Musk and the Mask

In The Atlantic yesterday, staff writer Megan Garber reviewed director Alex Gibney’s new documentary Musk, a portrait of a man whose influence has grown to something approaching systemic power.¹ The film isn’t out yet, and I haven’t seen it, but one device that Garber describes interests me more than the portrait itself. When trillionaire Elon Musk declined to participate, Gibney’s team compiled a database of his recorded public statements and used it to create an animated version of him. The avatar speaks Musk’s actual words, but in contexts where he was never asked the questions it appears to answer. According to Gibney, the result is intentionally ironic: a simulation of a man who has himself speculated that reality might be a simulation.²

»The avatar borrows words; the model invents them. But both are attempts to answer the same question – what would this person say now? – and neither has access to the one thing that could actually answer it: the person’s own encounter with a new situation«

This isn’t just a documentary using archival footage or a montage of quotations. The avatar gives the words a body, a voice, and the appearance of presence. It makes it seem as if Elon Musk is there responding, when in fact he wasn’t there and didn’t respond.

This made me think of British sculptor Marc Quinn’s Self, a series of casts of the artist’s head made from his own frozen blood.³ The sculpture is constructed from Quinn and resembles Quinn, yet it is clearly not Quinn. It can’t act on his behalf. Gibney’s avatar is more complicated because it can speak. Its words are really Musk’s, but someone else turned them into answers to questions he was never asked. The result is neither Musk nor not Musk. It’s a version of him assembled from material he supplied, made to perform in a different context.

This is where the device opens onto something larger, even though the mechanism isn’t quite the same. Gibney’s team selected and rearranged words that Elon Musk had actually said. Today’s large language models do something stranger: give them enough of a person’s writing, speeches, and interviews, and they learn a pattern they can extend into sentences that person never actually said. The avatar borrows words; the model invents them. But both are attempts to answer the same question – what would this person say now? – and neither has access to the one thing that could actually answer it: the person’s own encounter with a new situation.

A living person responds to changes in the world. But if a machine were trained on Stephen King’s novels and continued producing novels after his death, it would be drawing on a body of work formed in response to a world that no longer exists. King’s response to whatever came next might have differed from anything in the training material. No statistical refinement, however exact, could supply that missing relationship with the present. The model could produce a genuinely new novel, but it couldn’t establish that King would have written it, let alone that he did.

Would it be wrong to publish such a novel? Not necessarily, provided its origins were clear. The question is whether it should be presented as a novel by King. A work in someone’s style and a work authored by that person are different things. Authorship involves making particular choices, taking responsibility for them, and presenting the result to readers under one’s name. That’s not the same as saying that machines can’t create, or that they can’t be creative. It’s an argument against confusing the capacity to generate with the identity of the person being imitated.

The same gap seems to exist in Gibney’s film, even though Elon Musk is alive. The wealthiest person in the world continues to respond to a changing environment. His recorded words, detached from the circumstances in which he spoke them and inserted into new ones, can’t stand in for those future responses. Are the avatar’s words authentic? Yes. But real words don’t make the performance real. And what about questions of permission and compensation when someone’s recorded life becomes the raw material for a synthetic likeness?

The word »persona« offers a fitting image here. Its deeper etymology is disputed, but it might have come to us through the Latin word for an actor’s mask.⁴ The mask is shaped as a face, but someone else can animate it, and the voice spoken through it doesn’t have to belong to the person it represents.

Musk’s avatar is a mask made from Elon Musk’s own words. The question isn’t whether those words belong to him. It’s who gets to make them speak.

References