You find a track on YouTube that feels slightly wrong. The hats shimmer oddly, the vocal consonants seem soft, and the second chorus is suspiciously identical to the first. "It sounds off" is not a method.
Other producers are noticing the same thing. A thread titled "Internet is dead" in the Ai for Music section of the Audiosex forum carries the tags "junk" and "youtube" (source). The thread lists no test method and no accuracy data, and this article makes no accuracy claims for any detector. Instead, it gives you a repeatable exercise that trains your critical listening and produces a checklist, not a verdict.
What this test can and cannot do
No ear test or analyzer reading proves a track is AI generated. Every cue below has a human explanation, from lossy codecs to brickwall limiting to a producer who loves copy-paste arrangements.
Treat the exercise as evidence gathering. You score how many cues line up, how strongly, and how confident you are in each one. Then you combine that with the non-audio checks at the end.
Step-by-step: the blind A/B exercise
Prepare your audio
Start with calibration. A cue only means something if you know how it behaves on tracks whose origin you already know. Gather a few human-made tracks you trust, plus any AI-generated tracks you made yourself or whose origin you know, in a genre similar to the suspect's. Use only audio you have the right to use.
Match the format. If the suspect came from YouTube, it went through a lossy codec, so run your known tracks through the same path (or an equivalent lossy encode) before comparing. Comparing a lossless reference with a lossy suspect produces false positives by design.
Run the exercise
- Set up the project. Create a session at 44.1 or 48 kHz and import every file onto its own audio track. Pick the same musical section from each, such as the first chorus, and loop 20 to 30 seconds of it.
- Match loudness. Put a LUFS meter on the master and read short-term loudness for each looped section. Trim with clip gain or a utility plugin until every track sits within about 0.5 dB of the others. Trim before any limiter, and check true peak so nothing clips. No meter? Match by ear, flipping between tracks until neither one feels louder.
- Add your analysis tools. Insert a spectrum analyzer on the master. A reasonable start is an FFT size around 8192, moderate averaging, and a slope of roughly 4.5 dB per octave. Add a utility or mono button you can toggle, plus a correlation meter or goniometer if your DAW has one.
- Blind the files. Rename each file with a random number, or have a friend do it and keep the key. Put the suspect and your known tracks in the same pool, and drop the clips in alphabetical order so you cannot tell which is which. Alternatively, let a second person control the track mute buttons while you listen without looking at the screen.
- Listen first, look second. Make two or three passes with the analyzer hidden or minimized, and fill in the checklist by ear. Then reveal the analyzer, check the mono toggle, and add your visual notes.
- Reveal the key and check your known tracks. Do this only after you have scored every clip. If the cues you flagged as "AI-sounding" also appeared on your known human tracks, those cues are unreliable for you in this genre. Cross them off.
- Compare the suspect. Weigh its totals and confidence ratings against your calibration set, using only the cues that survived step 6.
Keep listening volume moderate and take breaks. Ear fatigue can make everything sound artificial. We cover related ground in serato sample alternative: chop breaks in your daw.
The listening checklist
Copy this into a document or print it. Score each cue 0 (absent), 1 (possible), or 2 (clear), then rate your confidence from 1 to 5.
| Cue | What to listen or look for | Could also be | Score (0-2) | Confidence (1-5) | |---|---|---|---|---| | Transients | Drum and pluck attacks that feel soft, smeared, or oddly uniform from hit to hit | Lossy codec pre-echo, heavy limiting, transient shapers, tape or saturation | | | | High frequencies | Grainy, swirling, or metallic content in the top octaves; a hard shelf where the analyzer drops off | Codec lowpass, vinyl or tape emulation, exciters, a dull source recording | | | | Stereo image | Wide elements that collapse or change tone in mono; phase wobble on the correlation meter | Stereo wideners, chorus, deliberate wide pads | | | | Vocal consonants | S, T, and K sounds that blur, lisp, or vary strangely; missing or repeated breaths | Heavy de-essing, pitch correction, low bitrate, budget mic chain | | | | Arrangement repetition | Sections repeating with no variation in fills, timing, or performance; abrupt structure with no development | Loop-based genres, copy-paste DAW workflow, sample packs | | | | Mix consistency | Instruments that shift in level or character without a musical reason | Automation, sidechain, aggressive mastering | | |
A single high score proves nothing. Several cues scoring 2 with decent confidence, while your known human tracks scored low, make a stronger signal. It is still a signal, not proof. This pairs well with go deeper on internet speed for remote music collaboration explained.
Why these cues give false positives
Lossy compression is the biggest confounder. It can lowpass the top end, smear attacks, and add swirly artifacts to cymbals and breaths, which are the same things people blame on AI. That is why the preparation step demands a matched format.
Heavy mastering is the second. Aggressive limiting flattens transients and can make a whole mix feel uniform. Stereo wideners can fail the mono check on perfectly human records.
Arrangement repetition is the weakest cue in loop-based genres. Plenty of human-made electronic and pop tracks repeat sections exactly, so score that cue lightly unless the genre normally demands variation.
These are inferences from how the processes work, not measured results. Your calibration set is how you check them against your own ears and your own genre.
Provenance checks that matter more than your ears
After the audio test, gather evidence that has nothing to do with waveforms.
Also read: see rotary dj mixer vs standard mixer: which suits you?
- Credits. Look for named writers, performers, engineers, and a label or distributor. Missing credits are a flag, not proof.
- Release history. Check whether the artist has a catalog that predates the suspect track, with a consistent sound and a real presence across platforms.
- Stems or project files. Ask for them. A human producer can usually send stems or a screenshot of the session. Refusal proves nothing, but a quick yes counts for something.
- Live or process footage. Performance clips, studio videos, and sample-credit documentation help, though they can be faked too.
- Platform labels. Note any AI disclosure the platform shows, and remember that the absence of a label does not mean a track is human-made.
If a sample or reference matters to your project, the decision is simple. Use tracks with clear provenance, and skip the ones where your checklist and your provenance checks both come up empty.
What to expect next
Your first few rounds may show how often your gut reacts to codec damage and loudness differences instead of anything about the music. After a handful of blind sessions, you should have a personal list of cues that held up on your known tracks and cues that did not. Keep that list and update it as tools and genres change. Treat every conclusion as a confidence level, not a yes or no.
