In February 2025, the Beatles’ “Now and Then” won the Grammy for Best Rock Performance. It is an arrangement built on a cassette John Lennon recorded at home in New York in the late 1970s: piano and voice, nothing else. The other three had in fact tried to finish it in the 1990s, but the technology of the time couldn’t pull his voice out from under the piano, and they gave up. Then, in 2021, a machine-learning tool built by Peter Jackson’s team separated the voice from the piano; Paul McCartney and Ringo Starr added their parts, and the song was released in 2023. McCartney has said it again and again: nothing was synthesised. The voice was separated, not made.
In September 2024, SOPHIE’s self-titled album was released. When she died in January 2021 it was nearly complete; her brother Benny Long, who had worked beside her on it for years, finished the rest.
In June 2022, three songs disappeared from Spotify and Apple Music — tracks from Michael (2010), the first album released after Michael Jackson’s death. For years his mother, his children and others had doubted that the voice on them was his. The estate that manages his music and Sony said the removal had nothing to do with authenticity; it was “the simplest and best way to move beyond the conversation.” Jackson’s posthumous music is held by an estate run by the executors named in his will; in 2024 Sony bought half of the catalogue for more than $600 million, and the two now run it together.
On New Year’s Eve 2019, on NHK’s Kōhaku song contest (Japan’s biggest annual TV music event), Hibari Misora — the biggest female star singer of postwar Japan, who died in 1989 — sang a new song. The voice was built by AI; the lyrics were ones she had never sung. Her family had agreed. The reaction split anyway: “I got to hear her again” and “this is desecration,” side by side on the same night.
This is what I keep snagging on. Using the voice data of a singer who has died to release a new song — is that all right? And whose voice is it, exactly?
What Hiroki Azuma wrote
A book from 2001 helps here: Otaku: Japan’s Database Animals, by the philosopher Hiroki Azuma (English translation by Jonathan E. Abel and Shion Kono, University of Minnesota Press, 2009). Using otaku anime consumption as his example, Azuma wrote this.
People used to receive a work as a single story — an author, a beginning and an end, something it wanted to say. In the nineties that changed. People began pulling only the elements they liked out of the “database” behind the work. A character could be broken down into parts — cat ears, a maid outfit — and once you had the parts, you could build the next character without an author. Azuma called this “database consumption.”
Today’s version is the Spotify playlist. Chill, Rainy Day, Focus. You take the element — a feeling — without registering whose song it is. Fewer people listen to an album as an album; more listen by tag. Azuma wrote this before TikTok and before Spotify existed.
Azuma described the consumer. IP is the owner’s side of the same thing
But Azuma was writing about how people receive. The database of 2001 was something fans had built on their own; it belonged to nobody.
What I’ve been sensing lately — call it IP-ification — is the same phenomenon seen from the other side. If something can be broken into elements, each element can be owned. If it can be owned, it can be run. And the people who run it are not consumers but whoever holds the rights.
The easiest rights for a third party to run are those attached to an artist who has died. No new work. No broken contracts. No ageing. No objections. Michael Jackson — the moonwalk, the white glove, that voice, the red jacket — is a bundle of elements you can consume completely without him. The estate and Sony keep him in content after death so that today’s teenagers meet him, and so that those teenagers, grown up, will buy him again as the sound of back then. Nostalgia, planted in advance.
The reason isn’t mysterious. Not “we can’t imagine a future” — just that it’s easy. Building recognition for someone new costs more than recycling what everyone already knows.
From use to inheritance
Artists have been used by power in every era — Rome, the aristocracy, wartime propaganda. That’s a sad story. But I don’t think “the artist won because the name survived.” What survived is the work. The person wasn’t repaid.
What’s changed, I think, is the form of the use. It used to be commissioning: at least the artist was in the room. Now it’s holding. When Hibari Misora is made to sing a new song, she isn’t at the table.
Line up the four cases from the top of this piece by how much the artist was actually present, and you get a ladder. Lennon sang and recorded it himself; the other three, as far as we’re told, only made the recording audible. SOPHIE had nearly finished the music, and the brother who had been working beside her completed it. On the three Jackson tracks, nobody could be sure the voice was his at all. For Hibari Misora, neither the voice nor the words had ever come from her. Each step down, the artist is further from the table. The yardstick for what counts as “her work” seems to be that distance, not the law.
Rights: what permission doesn’t cover
The law, briefly (I’m not a lawyer, and this was checked before publication). In Japan, a voice as such isn’t protected by copyright. The right to control commercial use of one’s name and likeness — publicity rights — derives, in case law, from the personal rights of a living person, and the prevailing view is that it doesn’t extend to the dead. So the voice of someone who has died is close to a legal blank in Japan; in practice, the family and the rights holders negotiate. The United States varies by state; Tennessee passed the ELVIS Act in 2024, explicitly covering AI-generated voice.
Even with the law settled, though, I doubt the snag goes away.
Re-releasing a recording the singer once approved is one thing. Making her say words she never said, in her own voice, is another. For the second, “consent” can’t even be formed — she died before the technology existed.
And the family’s permission doesn’t fill that gap. When the Anthony Bourdain documentary Roadrunner (2021) used an AI version of his voice, the director said the estate had approved; his ex-wife said otherwise. Nobody could answer who gets to speak for the dead. With Hibari Misora, the family had agreed, and the reaction split anyway. What people were objecting to wasn’t the absence of permission.
Ownership and inheritance can’t be separated from how we see death
This is the part I most want to think about. The very idea of inheriting a voice as property assumes one particular view of death.
Japanese law already contains a different assumption. Article 897 of the Civil Code: genealogies, ritual implements and graves pass outside the ordinary estate, to whoever presides over the family rites. They aren’t divided among heirs. The law itself concedes that what belongs to mourning can’t be split by the logic of property.
Is a voice property, or is it closer to the family altar? Practice treats it as property. But maybe the intuition that refuses is answering to the altar.
One more thing. Japanese mourning had forgetting built in. Japanese Buddhism has a custom of memorial services at set years after a death — the first, second, sixth, twelfth and so on — at which a priest is called to pray for the dead. The last of these is the thirty-third memorial, held in the thirty-second year, and in most schools that is where the individual rites end and the dead dissolve into the collective “ancestors” — the form the folklorist Kunio Yanagita described in About Our Ancestors (1946). Preserving a proper name forever was never part of the design.
What’s happening now runs the other way. Grave closures, tree burials, scattering ashes: the remains are being let go. Meanwhile voices and faces are managed as assets and run with no end date. We’re releasing the body and keeping the data.
Letting go stopped being automatic
Technology has made feelings we ought to release harder to release. More precisely: letting go stopped being automatic.
Things used to disappear on their own. Photographs faded, letters got lost, a voice lost its detail in memory. The rate of decay was the rate of mourning. Digital doesn’t decay, so time no longer does the work. Deleting the chat history of someone who has died, closing the account — these became jobs someone has to decide to do.
But not being able to let go was always the human default. The thirty-third memorial isn’t a natural event; it’s a deadline people set, precisely because left alone we don’t let go. “Let’s end it here.” What broke isn’t the capacity. It’s the device.
Not who owns it. When it ends.
So maybe what’s needed isn’t a new right but an end date — the way copyright has a term — for voice data too. Design for when it ends, not for who owns it. The language of mourning, not the language of property.
Maybe making the dead sing a new song isn’t something to judge only on whether it’s right or wrong. The one thing I can say is that a system where “the family agreed” ends the conversation is dropping something. What it drops is the fact that the singer isn’t at the table — and the feelings of the people who knew her alive.