Does this sound like AI?
Something felt off and now you cannot unhear it. This page is the honest version of that question: which cues actually hold up, which ones are just modern production, and what to check before you say anything out loud.
Last updated August 2026
Short answer
Does this sound like AI?
Trust the impression as a prompt to check, not as a conclusion. Ear-based judgement reliably catches incompetent generated audio and performs poorly on competent generated audio, and it produces false alarms on loud, quantised, heavily produced human music. Run the audio through a measurement, read which measurement drove the score, then ask the person for stems — provenance outranks everything acoustic.
What this cannot establish →First: what kind of audio is it?
- A full song — the free AI music detector on this site is built for exactly that. It runs in your browser, nothing is uploaded, and it shows you the reasoning behind the number.
- A voice, a voice note or spoken words — the wrong tool. Read how to tell if a voice is AI generated instead.
- A short clip from a social video — get the original file first. A re-capture measures the encoder more than the music.
Cues that tend to hold up
None of these is proof. Several of them together, in the same recording, is a reason to ask a question.
- Lyrics that scan perfectly and name nothing. Correct meter, generic imagery, no specific place, person, date or private detail anywhere in three minutes.
- Section joins that are too exact. Transitions that land on the bar with no fill, no push, no player deciding anything.
- Instruments that never change articulation. The same guitar attack, the same hi-hat, the same piano touch from first bar to last.
- Uniform delivery. Repeated lines phrased identically, no breath variation, energy that never builds or sags across a verse.
- Tonal balance that is identical everywhere. Verse, chorus and bridge with the same spectral shape, as though the whole thing was rendered in one pass.
Cues that fool almost everybody
These are the reasons careful people accuse human artists. Each one is a production choice made on the majority of commercial releases.
- Loud and flat. Low crest factor means aggressive mastering, not synthesis.
- Nothing above 16 kHz. That is a lossy encoder’s frequency ceiling. It says how the file was compressed, not how the music was made.
- Perfect timing. Quantised programming has been standard for thirty years.
- Perfect pitch. Pitch correction is on nearly every commercial vocal.
- Repetitive structure. Genre convention in template-driven pop, house, drill and library music.
In our own measurements of human synthesiser renders, three of twelve read above 70% — see the electronic false-positive study. If the track is techno, trance, hyperpop or a game soundtrack, raise your threshold before concluding anything.
Then check the things that actually carry weight
- Provenance. Stems, a project file with a plausible edit history, an unmastered rough, dated drafts, session video. Any one of these outranks every measurement on this site.
- Release context. Twelve releases in a month across unrelated genres, with identical mastering and no live footage, tells you more than any spectrum.
- Metadata. Tags often carry encoder strings and tool names nobody thought to clear. Content Credentials or a C2PA manifest is cryptographic provenance and beats statistics outright.
- Acoustic analysis. Run it on two different excerpts and compare — that is the largest measured source of movement in a reading.
- Your ears. Recorded first, weighted last, because a score anchors perception hard.
The full ordered process, with the mistakes left in, is written up as a ten-step walkthrough.
How to say it without overstating it
“The audio shows characteristics associated with generated music, and no provenance was supplied on request” is accurate. “This song is AI” is not, and in a dispute it is the sentence that gets quoted back at you. Where your inputs disagree, undetermined is a real result, and most single-track investigations honestly end there.
Questions people ask
Does this sound like AI — can a tool just tell me?
A tool can give you a probability from signal properties, not a verdict. Our detector reads full musical mixes and reports how AI-like the measurements are, along with which measurement drove the score. That is useful as one input among five: provenance, release context, file metadata, acoustic analysis, then your own ears — in that order of weight.
What makes music sound artificial even when it is human?
Loudness-maximised mastering, quantised programming, sample libraries reused across sections, template arrangements, and heavy pitch correction. All five are ordinary production choices, and all five push acoustic measurements in the same direction as generated audio. Electronic and synthesiser-based human music is the single biggest source of false positives we have measured.
I only have a short clip from a video. Is that enough?
Usually not. Screen recordings and short clips lose the top octave, the micro-dynamics and the stereo relationship that analysis depends on, and excerpt choice alone moved our readings by about 4.4 percentage points on average in robustness testing. Get the original file and forty-five seconds or more of the full arrangement.
Can I tell which generator made it?
No, and neither can we. Nothing in the audio reliably identifies a specific product, and any tool naming one from sound alone is guessing. Attribution comes from metadata, provenance manifests or an admission — never from a spectrum.
It sounds like AI and the artist will not explain. What now?
You have an absence of documentation, which is not the same as proof of generation. That is a reasonable basis for a private commercial decision and not a basis for a public accusation. Ask for stems, a project file or dated drafts first; a working musician can usually produce them in minutes.
Ready to measure it? Run a free check in your browser — no account, no upload, nothing stored.