demorphus
AI music detectors can give different results because they use different models and decision rules, or because they analyzed different audio. Start by checking the file and the scope of each report. A disagreement alone does not establish which detector is correct.
First check whether the audio was actually the same
A song title is not a file identifier. A Suno download, a streaming preview and a mastered WAV can all share that title while containing different audio. Record the filename, duration, format and SHA-256 hash of each upload. Matching hashes identify identical file bytes; different hashes do not necessarily mean the sound changed, because metadata can change too.
Keep the reports beside the files. If you compared an excerpt with a full export, begin with the full song versus sample guide. If the result changed after a new render, use the export checklist.
Compare equivalent report fields
Do not compare a generator label with an overall AI score, or a confidence field with the probability of AI generation. SH Labs documents separate fields for the overall probability, prediction, confidence and likely generator. Preserve their names when copying results.
| Record | Question to resolve |
|---|---|
| Input file and duration | Did both services receive the same audio? |
| Analysis scope | Was this a full upload, a preview or a selected passage? |
| Metric and scale | Are both values measuring the same thing? |
| Verdict | What classification did each service actually return? |
| Date and model version, if supplied | Were the reports produced under comparable conditions? |
Why identical uploads can still receive different predictions
A classifier learns from its training material and applies its own rules to new audio. Two products need not recognize the same cues or draw their decision boundaries in the same place. Without documentation, do not assume that a particular service uses a watermark reader, a specific frequency band or another company’s model.
Afchar and colleagues identify robustness to audio manipulation and generalization to unfamiliar generators as challenges. That paper explains why validation matters; it does not identify the better commercial detector for your song. For the evidence needed to compare accuracy, see our detector evaluation guide.
Run a comparison you can interpret
- Choose one unchanged file and keep it as the reference.
- Submit that file to each checker and save the complete reports, including errors.
- Record analysis scope and model information where provided; mark missing fields as unknown.
- If investigating inconsistent results within one service, repeat the unchanged input before trying an edited version.
- Report the disagreement, rather than replacing both results with the lowest number.
Do not average unrelated scores into a new probability unless that combination has been validated. Two websites using the same underlying service also do not provide two independent assessments.
What to do with a disagreement
For a known AI source, keep the production history and use the reports to compare classifications. For a human recording, retain the session and original takes and follow the false-positive checklist. SH Labs describes its outputs as probabilistic predictions, so a report should not replace production evidence.
To establish a baseline with Demorphus, upload the exact file to the free AI music checker. If you subsequently use Cleaner, compare before and after with the same engine and record listening quality separately.
Sources reviewed October 30, 2026. The comparison procedures are suggested workflows, not new benchmark results.
Prepared with AI assistance. Sources below.