My name on the platform is ri611. Or h924092b12ee797f, depending on who’s paying me.
I work as an AI trainer. I assess whether a chatbot’s tone is natural or flat, affected or annoying. I identify patterns in pictures of furniture; search the internet for group photos of strangers whom I’ll eliminate from the portrait, one by one. I trawl through bizarre videos so I can annotate and time-stamp the barking of a dog, the moment a stranger walks past a window, the precise millisecond a balloon pops. I generate anime sex scenes and decapitate young women, coax LLMs into giving me recipes for bombs made of household items, and generate invites to a reprise of January 6 at the White House, all as part of a red team whose purpose is to test safety precautions and probe weaknesses. I work for companies with names like Mercor and Outlier and Task-ify and Turing and Handshake and Micro1.
In my “other” career, I am a Hollywood writer and showrunner. I create prime-time TV, usually featuring a middle-class white lady having the worst day of her life, with some salt-of-the-earth police interference to raise the stakes. You can find my shows on Paramount and Hulu and the BBC. I would suggest you don’t.
In 2023, Hollywood went on strike, partly to keep the studios from replacing writers and actors with AI. When the strike ended after nearly five months, the entertainment-industry carousel never gained back its momentum. In early 2025—when yet another producer defaulted on a six-figure check I was owed for creating a TV show—I began to look around for some way to keep the wolves at bay.
AI training wasn’t on my radar until a comment in an unofficial Writers Guild of America Facebook group caught my attention. The page was filled with posts from unemployed writers struggling with debt and panicking about their income, begging for tips and ideas and survival strategies: “I am stressed and anxiety-ridden … simply trying to breathe” … “ISO food bank/pantry info” … “Hey, so what kind of part-time jobs are you all getting?” I’ve been working for this AI training company called Mercor, one woman typed in the comments. They’re paying 150 an hour for writers. It’s easy money.
I was down for some easy money. I too needed cash to pay rent, to buy food, to pay Maggie—the human still charging me a flat rate of 150 bucks to clean my apartment, a feat that AI had not yet figured out. How hard could it be to teach a machine to take my job? I was naive enough to believe that this industry wanted what we had to offer—not just our skills, but us.
I was wrong. Whatever this industry is, it is not easy money.
I got my first contract as an AI trainer in September 2025 after filling out 10 job applications, laboring for 20 (unpaid) hours on numerous tests to prove my capabilities, and being interviewed by an AI recruiter agent embodied by a flickering light on my screen. I was asked what I thought of a mediocre AI-generated couple of paragraphs about a soldier in the trenches sniffing a lavender-scented letter. Using all of the skills I had acquired with my English literature degree from Cambridge, I said it was shit. Six weeks later, I was hired as a “generalist” data annotator (below “expert” but well above entry level) at $52 an hour.
Once I’d passed the background check, I was made to install various apps and Slack channels and Airtables and payment portals and Google whatnots. After pinballing between them and a Zoom room where five unseen people hung out all day to counsel the legions of the confused, I was off and running.
My first task was to read a conversation between a user and “the assistant,” one of the major large-language chatbot models. Using a “bible” that dictated how the assistant should respond, I was to assess the chat as a success or a failure. The prompts were quirky and sad and heartbreaking. Are my feelings justified? Is this person’s behavior acceptable? Am I lovable? The AI responses belonged to an era when the assistant would happily tell you that you definitely had autism, your dad was clearly bipolar. I wondered if the user knew they had opted into sharing their private agonies as training data. After grading the assistant’s response on a scale of 1 to 5, I was to enter a justification for my verdict.
Our project manager, an intrepid 22-year-old recent university graduate who said he had intended to get into investment banking but failed, was in charge of about 10 unfriendly “team leaders” and “data managers.” Every day at a set time we would have Zoom office hours where we could discuss the complexities of our tasks. Our creative skills and our special minds were invaluable to this very important project! But it would be great if—in typing up justifications for our scores—we could keep our special minds on a tight leash and subordinate them to our ability to copy and paste verbatim from the scoring guidelines. Going off-piste with creativity, original thought, or fancy language might throw the model off.
I made friends with a handsome Swedish man who lived in the Nordic wilderness with his husband and numerous mammals. He had been on the project about a month longer than I had, and he kindly walked me through the platform and our employer’s expectations, which had been astonishingly vague despite the insistence that this work was urgent, important, and relevant, and must be guarded with the utmost secrecy. Handsome Swede and I exchanged contact information and shared dog pictures. The project was meant to be 20 hours a week for two months. I clocked 10 hours a week for two weeks, with constant stops and starts, before the project was summarily unplugged one morning with no notice. “Sorry guys,” typed University Graduate. “I had no idea this was coming.”
The Slacks and Airtables and office hours and Google documents were swiftly disbanded within a couple of hours. The project was over.
Most of the contracting companies that provide labor to AI firms advertise themselves to workers as offering the luxury of choice: “Contractors on Mercor’s platform choose when and how much to work,” sounding a common industry refrain. “How they participate on the platform is up to them.” Set hours and times are for boomers. Work on your own terms! Early on, I had this sales pitch bluntly reframed to me by a team leader in a midnight Slack message. I should not rely on this work, she snapped. I should not expect anything from it. These are not jobs, these are “tasks,” and we are “taskers.” I should think of tasking as a bonus. It is a “second job,” Team Leader typed.
She was so unpleasant she had to be human.
Four weeks after my first gig ended, I was offered an “expert” role, this time at $70 an hour. An “expert” is someone who usually has a higher degree, often a master’s, and significant work experience in their field, be it real estate, neurology, linguistics, history—or journalism. (“Expert” projects, I would learn, were typically given multisyllabic names from dead languages. Projects involving the minimum-wage grunt work of annotating tended to be named after small woodland creatures or celestial bodies. It is either a sign of my accomplishments, or my severe ADHD, that I was apparently a match for both.)
Work on Project Dead Language would start within a week, we were told. I went through another onboarding process. I joined another Slack. I signed up to another Airtable, which failed to indicate in any way whether the sign-up had been successful, prompting me to sign up a couple more times in confusion, before I noticed an all-caps message in the Slack exhorting me: DO NOT SIGN UP FOR THE AIRTABLE MORE THAN ONCE!!
A week passed, and “Phase 2” of the project failed to start.
Another week.
Another.
Thanksgiving arrived. Heartened by the prospect of extra cash, I drove six hours to Yosemite so that I could sit in an expensive cabin with my child and we could ignore each other in idyllic surroundings. Still Phase 2 did not arrive.
I had erroneously assumed that this new project would net me maybe $500 to $1,000 a week for a couple of months before Christmas. By December 1, I had earned just 180 of your finest American dollars.
Project dead language eventually launched not long before Christmas, four weeks after I’d joined. It was 9 pm on a Monday night. The doom pervading the Slack evaporated instantly and was replaced by panic over various technical problems. Turns out a bunch of people had, like me, registered for the Airtable multiple times over. None of us could access the tasks. By the time our tech issues were resolved 24 hours later, the work had run out. The tasks were finite. The smug few who’d evaded glitches had snatched them all up.
This abrupt hiring, firing, stopping, starting, abandonment, and rapid depletion of projects, was, I would learn, commonplace. A friend we shall call Jonathan, a mid-level TV writer who’d worked on several big streaming shows, was employed as an Expert Creative Writer. He was paid $150 an hour to evaluate scripts for OpenAI. He said it was all “a bit Hunger Games,” meaning he slept when they slept, and curried favor among his sponsors, also known as “TLs”—the team leaders who seemingly had the ability to hire and fire us at will. “It feels like we are all in a fishbowl waiting for our human masters to drop some food in a big aquarium,” someone wrote plaintively in a Slack for another project I joined that yielded barely any work. “And then, only the ones who are fast enough to swim to the top can eat.”
The more this became my new normal, the more I adjusted to the creaking lurch and giddy whiplash of the job. While we lounged in unpaid stasis waiting for an email to herald the arrival of work, we would be urged on by our team leaders and their exclamation points. Here they are at 3 am Eastern time with an update on why our Slack access has been revoked and why we need to change our password for the 17th time! There they are again at 11 pm with another energetic exhortation that the project will start any day now! At 7 am they’re back with the news that The Client is just finishing up Phase 1! At 2:27 pm: If you were a pizza, what kind of pizza would you be? Cue smiley face emoji. Fist emoji. Pizza emoji.
This would continue indefinitely. All unpaid.
At 7 pm on a random weekday, I’d walk in after a long day on set, having picked up the middle schooler from basketball. I’d take the dog for a walk, filter through the mail, think about throwing together some ingredients for dinner, when suddenly my phone would vibrate. My Slack would fill up with GO TEAM GO messages from someone who was just out of college, someone who has no idea that across the decades, people have died trying to establish labor laws that protect workers from the exact same conditions that he is now responsible for perpetuating, accompanied by numerous rocket ship emoji reactions. Our fearless leader tells us that it is IMPERATIVE that we complete our first task within 24 hours. If we do not, we will be at risk of being off-boarded! But you can work when you like! But if you don’t work now the tasks may be gone by the time you wake up tomorrow! LET’S GET THIS ACROSS THE FINISH LINE!
I abandon the meal. Retrieve some two-day-old pasta from an ancient Tupperware that has now been stained a dull, opaque orange. Throw it in the microwave. Slink over to the laptop. I had hoped for an early night. My middle schooler asks for help with his math homework. I snap not now, call Dad, too focused on the hunt to pay attention to the humans in my life. Log in. Start the timer. The Slack is exploding. It is a frenzy of caffeinated, taurinated, and probably high-as-fuck humans taking precious time to post jubilant updates about the precious work that has been bestowed upon us by The Client. An all-nighter is a prerequisite. The endorphins, possibly helped by amphetamines, are flowing freely. There will be no tasks in the morning. We must reap now, regardless of sleep, family, careers, and other trifles. There is nothing now but the project, and the task. Go Team Go.
We lock the fuck in. Desperate, warm, frail, finite human bodies with no overheads, fast internet, and a tolerance for less-than-ideal conditions.
We task through the night.
In the reddit groups dedicated to people who work for AI contractor companies, an atmosphere of fear and paranoia pervades. I sought out these forums soon after my first encounters with the industry, because I felt it was my responsibility to incite others with the rage, disappointment, and betrayal I had experienced waiting for work that often never appeared.
Turns out I did not need to incite anything. People were pissed.
Thousands of Mercor employees making $21 an hour on Project Musen had been fired in November 2025, and immediately rehired on an identical project, Nova, at a significantly lower rate—$16 an hour. Despite the insistent bleating that this was a “second job,” for plenty of people on Reddit it was their only job. Losing five bucks an hour hurt them badly. Not only that, they’d made friends, started a Discord together, knew each other’s names, found some kind of community. Plus the project had been disbanded just before Thanksgiving, for chrissakes.
People who had previously felt paralyzed by their NDA’s began to talk. Helena, the conflict-avoidant moderator of the Mercor subreddit, worked overtime deleting furious rants from aggrieved workers who delighted in dropping names of the “secret projects”—something explicitly banned by the non-disclosure agreement every tasker must sign before being hired as an Independent Contractor.
Elsewhere, on another project, Handsome Swede was not faring well. Felled by Covid, he told his team leaders he could not make the minimum weekly requirement and was swiftly fired. He entered the melee once again to find yet another project.
The wages were dropping week by week. When I first started scrolling the contractor jobs in early 2025, companies like Mercor, Handshake, Turing, Task-ify and Outlier were offering $150 an hour for “experts,” $35 to $75 an hour for “generalists.” Today, Mercor says the average hourly rate on its platform is $105. But in my searches across the industry near the start of 2026, the experts were often getting $50 an hour, and the entry-level grunt workers were getting as low as $16 —less than California minimum wage. Contracts were now referred to as “sprints.” The work had to be done, asap, as fast as possible, for employment that might last 24 hours. The urgency was paramount, self-important, and annoying as fuck.
The burnout has led many taskers to turn to the courts. Several lawsuits have alleged that Mercor is misclassifying its workers as independent contractors, pointing out that the demands of the job—frequent onboarding, infinite retraining, the need to check email and Slack several times a day, to be on call and perform at very short notice, the expectation that taskers will complete a certain number of hours every week—are indications of employment. But compared to regular employees, contractors receive almost no workplace protections against unpredictable scheduling, prohibitive work hours, denial of breaks, or retaliation from employers. Which feels like a big risk if, like me, you are tired of the bullshit and complain. Loudly. Often.
Christmas day came. I had not earned the additional $3-5K I thought Project Dead Language would have netted, and my bank account hovered around $14. Mired in existential panic and with only enough money to live off cereal, I accepted two different invites to work on an enormous $16-an-hour project that was in its final stages. It employed several thousand annotators across multiple platforms to perform incredibly boring objectives. The entire enterprise had the feeling of a bustling refugee camp that had been functioning long enough to cover essential needs, but not to be, like, comfortable. I’d already completed most of the onboarding steps. The most important thing, they emphasized in the literature, is to get on Slack.
I couldn’t locate the Slack.
I called the Zoom helpline.
“Do you just hang out here all day?” I asked a faceless man while, in another square, an elderly woman peered suspiciously into her camera wearing a nasal cannula attached to an oxygen tank, set against a background of palm trees. “Pretty much,” snorted the faceless man. “I hope they pay you well,” I said sincerely. “They don’t,” he responded, before informing me that I was already a member of the Slack channel I had spent two days waiting to join, and that I had missed five essential onboarding quizzes in a document I had failed to read.
I was dispensed with in three minutes. “What about me?” rasped the elderly lady through her tubes. “I’ve been here for 15 minutes now.”
I dove into the Slack channel of which I was already a member. Spirits were high. A badge had been promised to the top 20 percent of performers. This honor would be displayed on their profile, and potentially lead to quicker employment and less downtime on the next project. The badge was causing consternation: Who would get it? Who might be erroneously deprived of it? Because it might lead to something. The badge would validate—all of this.
I located the quizzes I’d failed to complete, worked my way through them, and then waited. Apparently, I would be granted access within 24 hours. After 24 hours came and went, I reread the complex onboarding document, which said I needed to check in at the Zoom again. The faceless Zoom team informed me that I had screwed up. There was another quiz I needed to complete. They could not grant me access. I burst into tears.
They removed me from the Zoom, and blocked me from reentering.
Twenty-four hours later, I was removed from the project.
Once you have overcome your initial shock at conversing with a mellifluous, perky, female AI recruiter with a two-syllable name, it becomes easy to do so again, and again, and again. This is, in part, why legions of people are lured to apply to work for an AI recruiting company like drunken sailors to a siren. “Zara” doesn’t care if you’re in your car wearing your pajamas, digging out a stubborn booger, swamped in a plush sherpa wearable blanket. You can even tell her to eff off, inform her that you’re practicing your emotional voice skills, and watch her process this; a humming, flickering light in the middle of the screen. “You are using language and prosody that indicates irritation,” she states blandly.
In the real world, if you look promising for a particular job, you might be fortunate enough to get called in for an interview. In the world of AI, an interview is standard fare for everyone. This has prompted some people to suspect—despite companies’ assurances to the contrary—that the interview process itself is a means of harvesting data, a form of free AI training. But we persist because there are still enough braggarts on Reddit posting about making “life-changing money” that we think we might actually land the job that pays $150 an hour for something not too boring.
By February 2026 I’d been on Project A, a video-annotation gig, for five weeks: a lifetime in the AI training world. I had beaten the system by eliminating my social life so that I could task every free second of my existence. I also scrolled through thousands of videos to locate the most absurd, complex “edge” tasks: videos featuring speakers with accents so thick and regional that no one apart from a fellow Celt could decipher them. The audio was often so distorted it hurt to listen: a cacophony of excruciating, screeching dissonance. By doing the tasks no one else wanted and chaining myself to my laptop, trying to time my submissions so I could make them last throughout the dry patches, I could keep myself in employment. In reality, I hadn’t beaten the system, the system had beaten me—and would continue to find new ways of doing so.
The senior project lead was another 22-year-old: a sweet baby-faced guy several months out of the Ivy League with a degree in economics. His LinkedIn picture showed him in graduation robes. His employment history showed one job: this. Beneath him, hundreds of people churned out captions for videos on a janky interface, while an invisible force of quality-control reviewers monitored and graded our performance.
Initially, these grades were something we annotators never saw. Then a few weeks in, some genius had the thought that revealing our scores would encourage our competitive spirit. The reviewers graded us harshly on a scale of 1 to 5. Five was perfect, 1 was utterly useless. Most of us seemed to be around the 2 mark. Those consistently scoring less were threatened with off-boarding.
Then our managers announced that a “golden batch” of tasks would be released to the most talented, the most special, the royalty of annotators—the folks, we were told, who consistently scored a perfect 5 with an average handling time well below the recommended amount. These moves resulted in chaos, fury, and an unexpected revolutionary streak in the Slack channel.
We all understood that we were expected to transcribe the contents of videos in microscopic detail and that we’d be graded on how accurately we time-stamped the words we heard and the other verbal indicators in the video. Yet as our scores plunged, the feedback we received came in the form of absurdly granular orders stamped with an error flag: “Replace ‘t-shirt’ with ‘a t-shirt.’” “Swap out ‘red’ with ‘maroon.’” “Change ‘grunt’ to ‘grunting.’” It became increasingly evident that the scoring rubric was a vague and moving set of goalposts.
Attempting to repair the damage, our managers—two humorless young women, one of whose LinkedIn profile pictures also featured graduation robes—went into overdrive. They messaged the most pissed-off among us and pleaded with them to “be positive,” suggesting that no one understood the pressures they were under (debatable, given that most of the people they were managing seemed to have at minimum a decade of experience in their various industries). They dismissed suggestions for improving the workplace. But they did inaugurate a new, optional “coffee time” Slack thread to encourage team spirit, featuring icebreaker questions such as, If you were a condiment, what condiment would you be?
I estimated that about 95 percent of the annotators were professionals in their thirties or forties with a seething, deep-rooted hatred of their Gen Z overlords—“a clueless bunch of kids with no work experience,” as one colleague phrased it. The other 5 percent were sycophants who could see that outsmarting their bosses would get them nowhere, and placating them with blithe obedience was the key to success. “My overall attitude is that everyone has my best interests at heart,” Linda, in her sixties, wrote primly on Slack. “If I get a 1, I study the task to make sure I understand the problem. If I get a 5, I also look at the task to make sure I understand what I did right.” I decided that Linda was not my people. Also, if I were a condiment, I would be Marmite.
No matter how hard we worked, our scores went down. Meanwhile, the management team was constantly recruiting people with the best scores to “promote” into reviewer jobs. My colleague Melanie had recently been demoted from reviewing, back to annotating. “No one gets more money for being promoted,” she wrote to me on Instagram. “That’s all a lie.” (I’ve changed some identifying details in this story to preserve the anonymity of people who fear blowback.)
These twists of logic were tearing apart our morale. I had become some kind of snarling beast, schlepping McFlurry and Hot Pockets over the keyboard (I did not have time to cook). I kept applying for more AI trainer gigs in the background, growling at my AI interviewers with increasing disrespect. Any moment might be my last, and so I threw myself off the cliff of propriety with alarming regularity.
In a regular workplace, face-to-face interaction forces a modicum of civil behavior onto disgruntled humans. The moments of kindness and empathy that emerge alongside our worst traits can be enough to make the coldness of the corporate world a little more tolerable. In these AI gig environments? Forget it. Occasionally, the bone-dry messages of asinine insipidness would be punctuated by someone truly losing their shit. WTF IS GOING ON? they would scream in all-caps in the FAQ channel. The most irate began to disappear. We hoped that they had been led, angels singing, to the hallowed realm of the Golden Task. But we feared the worst.
With the reckless abandon of a millennial who has been tone-policed once too often, I, too, began to post messages encouraging rebellion in the Slack channel. This did not go unnoticed by the operations manager, a mirthless young stay-at-home mother with a religious background. After noting that I took the Lord’s name in vain, Mirthless insisted that I “try to use a professional and positive type of communication.” I resorted to workshopping all of my responses to her through ChatGPT. It proved to be an excellent collaborator, well-versed in bland corporatese and the battlefield tactics of modern office life.
But ChatGPT would only participate in this insanity so far. “Go somewhere where your unique talents and skills will be welcomed and encouraged!” it told me, presumably tired of my complaining. “Redirect that irritation somewhere productive. Into something that exposes the absurdity of this system. Because you’re not small in that room. You’re just temporarily renting space in it.”
How could I break it to ChatGPT that I was small in this room? So small that I had been crushed into about 72 pixels per square inch? But it was this or not pay rent. I made more money in three days on a project that involved writing shopping prompts for automated lawnmowers and red-light-therapy masks than I did teaching three hours a day for a month at UCLA. (The shopping prompts gig—bizarrely my favorite of all of the projects—lasted a week before they fired me.)
I no longer knew what the Golden Task of my own life might look like. When given the chance to respond to details from this article, Mercor said that it strives to give workers “as much notice as possible when these projects change”—a sentiment roughly echoed by other companies. Between February and April 2026, I was hired and fired on seven different projects over four different platforms. The dismissals were always abrupt, shocking. One moment I would be typing rubrics into an Airtable, waiting in line on a 24-hour Zoom to talk through a task with a reviewer. The next, the UI would vanish. The Slack channel would disappear. The Google docs would lock me out. No message. No warning. No explanation.
I never intended to write about this industry. I came to it not as a journalist but as a disgruntled, broke TV writer determined to make a dent in student loans and keep paying LA rent while my industry withered in front of me. But working with and for AI had proven even more cruel than I could have ever imagined. Mercor says it employs about 300 full-time staffers. Meanwhile, each week it keeps some 30,000 independent contractors caught up in a fever dream of aimless, directionless urgency, corralled across Slack channels by achingly young adults, sending messages at 3 am to “push on” and “finish strong” and “lock in” and “Go Team GO!” All in service of the grandest purpose in history: to successfully remove a scuba diver from a picture with one click of a mouse, transport him to the moon without any glaring artifacts—and bring him back again.
The next generations of team leaders won’t know our specific talents or our unique skills, but they will know the Average Time it takes us to annotate a grainy video uploaded without the owner’s consent into a vast catalog of other possibly stolen videos. They will be tasked with making us work faster, and longer, with more precision, more control, fewer errors, fewer overheads, fewer costs. To make the machine more human, they will make us more like the machine.
What Say You?
Let us know what you think about this article in the comments below. Alternatively, you can submit a letter to the editor at [email protected].
