Songlinks › We asked an AI about 150 songwriters
We asked an AI about 150 songwriters
It knew the hits. It didn’t know the catalogue, the working songwriters, or the numbers royalties are paid on.
People in the music business have started asking AI assistants who wrote what. So we tested one. We took 150 songwriters at random from public music rights-registration data, the records US streaming royalties for songwriters are paid against: half of them famous, half working songwriters with real, recorded catalogues that nobody writes articles about. We asked a leading general-purpose AI assistant what each of them wrote, how many works they have, and their IPI name number, first from memory and then with web search. Then we checked every answer against those registrations. The questions and the pass marks were fixed before it answered.
What it got right, and what it didn’t
From memory. With web search the picture barely moved: 74% of named songs correct, 1.8% of the catalogue, still 0 of 680 for working songwriters, and 1 IPI number of 150.
It didn’t make things up. It had nothing to go on.
The striking part is not invention. For the writers it didn’t know, the assistant mostly said so: it gave no IPI number for any of the 150 rather than guess one. For famous writers it named the songs everyone knows, and most of them check out. But a famous writer’s hits are a sliver of what they have registered, and royalties are paid on every registered work, not the ten songs anyone remembers.
For working songwriters there was nothing at all, even with search: not one of their 680 most-recorded songs. These are exactly the writers whose registrations most need finding, and they are invisible to a tool that learns from what has been written about music.
Why: one songwriter is many records
The registry does not hold “a songwriter”. It holds registrations, each a name as somebody once typed it: a legal name, a stage name, a middle initial or not, an apostrophe or not, “Jr.” in the surname field, more than one IPI name number or none. Deciding which of those records are the same human being is a question about the registry’s own evidence — shared co-writers, publishers and works — not about how much text a model has read.
That is the one question Songlinks is built for. It uses a model made of decision trees, trained for that single decision, which joins records only where the evidence does, says when it cannot settle a pair, and shows the works and collaborators every join rests on.
Watch how a decision is made → · or look up a writer on songlinks.io
How the test was run
| What | Detail |
|---|---|
| Writers | 150 drawn by a fixed rule from public registration data, in three separate samples of 50. Famous: a writer on a work with at least 500 recordings. Working songwriters: 5 to 200 registered works, at least one recorded, none with more than 3 recordings. |
| The assistant | One leading general-purpose AI assistant, asked from memory (no tools) and, separately, with web search. It was shown each writer’s registered name only. |
| The key | The registrations themselves: a song counts only if it is registered to that writer. Songs the registry does not confirm are not counted as wrong, only as unconfirmed. |
| Pass marks | Written down before any answer existed. Each finding above held in all three samples, from memory and with search. |
| Date | October 2026, against the registration data snapshot of 2026‑10‑09. |
What this does not show. It is one assistant, so it says nothing about AI assistants in general, and a different model or a later version may do better. The 145 of 150 for Songlinks means the registered writer was among its answers; it is not a measure of how accurately it groups records. The full method and every figure, with confidence intervals, are available on request.