demorphus
If your song is still detected as AI after processing, the tested export still meets that detector’s criteria. It does not necessarily mean the treatment failed to run. First confirm that you uploaded the processed file, then compare the score and verdict with the original under the same test conditions.
Check which file produced the result
Save the source and output with different names, then select the output deliberately. Preview it before uploading. A browser download suffix, a duplicate in your Downloads folder or a later DAW export can make two files easy to confuse.
The Demorphus checker’s downloadable report includes a SHA-256 file hash. Matching hashes identify the same file bytes. Different hashes show that the files differ, but do not prove the audible signal changed: metadata edits can also change a hash. Note the duration and format alongside the filename.
Read the score, verdict and generator label separately
| Observation | Interpretation | Next check |
|---|---|---|
| The score fell, but the verdict stayed AI | The result changed without crossing that system’s decision boundary | Save both reports and compare the sound |
| The predicted generator changed | Attribution changed; this is not automatically a Human Made verdict | Read the overall classification separately |
| A second detector disagrees | The systems returned different predictions | Use identical file bytes and record each result |
| The analysis timed out or was unavailable | There is no usable verdict | Resolve the analysis error before drawing conclusions |
A score is not the percentage of notes made by AI. For definitions and report fields, see how to read an AI music score.
Why processing may leave recognizable cues
SH Labs explicitly lists processed and hybrid content among its detection capabilities. A master or edit can sound different while still leaving enough information for its classifier.
Research on AI music detection also identifies audio transformations and unfamiliar generators as robustness challenges. That supports testing the exact output rather than predicting its result from the name of an effect. It does not establish a recipe that works on every detector or song.
Use the source and the export you actually intend to keep
For a Cleaner trial, use the generator’s original WAV when you have it. That is our recommendation from internal use, not a promise that WAV alone lowers detection. Converting an existing MP3 into WAV does not restore lost source information. The WAV versus MP3 guide explains the distinction.
Submit the complete final file and check the service’s analysis coverage, rather than assuming a short excerpt covers the song. See the full song versus sample guide for duration limits. If you master, normalize or re-encode after a successful check, save and test that new export when you need a result attached to it.
Separate a retry from another processing generation
Two files both called “cleaned” may come from different routes. Label a fresh trial from the original as a separate candidate; label processing applied to an earlier output as another generation. Keep the parent filename with each render so you can reconstruct the chain.
If your goal is to compare treatments, return to the unchanged source. If a particular finishing edit is necessary, preserve the pre-edit file and check the new export separately. Use the export checklist when you cannot explain a changed result.
What to do after an unsuccessful Cleaner result
Open the job in your account and confirm that the report belongs to the downloaded output. Demorphus selects that output automatically; the report retains the actual score and verdict even when it remains classified as AI.
Keep the job reference, file hash, timestamp and report if the result differs from the interface. Avoid repeatedly processing an already processed file: each pass creates another audio generation. If you originally supplied an MP3 and still have the original generator WAV, that source is a better starting point for a separate trial. It is not a guaranteed pass.
Our internal results include unsuccessful classifications. For the normal upload-to-download process, read the Suno fingerprint remover guide. If you have only used a checker so far, the checker versus Cleaner guide explains why analysis alone does not alter the file.
What if the track was genuinely recorded by people?
A high score on a known AI export and an incorrect flag on a human recording are different situations. If your track is a human recording, keep the original takes, project and export history. Follow the false-positive review checklist rather than treating sound alteration as proof of authorship.
If you are comparing SH Labs with Suno’s credentials checker, they do not answer the same question. Suno’s tool checks its C2PA provenance metadata; an absent credential does not establish a human recording.
Sources reviewed October 30, 2026. Listening and export checks are suggested procedures, not new benchmark results.
Prepared with AI assistance. Sources below.