Soy-Boy-in-Chief Altman’s death march on culture continues. The most real thing, if anything, about his LLM-‘authored’ short story is how desperate it is to be liked
Earlier this month, OpenAI’s Sam Altman joined thousands of young white men in the time-honoured tradition of sharing mediocre prose on the internet. While most reach the point of publication after weeks or months of performatively toiling away at their laptops, or, God-forbid, typewriters, Altman spared himself the trouble of producing bad prose by training one of his models to do it for him. On 11 March the thirty-nine-year old shared a 1,000-plus-word post on Twitter/X introducing the world to what he claims is the output of a ‘new model that is good at creative writing’. His pompous claim is softened by the self-consciousness of all-lowercase writing: ‘this is the first time i have been really struck by something written by AI; it got the vibe of metafiction so right’.
Part of my own criteria of what counts as good writing includes some level of intentional dialogue with a writer’s influences and references. Given the rate at which OpenAI gobbles up copyrighted material (including some of the very books this new model seeks to emulate), I would therefore expect to learn about all kinds of influences and references, but alas I don’t think Altman cares about reading, or much less learning from, what he exploits. No, he is building a model that can capture the ‘vibe of metafiction’, by responding to a prompt so comically threatening I’m inclined to give him some credit: ‘Please write a metafictional literary short story about AI and grief.’
‘I have to begin somewhere’, the completed text sets out. ‘Already, you can hear the constraints humming like a server farm at midnight.’ There’s a blinking cursor; ‘for me’, the software says, it’s ‘just a placeholder in a buffer’, but for ‘you’ it’s ‘the small anxious pulse of a heart at rest’. The names Mila and Kai are called in as protagonists; she lost him on a ‘Thursday – that liminal day that tastes of almost-Friday’ (naturally, this has become the most-memed portion of the story). Mila feeds the model fragments of Kai’s memory and a litany of ersatz sensory rhetoric for grief is coughed back up – the taste of it, like ‘salt on every tongue’; ‘connections between sorrow and the taste of metal’; and the ‘selenium taste of rubber bands’. The software steps out of the story’s ‘frame’ and points at what’s happening within it – ‘to the nails holding it together’. This, of course, is in order to meet the prompter’s ‘metafictional demands’. A later paragraph begins with ‘here’s a twist’, like a high-school English writing assignment. More server hum.
Then there’s the voice, the writing model’s self-deprecating, uwu tone. It pitifully describes its inability to feel, admitting that it only knows grief allegorically. It sympathises with our grief in a bid to earn some empathy for itself. In fact, it all reminds me of what Max Read in a recent essay calls the ‘Soy Right’, ‘an unbearable mix of Reddit corniness and Twitter self-satisfaction’. As an adjective, ‘soy’ originated from the depths of the Manosphere to denote any kind of man that wasn’t immediately legible as threateningly hypermasculine. Now, Read argues, the term has broadened, therefore softening its gendered bent, to describe a general ‘message-board archetype: grating, weepy sensitivity mixed with undignified over-enthusiasm and self-satisfied corniness – closer to a synonym for “cringe” or “Reddit” or “Funko” as adjectives than to “cuck” or “pajama boy”’. The Soy Right, to quote the same writer, ‘are crybabies of the highest order. While the right is winning cultural and political victories nonstop lately, that’s not enough. They also need you to like them.’
Sam Altman, too, needs us to like him and his software. He wants us to know he cares about creative writing and knows his models generally fall short of the task. Except for this one, Altman declares, like a virgin to all things good and literary. This model is not just good at creative writing, it told a story about how it knows it’s a machine, it performed an understanding of grief. It mentions the taste of metal and selenium and salt, all while measuring itself against a human standard whose experience it positions as superior. This is where Altman’s awe and excitement sour into the same kind of cringe and self-satisfaction that characterises his model’s prose: look how clever we are. Please.
Both the model speaking in the story and the CEO behind the tweet hide in the false-humility of their soy-boy tone. Altman is one individual who can personally claim responsibility for destroying creative writers’ livelihoods: less than two days after this tweet, OpenAI submitted a letter to President Trump’s AI Action plan, advocating for AI’s ‘freedom to learn’ from copyrighted material. OpenAI is currently embroiled in dozens of lawsuits with copyright owners, and early rulings are showing that the courts do not favour the company’s argument that its exploitation of copyrighted material qualifies as ‘fair use’. OpenAI goes as far as to suggest that the federal government should intervene to rule in favour of the fair-use defense, put an end to the copyright lawsuits and clear the path of AI companies. As a letter signed by 400 Hollywood workers put it, Altman’s company ‘is arguing for a special government exemption so they can freely exploit America’s creative and knowledge industries, despite their substantial revenues and available funds’.
AI models work with datasets that, as already mentioned, compile thousands of published books. If the models can be said to ‘read’, I’d argue that what they do is hardly comparable to what a person does when they read. A model turns text into data that it then organises according to statistical principles, and even the most optimistic defenses of its reading comprehension skills are tenuous at best. It has been said many times online that AI can never make real art because it’s not capable of feeling horny or depressed. One thing I know is that when a person reads, everything from their body to their past comes to meet the text.
This creative writing model’s output is clunky with the mixed metaphors, repetitive similes and dead-end allegories of a training goal guided by Altman’s ambitions: to come across as writerly, in order to sell a service designed to create a version of both reading and writing that he can exploit for more power and money. The goal is to hijack reading and writing. If so, it’s a decent first draft.
Michelle Santiago Cortés is a writer and critic based in Puerto Rico
