debtOr for a few hours a week, I write for a tech company worth billions of dollars. Joining me are published novelists, budding academics, and other freelance journalists. The workload is flexible, the pay is higher than we’re used to, and there’s no shortage of work. But what we write is never read by anyone outside our companies.

That’s because we’re not writing for humans, we’re writing for AI.

Large-scale language models (LLMs) like ChatGPT have made it possible to automate huge parts of our linguistic lives, from summarizing any amount of text to drafting emails, essays, and even entire novels. These tools have become so good at writing that they have become synonymous with the very idea of artificial intelligence.

But before we risk god-like superintelligence or catastrophic mass unemployment, we first need training. Rather than automating our lives with these fancy chatbots, tech companies are contracting us to help train their models.

The core of the job is writing fictitious responses to questions for a hypothetical chatbot. This is the training data that needs to be fed into the model. Before the “AI” can even try to generate “good” sentences, it needs examples of what “good” sentences look like.

In addition to providing our models with this “gold standard” material, we also help them avoid “hallucinations” (a poetic way of saying lies) by using search engines to give them examples of citing sources – without seeing such texts, the models cannot teach themselves.

Without better language data, these language models simply cannot be improved: their world is our language.

But wait a minute: haven’t these machines learned billions of words and sentences? Why do they need physical scribes like us?

First, the internet is finite. And so is the sum of all the words on every page of every book ever written. So what happens when the last pamphlet, papyrus, and prolegomenon is digitized and the model still isn’t perfect? What happens when there are no more words?

The date for the end of language has already been determined. Researchers Announced in June “If current trends in LLM development continue,” this is expected to happen between 2026 and 2032, at which point “models will be trained on datasets roughly the same size as the available stock of publicly available human text data.”

Focus on the words humanLarge-scale language models do little more than generate prose, and many of them are already publicly available on the Internet. So why not train these models on their output (so-called synthetic data)? The cyborg Internet, jointly created by us and our language machines, could expand infinitely. But no such luck. Training current large-scale language models on their output won’t work. “Learning indiscriminately from data generated by other models leads to ‘model collapse’, a degeneration process in which a model forgets the true underlying data distribution over time,” Ilya Shumailov and colleagues write in the paper. NatureIn other words, they tend to go off the rails and produce nonsense. Giving something its own stench leads to debilitation. Who would have thought?

Shumailov explained that whenever a model is trained on synthetic data, it loses awareness of the long tail of “minority data” (rare words, unusual facts, etc.) that it was originally trained on. The breadth of knowledge is lost and replaced with only the most likely data points. LLM is essentially a sophisticated text prediction machine, so if the original digital data was already biased (mostly English-language, mostly US-centric, full of unreliable forum posts), this bias is only repeated.

When AI-generated synthetic data isn’t enough to improve models, something else is needed. This is especially true for Concerns grow The much-praised model will likely be unable to be improved upon before it becomes useful in practice. Sequoia is AI companies need to close a $500 billion revenue gap by the end of this year to keep investors happy. AI machines may be hungry, but so is the capital to back them.

OpenAI, the trillion-dollar Microsoft protectorate behind ChatGPT, recently signed a licensing agreement with the company. Hundreds of millions of dollars From News Corp Financial Times.

But it’s not just a matter of accumulating original words: these companies need the kind of writing that their models try to emulate, not simply absorb.

This is where human annotators come in handy.

IFritz Lang’s 1927 classic film Big citiesThe ancient Canaanite god Moloch is reincarnated as an insatiable industrial machine: technology that works us, not for us. Factory workers meet its ever-increasing demands by charging at dials and pulling levers. But they cannot keep up. The machines hiss and explode. And we see workers abandon the act of feeding and walk straight into the mouth of Moloch’s furnace.

When I first took on the role of AI annotator, or more precisely, “Senior Data Quality Specialist,” I was very conscious of the irony of my situation. Large language models were supposed to automate the work of writers. The more the models improved through our work, the faster our careers would decline. And I was, feeding our own Moloch.

In fact, if there’s anything that this model accomplishes quite well, it’s the kind of digital copywriting that many freelance writers do to earn a living. Writing an SEO blog about the “Internet of Things” might not require a lot of research, pride, or skill, but it usually pays a lot more than writing poetry.

Working as a writer at an AI company is like being told Dracula is coming to visit and instead of running away you stay home and set the table. But our destroyers are generous, and the pay is big enough to justify the alienation. If our division goes up in smoke, we’ll just go up in smoke.

Therein lies the ultimate irony: we have a new economic phenomenon that rewards, encourages, and truly values writing. And yet, at the same time, it is seen as an obstacle, a problem to be solved, an inefficiency to be automated. It’s as if we’re being paid to write in sand, to whisper secrets into a block of butter. Even if our words could cause harm, we wouldn’t realize it.

But maybe it’s folly to treasure such mundane technology: After all, how many people are actually worth impacting?

Francois CholetThe author of a best-selling computer science textbook and creator of the Keras training library (which provides the building blocks for researchers to create their own deep learning models), said he estimates that “it’s probably about 20,000 people employed full time” just to create the annotated data to train large-scale language models. Without human input, he says, the model output would be “really terrible.”

The goal of the annotation work I and other researchers are doing is to provide gold-standard examples for models to learn from and imitate. This goes a step beyond the annotation work we’ve done unconsciously so far. If you’ve ever faced a “Captcha” problem that asks you to prove you’re not a robot (e.g., “select all tiles with a picture of a traffic light on them”), you’ve actually just been doing Unpaid labor for machinesHelp teach them to “see.”

As a student, I remember repeating words like “left” and “right” at a laptop for hours on end to help develop self-driving cars. I was paid a few hours’ worth of money for each satisfying utterance, not even close to minimum wage, so I gave up.

The role today is different and a key part of the LLM’s development. Alex Manthey, head of data at Context AI, is one of the people hiring writers to improve the models. She says: observer This practice is “mission-critical” and “requires human intervention to review,” [the model’s output] The human touch that “makes sense to the end user” works: “There’s a reason why every company spends so much time and incredible amounts of money trying to make this happen,” she says.

According to Sholet and Manthey, employment in this field has recently shifted from controversial, low-paid jobs in developing countries to more specialized, higher-paid roles. As models improve their ability to produce text, the quality of training data required also improves, and wages rise accordingly; some remote annotation jobs pay writers more than £30 per hour. Third-party annotation vendors such as Scale AI (valued at $14 billion) are also capitalizing on this shortage of high-quality training data.

A selection of current job adverts for AI annotation roles in the UK involve a variety of tasks, including: “Create responses that will serve as the ‘voice’ of the future AI,” “Provide feedback to help AI models become more useful, accurate, and safe,” “Write clear, concise, factually and grammatically correct responses,” and “Coach the AI model by assessing the quality of AI-generated writing, reviewing the work of peer writing raters, and creating unique responses to prompts.” If chatbots can write like humans, so can we.