalone page, so their notes are published here in full instead.

Riffusion

Begin with the workflow and records, then use audio classification as supporting information—not platform attribution.

Product context
Generated musical clips and songs, depending on version.
Useful evidence
Project history, exports and creator disclosure
Public detector output
Category, component indicators, confidence and timeline
Attribution supported
No

Start with the production path

Historical descriptions of a product may not represent its current model or interface.

A rendered master flattens that history into one waveform. Classification can assess the waveform, but it cannot recover prompts, ownership, consent or the division between generated and performed work.

Evidence to request before drawing a conclusion

Date any technical claim and identify the exact model release; old artefact lists age quickly.

  • The original export and its date
  • Prompts, generation history or platform records where available
  • DAW sessions, stems and edits made after export
  • A direct account from the creator

What the report actually contributes

The current site sends the complete file to a specialist third-party service. It returns a primary category, separate vocal and instrumental percentages, confidence and sometimes window values. The application does not blend those fields into a platform score.

The detector does not test for a Riffusion-specific signature.

Use the answer in proportion to the stakes

For curiosity, a result may be enough context. For moderation, contracts, copyright, employment or education, obtain records and human review. An inconclusive result is not a concealed accusation, while a not-AI result is not a certificate of human authorship.

Mubert

Begin with the workflow and records, then use audio classification as supporting information—not platform attribution.

Product context
Background tracks and production-ready music beds.
Useful evidence
Project history, exports and creator disclosure
Public detector output
Category, component indicators, confidence and timeline
Attribution supported
No

Start with the production path

Material may be created to fit a duration or use case and then edited into a larger project.

A rendered master flattens that history into one waveform. Classification can assess the waveform, but it cannot recover prompts, ownership, consent or the division between generated and performed work.

Evidence to request before drawing a conclusion

Licensing records and export history matter more than repetition in the audio.

  • The original export and its date
  • Prompts, generation history or platform records where available
  • DAW sessions, stems and edits made after export
  • A direct account from the creator

What the report actually contributes

The current site sends the complete file to a specialist third-party service. It returns a primary category, separate vocal and instrumental percentages, confidence and sometimes window values. The application does not blend those fields into a platform score.

Looping is common in human-made electronic and library music, so it cannot identify Mubert.

Use the answer in proportion to the stakes

For curiosity, a result may be enough context. For moderation, contracts, copyright, employment or education, obtain records and human review. An inconclusive result is not a concealed accusation, while a not-AI result is not a certificate of human authorship.

Seed Music

Begin with the workflow and records, then use audio classification as supporting information—not platform attribution.

Product context
Integrated vocal and instrumental music.
Useful evidence
Project history, exports and creator disclosure
Public detector output
Category, component indicators, confidence and timeline
Attribution supported
No

Start with the production path

Published demonstrations describe capabilities, but a third-party file rarely carries enough information to establish that it came from the demonstrated system.

A rendered master flattens that history into one waveform. Classification can assess the waveform, but it cannot recover prompts, ownership, consent or the division between generated and performed work.

Evidence to request before drawing a conclusion

Use primary technical documentation for capability claims and provenance for any individual track.

  • The original export and its date
  • Prompts, generation history or platform records where available
  • DAW sessions, stems and edits made after export
  • A direct account from the creator

What the report actually contributes

The current site sends the complete file to a specialist third-party service. It returns a primary category, separate vocal and instrumental percentages, confidence and sometimes window values. The application does not blend those fields into a platform score.

No Seed Music-specific evaluation has been completed here.

Use the answer in proportion to the stakes

For curiosity, a result may be enough context. For moderation, contracts, copyright, employment or education, obtain records and human review. An inconclusive result is not a concealed accusation, while a not-AI result is not a certificate of human authorship.

MiniMax

Begin with the workflow and records, then use audio classification as supporting information—not platform attribution.

Product context
Generated song material and full arrangements, depending on product version.
Useful evidence
Project history, exports and creator disclosure
Public detector output
Category, component indicators, confidence and timeline
Attribution supported
No

Start with the production path

Music can be conditioned by text or references, then altered after export.

A rendered master flattens that history into one waveform. Classification can assess the waveform, but it cannot recover prompts, ownership, consent or the division between generated and performed work.

Evidence to request before drawing a conclusion

A reference can shape style without leaving a platform-readable marker in the exported file.

  • The original export and its date
  • Prompts, generation history or platform records where available
  • DAW sessions, stems and edits made after export
  • A direct account from the creator

What the report actually contributes

The current site sends the complete file to a specialist third-party service. It returns a primary category, separate vocal and instrumental percentages, confidence and sometimes window values. The application does not blend those fields into a platform score.

The detector neither identifies reference use nor attributes output to MiniMax.

Use the answer in proportion to the stakes

For curiosity, a result may be enough context. For moderation, contracts, copyright, employment or education, obtain records and human review. An inconclusive result is not a concealed accusation, while a not-AI result is not a certificate of human authorship.

Mureka (Sonauto)

Begin with the workflow and records, then use audio classification as supporting information—not platform attribution.

Product context
Generated tracks with configurable style and vocal elements.
Useful evidence
Project history, exports and creator disclosure
Public detector output
Category, component indicators, confidence and timeline
Attribution supported
No

Start with the production path

Names, features and models can change as products are renamed or updated.

A rendered master flattens that history into one waveform. Classification can assess the waveform, but it cannot recover prompts, ownership, consent or the division between generated and performed work.

Evidence to request before drawing a conclusion

Record the product name, model version and export date at creation time instead of reconstructing them later from audio.

  • The original export and its date
  • Prompts, generation history or platform records where available
  • DAW sessions, stems and edits made after export
  • A direct account from the creator

What the report actually contributes

The current site sends the complete file to a specialist third-party service. It returns a primary category, separate vocal and instrumental percentages, confidence and sometimes window values. The application does not blend those fields into a platform score.

Smoothness, loudness or style are not reliable product identifiers.

Use the answer in proportion to the stakes

For curiosity, a result may be enough context. For moderation, contracts, copyright, employment or education, obtain records and human review. An inconclusive result is not a concealed accusation, while a not-AI result is not a certificate of human authorship.

Why we don’t name a generator

Identifying the specific tool behind a track would require a classifier trained on labelled output from every platform involved, kept current as each one ships new model versions. No one has published a model like that with credible held-out results, and we are not going to imply one exists by printing platform names on a result screen.

What these pages offer instead is context. Knowing that a given generator maximises loudness by default, or stitches sections together, tells you which parts of a report to trust and which to discount.

Keep reading