An AI music detector false positive is a human-made recording incorrectly classified as AI-generated. A high score alone is not proof of how a song was made. Preserve the tested file, check the tool’s classification scope and gather the recording history before changing the audio or challenging the result.

First establish what the detector is classifying

Keep two questions separate: who wrote the composition, and how the audio being analyzed was produced. A song written by a person but rendered by an AI generator contains generated audio. A flag on that export is not automatically a false positive about the recording.

Likewise, AI-assisted mastering, synthetic instruments, sampled performances and generated vocals describe different workflows. Record what actually happened. Do not infer a detector’s treatment of every hybrid case from its name or a single score.

Why false positives need real evidence

A 2026 research preprint on AI versus edited music evaluates edited recordings as a difficult comparison class and reports weaker performance on them than on the AI class. Its findings concern that research system and dataset, not a measured SH Labs error rate.

The earlier ICASSP paper on detection challenges discusses generalization and audio manipulation. These are reasons to examine real-world performance, not to conclude that every heavily edited human track will be flagged.

We do not publish an independently measured SH Labs false-positive rate here. A detector’s overall accuracy figure cannot by itself tell you the chance that your particular flag is wrong.

A simple example: accuracy is not the chance a flag is right

This is an illustrative calculation, not a benchmark. Suppose a collection contains 100 AI tracks and 900 human recordings. A hypothetical detector flags 90 of the AI tracks and incorrectly flags 18 human recordings.

It has 90% sensitivity and a 2% false-positive rate in this example. Yet 18 of the 108 flags are wrong: about 16.7%. The proportion of AI material in the collection matters, as do both kinds of error. None of these figures describes SH Labs or Demorphus.

What to save when a human recording is flagged

  1. Keep the exact flagged export and the full report, including date, score, verdict and detector version when available.
  2. Preserve original takes, stems and the DAW session. Keep earlier exports instead of overwriting them.
  3. Write a brief production history: recorded parts, sample sources, generated elements if any, and the processing applied.
  4. Check the same file with another identified detector if useful, keeping the results separate rather than averaging their scores.
  5. If a service acted on the result, send its support team the report and relevant creation evidence through its review process.

Changing a track until it scores lower would not establish that the first classification was wrong. For a recording-provenance question, the original files and session history are more useful than a newly altered export.

When detectors disagree

Check filenames and file hashes before comparing results. A WAV, an MP3 and a screen recording are not interchangeable test inputs. Record the tested duration and whether one service inspected only a portion of the track.

Different predictions are a reason for review, not a vote that settles authorship. Use the comparison checklist to keep the file, detector and date attached to each result. If the input is a known AI track that remains detected after processing, use the processing troubleshooting guide instead.

Check a file and keep its report

The free Demorphus AI music checker displays a SH Labs result for your uploaded file and lets you save the report. Use it as one piece of evidence alongside the production history.

A Human Made label is a detector verdict, not an authorship certificate. Read the score guide for the distinction before making a decision about someone’s work.

Sources reviewed October 11, 2026.

Prepared with AI assistance. Sources below.

Sources and documentation