How the checker works
Every report on this site is built the same way, from public data, with every step shown on the page.
Where the data comes from
We match each artist to its catalog on Deezer, which publishes every release with its date, label line and songwriter credits. We then look the act up on MusicBrainz, the open music encyclopedia, to see whether anyone has recorded where the act is from or when it started, matching the entry through its Deezer link where one exists so that two acts with the same name are never confused. We also look for a Wikidata entry carrying the same Deezer artist ID, which is strong evidence that real people have been documented behind the act. For a handful of acts that have been covered in the press we add a short "What has been reported" box with its sources.
The signals and their weights
Signals that point toward AI add points; evidence of a documented human act subtracts them. A total of 6 or more reads as strong AI signals, 3.5 to 6 as some AI signals, and below 3.5 the catalog looks like a human act. Acts with fewer than four releases read as too early to tell, because a debut single cannot show a pattern and "new, no credits, no records yet" describes every new indie artist too.
| Signal | Points | What it measures |
|---|---|---|
| Release burst at launch | 3 | Different release days in the first 90 days after the first release, with deluxe, remastered and sped-up versions counted once (4+ days or 8+ titles scores 2; 8+ days or 12+ titles scores 3). |
| Output pace | 2 | Distinct releases per month across the whole catalog (1+ scores 1, 2+ scores 2). |
| Catalog started recently | 1 | First release within the last two years. |
| No credited songwriters | 1 | No credited songwriters on the top releases, or only the act's own name. |
| Not linked to other artists | 1 | Deezer links the act to no other artists. |
| Thin public footprint | 1 | No MusicBrainz entry, or an entry with no home area and no start date. |
| Self-released | 0.5 | No label, or the label is the act's own name. |
| Described as AI on Wikidata | 3 | Wikidata describes the act as AI-generated or an AI artist. |
| Documented on Wikidata | -2 | Wikidata links this exact Deezer artist to an entry with a birth date, founding date, origin or members (minus 2). |
| Several credited songwriters | -1 | Two or more different people credited as songwriters (minus 1). |
| Long release history | -1 | First release five or more years ago (minus 1). |
How accurate is it?
We test every change to the scoring against a fixed set of acts before it ships. The set has two groups. The first is acts that have been publicly documented as AI-generated, in the press and on Wikipedia. The second is human artists chosen to be hard cases: some of the most prolific catalogs in music (Buckethead, King Gizzard & The Lizard Wizard, YoungBoy Never Broke Again, Merzbow), acts whose first release was in 2024 to 2026, and masked or anonymous acts like Sleep Token, Ghost and Daft Punk.
The human acts that scored highest, so you can see where the checker is weakest:
| Artist | Group | Points | Reading |
|---|---|---|---|
| Wishy | human | 1 | Looks like a human act |
| Jinsang | human | 0 | Looks like a human act |
| sombr | human | 0 | Looks like a human act |
| Teddy Swims | human | 0 | Looks like a human act |
| Ghost | human | 0 | Looks like a human act |
| Djo | human | -0.5 | Looks like a human act |
| Kidz Bop Kids | human | -1 | Looks like a human act |
| Chezile | human | -1 | Looks like a human act |
A small test set cannot prove the checker is right about any one artist, and new kinds of AI project will appear. That is why every report shows its evidence instead of only a verdict.
Why these signals
Generating a song with a tool like Suno or Udio takes minutes, so AI projects tend to arrive all at once: an "artist" with no history uploads an album, then another, then a run of singles, all inside a few weeks. Human acts almost never do this, because writing, recording and mixing take time. In our test of 77 human acts, the typical act released music on just one day in its first 90 days, and none put out more than five titles in that window. The five documented AI acts released on 4, 8, 12, 15 and 24 different days. Some AI projects also upload in batches, a dozen singles on the same afternoon, so the checker counts titles as well as days.
Songwriter credits matter for a different reason. Human songwriters register their songs so they get paid, and those names show up in the credits. AI projects usually list nobody, or only the act's own name.
What this cannot tell you
A catalog reading is evidence, not proof. A prolific human producer can trip the release signals, and a careful AI project can avoid them. We never state that an artist is AI on the strength of our own reading. If you represent an artist and think a reading is wrong, see corrections.
Why we do not run an audio detector
We tried. The open detectors we tested misread short previews badly, scoring Sleep Token and Morgan Wallen as likely AI. The full results are on our detector test page.