When Erin Hoover, the poet laureate of Cookeville, Tennessee, was getting her PhD in creative writing, her professor used to ask her: “Why did you write this?” Were you to ask an AI model today the same question – writing requests account for about a quarter of all prompts to ChatGPT – its answer might, truthfully, be: “Because Erin taught me to.”
Alongside her work as a poet and professor of poetry at Tennessee Tech University, Hoover is an AI trainer. She works for Mercor, a tech recruitment company that pays more than 30,000 “experts” upwards of $4m (£3m) a day to train AI models. It has hired almost five million such people since 2023. Mercor’s clients include OpenAI (which owns ChatGPT) and Anthropic (which developed Claude). Over the past year I have noticed Mercor advertising roles including “novelist” and “poet” on LinkedIn – not job titles you would typically expect to encounter on the app, let alone from a Silicon Valley start-up. The salary offered for both jobs was up to $150 an hour.
On an average working day, Hoover will be provided with a piece of AI-generated writing and have to explain to the model why what it has written is good or, more likely, where it has gone wrong. With Mercor’s experts in biology or maths, it’s easier to see how human trainers can quickly correct inaccuracies and improve a model’s output. But what determines quantifiably “good” poetry?
In 2024, researchers at the University of Pittsburgh found that non-experts more often than not preferred AI-generated poetry to the human-written variety. However, lay people are becoming adept at spotting the hallmarks of AI-generated writing: em dashes, or sentences structured “not X, but Y”. Writing experts can help AI models strip out these red flags, and teach them about prose and poetry as they might in a creative writing class. But the goal is not just that AI models learn the rules of rhyme and rhythm from these human trainers, it’s to assimilate their taste.
Taste has become a buzzword of the AI race in recent months. The president of OpenAI, Greg Brockman, tweeted in February that “taste is a new core skill”, and June brought the launch of Taste Labs, a start-up “building the data and infrastructure layer to give AI models and agents taste”. As AI learns to do jobs previously done by humans, it is now the more ineffable qualities of our interactions with art – why we are moved or engaged by some works of literature and left cold by others – that it is often attempting to capture.
A spokesperson for Mercor told me that writing experts are teaching AI models their “judgement”. The models learn the experts’ literary preferences and are rewarded when they meet them, like a poetic Pavlov’s dog. This is known as RLHF (Reinforcement Learning from Human Feedback) and was the default approach for fine-tuning ChatGPT. According to the Mercor spokesperson: “As models get better at the mechanics of writing, that judgement [from expert writers] becomes more valuable, and it’s why we’re seeing demand grow for skilled writers.”
In October, Brendan Foody, Adarsh Hiremath and Surya Midha, Mercor’s co-founders, became the world’s youngest self-made billionaires ever, at the age of 22. On a podcast in January, Foody, perpetually grinning and wearing a T-shirt with the Mercor logo, said that the aim of hiring poets is to train AI models to identify and spit out “good” poetry on command. “When one of the AI labs wants to teach their models how to be better at poetry, we’ll find some of the best poets in the world that can help to measure success via creating evals and examples of how the model should behave,” he said. Foody estimates AI models like ChatGPT will be able to write a poem “as good as the median Pablo Neruda poem” within the year. The idea that poetry can be quantified in “medians” and “evals” might be a sign that the stanzas of Silicon Valley will be a little different from those Neruda had in mind.
Duncan Brumby, a professor of human-computer interaction at University College London, compares AI models to the Wizard of Oz; seemingly magic, with a team of people working behind the curtain. “Behind these great tools like ChatGPT and the like, there’s often a lot of invisible human labour,” he says. In the early days of AI, this mostly involved poorly paid workers labelling images, captioning videos or moderating content. In their 2019 book Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass, Mary L Gray and Siddharth Suri write: “Beyond some basic decisions, today’s artificial intelligence can’t function without humans in the loop.” Seven years after Gray and Suri published their book, companies such as Mercor, ScaleAI, Handshake, micro1 and Outlier are all employing humans to teach AI how to do their jobs, which range from typical white-collar fare to niche areas of expertise. At the time of writing, Mercor is hiring a “Gujarati music expert” and a “culture specialist” among its 3,623 job openings.
It is 6am in New York when I speak with Clifford Evan, but he has already been awake for an hour, writing his novel while his children sleep. The book is a fantasy tale about a criminal who meets a girl with psychic powers. “He’s really a supremely talented artist, and she needs to get him back onto that proper path,” Evan explains. Evan has been a writer for 30 years, and the novel is part of a three-book deal he signed with a small publisher in 2025. After we speak, he will log onto his second job as an “expert writer” and “self-enrichment teacher” for Mercor. He has also worked for Outlier, another company that employs freelance experts to train AI.
“I would say it was kind of teaching [AI models] creative writing,” Evan says. This writing can range from instructions for a Roblox game to a speculative essay on ancient Babylonian deities. The “self-enrichment” aspect is a little more vague – his closest colleagues are an artist and a poet, but he says “self-enrichment” could equally be taught by a yoga teacher. Though this has connotations of spiritual enlightenment and growth, enriching an artificial self has its limits. “I feel like whatever is being written, there’s sort of a lack of soul to it,” says Evan.
AI trainers are assigned to projects with various clients, to train models for a range of purposes. The duration and demand of these projects can vary greatly. One project Evan was assigned to lasted eight months, but he says this was an anomaly; others have taken a matter of days. The world of AI training is “weird”, he says, “because it is kind of precarious, and you never really know what’s going on.” When I ask him about the overarching purpose of his job he pauses to think. “I don’t know if they ever explained that,” he says.
On Reddit, communities of trainers dispute whether AI training jobs are cash cows or scams. Some have earned more money in this line of work than in any of their previous jobs, while others have grappled with low pay, long hours and a frustrating lack of transparency. Although Evan is largely positive about his experience, he has had issues with payment. At some companies, trainers are given specific time constraints – often unfeasibly short – within which to complete tasks. If the timer goes off before they have finished, they risk not being compensated for their work. When Evan tried to complain about this, he couldn’t get through to a human manager, only an AI chatbot.
An expert from Spain, who has worked for various AI-training companies since 2024 and wished not to be named, says a boom in hiring has made working conditions worse. “Last year, projects became much shorter-term and companies started onboarding huge numbers of people, like hundreds of people at once. So over time projects could end literally within a week and good, long-term projects became less common,” she says. “Communication inside the communities also became chaotic.” This didn’t prevent her from introducing her partner and several friends to jobs as AI trainers.
In April, following a data breach, Mercor employees launched seven class action lawsuits against the company, accusing it of sharing workers’ data with clients without their consent. They allege that Mercor used recordings of candidate interviews to train AI models, shared applicant background checks with third parties and took screenshots of employees’ laptop screens (which showed bank account details and health insurance portals). A spokesperson for Mercor said: “A lawsuit can contain allegations and speculation.” In an earlier statement, the company wrote: “Of our nearly five million experts, only a very limited subset had sensitive information affected. There is no evidence that any of this data has been used fraudulently.” One California law firm suing Mercor has also launched lawsuits against Surge AI and Scale AI, claiming that the companies misclassify their workers as independent contractors in order to deprive them of benefits and minimum wage requirements. Surge has said the suit is “without merit”.
All the experts I spoke with said they feared they were being scammed when they first signed up to their job (where else can a poet earn $150 an hour?). Though most writers sign up to train AI models unenthusiastically and due to financial necessity, some are pleasantly surprised. Peter Valdes-Dapena had been an automotive journalist at CNN for nearly 25 years when he lost his job in 2024. “I was laid off about six weeks from my 60th birthday, so I had no illusions that it wasn’t going to be hard,” he says. Freelance work wasn’t paying his bills, and he saw an advert for a writer at an AI training company on LinkedIn. He immediately signed onto a project. “The first one, to be frank, it didn’t go well,” he says. Projects were unreliable and constantly changing. “It starts and then the client, whoever that is, wants to pause the project, and then they change the rules of the project and how you’re doing it.”
Valdes-Dapena’s second project, however, has lasted for nine months so far. He’s worked consistent 30-hour weeks and even made friends. “It’s got a really nice community around it of other writers who talk about our work and share our work with each other,” he says. He recently met up for a drink with colleagues who also live in New York. And the money from this project is good. So good that he doesn’t really write anymore outside of his AI work.
“I do worry about future generations and honing these skills,” he says. As a young journalist, he improved by writing badly and being edited well. Now, he is passing that training on, but to AI models rather than younger writers. Last year more than 3,400 journalists in the UK and US lost their jobs.
Valdes-Dapena’s work isn’t just trawling through AI slop; he finds it intellectually stimulating, albeit taxing. “I’ve always just been somebody who just writes well,” he says. “Never had to think about it, I just kind of had a feel for it. But this is causing me to need to take a deeper look at, actually, what is good writing? And what makes it good?”
Prof Brumby is not convinced that nuanced literary taste can be successfully learnt by AI models. “What makes something cool on the school playground? It’s not following the rules – it’s defined by the kids on the playground,” he says. “When kids start going to school, they spend forever trying to figure out what’s cool and how to be cool. They might buy all the right stuff, do their hair the right way, but then suddenly the top kid changes the rules, and that’s that. And I think that’s ultimately what we’re going to see happen here, right? These tech companies are ultimately run by a bunch of geeks who just don’t understand a lot of what’s going on.”
Yet every trainer I spoke with said they had experienced the models improving. “Seeing how quickly it becomes smarter was both freaky and fascinating,” says Tallen Gabriel, a 32-year-old poet from Brooklyn who worked for Mercor as a creative writing expert for just over a year. “It was almost like this personality was building in a way that you’re like… woah… you’re becoming more yourself.” This “self” is, in some respects, a funfair-mirror reflection of its trainers, whose own selves it is learning from. For Gabriel, this came with concerns. “I’m very anti-AI in art,” he says. “I’m just pretty anti-AI in general, which is ironic.” Working as an AI trainer was fickle and financially undependable, but so is being a poet. “I’m not working in a job that is super aligned with my values, but I also do need to make money. I live in Brooklyn. It’s very expensive here. I want to keep making art.”
Plus, there were moments when Gabriel felt he might be contributing positively to a force that was otherwise outside of his control. Occasionally, he claims, the models would encourage users to self-diagnose mental health issues. “There were some scenarios where the AI was giving kind of dangerous recommendations.” Gabriel would explain to the AI model why it shouldn’t do this, thus preventing it from occurring with a real user of the technology. “There were moments where I was like: ‘Okay, I actually think maybe I’m doing things that are helping the ethics of this a little bit.’ At least there’s that.” Other trainers would rather not think about the ethical quandaries involved in training AI. “I don’t necessarily think about it,” Evan says. “Because then you just kind of go crazy and you sort of lose yourself.” In not thinking about the darkness inherent in outsourcing art – the purpose of creating and consuming which is to enhance our lives – to artificial intelligence, what else are we losing?
Hoover, the poet from Tennessee, is a staunch optimist. She signed up to train AI wanting to understand how it could be used to write poetry, in order to pass this onto the university students she teaches. She says the AI models she trains “understand” rhyme, as well as the constraints of more formulaic styles of poetry. “There were moments where I thought: this is actually really interesting what it’s done.”
But she isn’t convinced that AI will master the elements of writing that don’t require a hard and fast rule. In her experience, AI models can’t fathom line breaks, or other elements of a poem that she believes come intuitively to human writers. When I ask ChatGPT to write me a poem, it generates five neatly rhyming stanzas within a second, titled “Where the Light Finds You”. They describe “heavy skies” and “tired eyes” and tell the reader: “Even in darkness, you’ll find your way.”
“I think what makes narrative poetry tick is not something that AI is able to really do yet. It produced some pretty terrible things for me when I tried,” Hoover says. AI tends to take writing instructions too literally. It knows good writing requires subtlety, so it becomes obsessed with silence and quiet. The poem ChatGPT generates for me contains “quiet tears” and “whispers”. And, while it directly addresses the reader, it doesn’t contain any human experience. There is something uncanny about being commanded to “breathe” by a technology that is incapable of inhalation.
Hoover would never use AI to actually write – nor would any of the other trainers I spoke with, aside from for research purposes – since most of her poetry comes from experience. “AI has never sat next to a lake; AI has never drunk a beer,” Hoover says, so how could it write about those things? The answer, of course, is that maybe, someday, she could teach it to.
[Further reading: A dream of Cambridge]
Content from our partners
Related
Join the debate
Subscribe here to comment