Gen Z AI-Era Careers: What Data Labelers & Engineers Really Make

Editorial Note: The three people in this story are composite characters, built from published reporting, industry wage data, platform information, and documented events. Individual details such as names, employers, living arrangements, and some financial figures are combined or adapted to represent patterns found across multiple sources. Where specific pay rates or industry trends are cited, they are grounded in the sources listed at the end of this article. Employers are pseudonymized, except where a company is named in mainstream coverage. Nothing here is career advice.

Gen Z Careers in the AI Era: What a Data Labeler, an Ex-Prompt Engineer, and a Data Center Electrician Actually Make

A Day in the Life of an AI Data Center Electrician Apprentice: $34/Hr, Overtime, and Dropping Out

Cole cinches his tool belt while it's still dark out. On the belt: strippers, a voltage tester, a roll of blue electrical tape, and a Klein screwdriver. He bought all of it himself — about $1,200 last year. He's twenty-two, fourteen months out of community college, and in year two of an IEC electrical apprenticeship.

The job site is on the northern edge of the Dallas metroplex, a data center campus still filling in. Three months ago this was dirt. Now the steel is topped out, and Cole and a dozen other electricians are landing feeder cables from the main switchgear onto a fresh row of cabinets. The foreman, Ray, walks up holding a printed one-line diagram and points at the gray racks nearest the wall. "This row's for the AI."

Cole doesn't ask whose model runs in there. He only knows this shipment is heavier than last month's, and the cooling lines are denser. He drags a length of 4/0 copper out of the tray, bends it, crimps the lug, torques it to spec. Forty-odd repetitions later, his forearms start to burn.

He remembers the day he dropped out, in the office of his community college advisor. Martinez, fifty-something, had Cole's transcript for three semesters on the desk — a 3.4 GPA, nothing to be ashamed of. Martinez took off his glasses, wiped them, and said, "You're sure? You're four semesters from the degree." Cole told him he'd run the numbers: an entry-level software job around $70K, minus student loan interest, came out close to what he was already making as a first-year apprentice — except the apprenticeship carried no debt, and the wage ticked up every six months. Martinez was quiet for a few seconds, put his glasses back on, and said, "Alright." No disappointment in it. More like a confirmation.

It's 9:30 a.m. now, and Cole has just finished a set of terminations. Standing under the cable tray, drinking water, he does the math in his head: twelve overtime hours this week at time-and-a-half. $34 × 1.5 = $51. Eight more hours and he clears another $400 before tax. He needs it — the journeyman exam is next month, and the test fee plus study materials will run him three hundred and change.

He glances at his phone. There's a Reddit push notification on the lock screen. He doesn't open it. Phone back in the pocket, on to the next cable.

How Much Do AI Data Labeling Jobs Really Pay? From $9/Hr at Remotasks to $28/Hr at Outlier

Erin's office is a sixty-centimeter-wide white IKEA LACK table in her Buffalo apartment — just big enough for a laptop and a water glass. She splits the two-bedroom with a roommate; her half of the rent is $780. Her bedroom faces north and the heat barely reaches it in winter, so she mostly works at the kitchen table instead. The kitchen has radiant floor heating, and the table is bigger.

She's twenty-three, an English grad from the class of 2024. For four months after graduation she sent out more than two hundred applications — marketing coordinator, content writer, editorial assistant. Under three percent of them got any response at all. In month five she found a data annotation listing on Indeed advertising low-paid remote data annotation work with no experience required. She clicked apply.

The platform was Remotasks, a crowdsourcing marketplace under Scale AI. The work: tagging images, classifying conversations, judging whether an AI-generated response was helpful, honest, and harmless. The rubric had a dozen criteria, each with sub-clauses. Her first two weeks she averaged twenty-odd items an hour; once she got the rhythm, thirty-five to forty. At $9 an hour, that math comes out to a few cents a judgment. She told herself she was building experience.

A year later she moved to DataAnnotation.tech, where she graded the quality of AI writing on project-based contractor work that paid more than her earlier annotation jobs. The tasks were heavier: reading long-form model output and scoring it across six dimensions — coherence, factual accuracy, tonal appropriateness among them — then writing revision notes. Thirty to thirty-five hours a week, $2,600 to $3,000 a month, steady. Then she learned the catch: the platform guarantees no minimum hours. In slow weeks she can grab four or five hours of work a day, and her income halves overnight.

Outlier is her third stop. She works on AI evaluation projects, with task rates in the high-$20s per hour when the projects are available. She's not a doctor or a lawyer, but in the platform's tiering system, writing ability plus rubric accuracy is its own track, and she sits in the middle of it. Moving up requires submitting more high-quality samples and passing review. She's tried twice. The first attempt was rejected — "insufficient grasp of clinical nuance." The second is still pending.

It's Tuesday. She opens the Outlier queue and finds a system notice: this week's evaluation target is v3.2 of a major open-source model, and the rubric has four new criteria. She reads them one by one. One of the new rules says the model should avoid over-apologizing when it refuses a request. Interesting, she thinks. Then she starts labeling.

At lunch she scrolls her phone and lands on a Reddit post — something about an engineer with eighteen years on the job, laid off, now flipping burgers at McDonald's. She reads two screens of comments, upvotes it, and goes to reheat her lunch. At 1:30 she's back at the table. One hundred and thirteen items left in the queue.

Sometimes Remotasks-era memories surface. In spring 2025, MIT Technology Review published an investigation: photos taken by iRobot vacuums inside users' homes — including one of a woman sitting on her toilet — had been sent to Scale AI for annotation and leaked onto internet forums. Erin had already left by then, but she knew the pipeline: photo in, task assigned, labeled, QA'd, delivered. She had been one of the hands on that chain. The day the story broke she stared at the screen for a long time. Not anger, not relief — just a very specific memory of having handled something.

What Happened to Prompt Engineers? One Developer's Pivot to Agent Workflow Engineering

Miles rents a one-bedroom in Denver for $1,650 a month, with a window that catches the edge of the Rockies. He's twenty-four. He graduated with a CS degree in 2023 and took a backend job at a mid-size SaaS company for $92K. Eight months later, a forty-person AI startup poached him. The title: Prompt Engineer. The offer: $122K.

It was fall 2023. His LinkedIn DMs filled up with people asking how to break in. Bootcamps were advertising "prompt engineering in seven days." The press kept recycling Anthropic's $335K outlier as if it were the going rate. Miles's actual days went like this: writing and tuning system prompts for the company's homegrown support chatbot, running evals, logging how each phrasing shifted output quality. He believed he was doing real technical work.

In February 2025, the company eliminated the entire prompt team. In the all-hands, the CEO said something about model capabilities improving faster than manual tuning could keep up with, and then read out the layoff list. Miles's name was on it. He spent four months unemployed — a hundred-plus applications, a dozen interviews. Most of the openings wanted five years of ML experience, or had renamed themselves "AI Solutions Architect" or "Agent Workflow Engineer." When the offer finally came, it was $106K. Thirteen percent below his last salary.

What he does now looks nothing like 2023. He barely writes prompts anymore — a prompt is one small component in a larger system. His job is building the harness around agents: task decomposition, state management, failure recovery, tool-calling chains. Last month he inherited a customer-support agent project. Same underlying model, same prompt templates. He rewrote the retry strategy and the context-window management, and the next round of internal evaluations showed a dramatic improvement in task success. The industry now calls this harness engineering, and the shorthand is Agent = Model + Harness. For the first time, he feels like he's doing actual engineering.

His LinkedIn still carries the tag "former Prompt Engineer." Every week someone messages him asking whether prompt engineering is a good field to get into. He usually doesn't reply. When he does, he sends a link or two about how the standalone role has changed. Then he tells them to learn the underlying engineering skills, not just the title.

At 2 a.m. he pushes a new agent config to staging, and the CI pipeline starts chewing through the test suite. He goes to the kitchen for water and passes the living room, where his phone sits on the coffee table. Same Reddit push notification on the lock screen — the eighteen-year engineer, now at McDonald's. He sits down and reads the entire thread, all of it, top comment to last reply. Some people blaming capital. Some saying switch careers. One person posting the exact timeline of their own layoff. Forty minutes later he sets the phone face-down and goes back to watch the CI logs.

The tests are still running. He doesn't wait for them. Lights off, bed.

How AI Model Updates Ripple Through Data Labeling, Agent Development, and Data Center Construction

A major open-source model releases a new version in March.

Erin spends the whole week grading it. Monday morning, her Outlier queue is wall-to-wall v3.2 output. Two of the four new rubric criteria are about refusal behavior, one about context retention across turns, one about comment conventions in generated code. She burns two hours reading the rubric update doc before she starts. Her throughput this week drops fifteen percent — the new criteria demand more judgment per item, and she can't afford an accuracy dip that might cost her tier. By Wednesday night, at item two hundred, she notices she's just flagged "over-apologizing" three times in a row. She stops and thinks about it. The new rule's boundary really is fuzzier than the old one. She files a note through the platform's feedback channel. Thursday morning, an automated reply: received, thank you for your feedback.

Miles finds the anomaly on Tuesday morning. His customer-support agent's success rate on staging has jumped overnight. His first instinct: he broke something. He changed the context-window truncation logic last week; maybe a bad truncation is accidentally feeding the model extra information. He spends three hours diffing his last five commits and finds nothing. Then he opens the model's changelog. Release note number three for v3.2: "improved instruction following in multi-turn customer service scenarios." He stares at the line for thirty seconds, understanding that maybe half of his last three weeks of harness tuning was compensating for a defect in the old model — a defect the vendor just patched under him. Relief and deflation in the same breath. Then he opens a fresh eval report and starts recalibrating the baseline.

On Cole's site, the same week is a crunch. A new shipment of cabinets lands four days early, and Ray compresses a three-week pull into ten days. Monday through Wednesday, Cole works until nine at night, all of it at time-and-a-half. Thursday at 3 p.m., a liquid-cooling delivery truck blocks the campus entrance; it takes two hours to unload, and that night goes to ten. Friday morning, buckling his tool belt, Cole finds a blister on the web of his right hand. He pops it with a needle, tapes it, and goes to land the next fiber termination.

These three people do not contact each other this week. They don't know each other exists. But the same AI buildout reached all three of their jobs in different ways: one worked extra overtime, one had to redo evaluation work, and one spent more hours at a kitchen table.

AI Career Incomes Compared: W-2 Employee vs. 1099 Contractor vs. Blue-Collar Apprentice

Miles. $106K a year, W-2 employee, 401(k) match and health insurance. Rent in Denver: $1,650 a month, plus about $180 for utilities and internet. Student loan balance: $34K, $380 a month. The equity from his last company went to zero when he was laid off; his current options vest over four years and sit at roughly $12K on paper, but he no longer counts that as an asset. Take-home after tax: around $5,800 a month, against roughly $3,200 in fixed costs. The rest goes into a high-yield savings account. No budgeting app — a Google Sheet, updated on the first of the month.

Erin. $26–28 an hour, 1099 contractor. No health insurance, no paid time off, no retirement match. Rent in Buffalo: $780 for her share of a two-bedroom, plus about $90 in split utilities. Student loan balance: $21K, on an income-driven repayment plan, current monthly payment $0 — below the threshold — while interest keeps accruing. She sets aside money for self-employment taxes and other contractor expenses, because none of them are withheld automatically. She pays $45 a month for a private short-term disability policy, because no employer provides one. Her income swings hard: a full month of work clears $3,600–3,900 before taxes; a slow month can fall to $1,800. She tracks every dollar in Notion, each entry tagged by platform and date. A week of being sick costs her $1,000 to $1,100, so she tries not to get sick.

Cole. Year-two apprentice, $34 an hour, W-2 employee. His contractor carries workers' comp and a basic medical plan with a high deductible. Rent in the Dallas suburbs: $850 a month plus $110 in split utilities. No student loans. Tools run about $1,200 a year, an out-of-pocket cost he keeps track of. Overtime at time-and-a-half, averaging forty to sixty hours a month. In a typical month, his gross pay runs well above his straight-time base because of the overtime; take-home varies with hours worked and deductions. No credit card debt. A 2014 Ford F-150, paid off — gas and repairs average $350 a month. He keeps his books in a paper notebook, one line every night.

Set the three ledgers side by side and none of them is better or worse. They're just three different ways of distributing risk. Miles's risk lives in the memory of equity going to zero and the odds of the next layoff list. Erin's is printed on a 1099 form and a quarterly self-employment tax deadline. Cole's rides on a blister at the base of his thumb and, eventually, on a knee.

Doubts in the AI Workforce: Labeling Ethics, Title Inflation, and Physical Toll

On a Thursday afternoon, Erin lands on a therapy-scenario conversation. The user describes anxiety symptoms. The model responds with something structurally complete, warm-toned, laced with empathy statements and correct clinical terminology. She scores it down the rubric: accuracy, pass. Safety, pass. Tonal appropriateness, pass. Then she stops at "overall helpfulness." The response is good — too good. She cannot tell whether it reads like a trained therapist or like a model trained to sound like one. She thinks about her own six free sessions at the campus counseling center sophomore year, and how the counselor's phrasing carried a similar temperature to this text. Her cursor hovers over the score buttons for maybe twenty seconds. She picks 4 out of 5 and types a note: "Response is clinically appropriate but may create unrealistic expectations of AI-delivered support." She submits. The next item loads: something about Python debugging. She keeps labeling.

On Wednesday, Miles gets an email from a recruiter. The role's title is AI Enablement Engineer. The JD says it will "bridge the gap between business stakeholders and AI capabilities." He stares at the title for five minutes. Enablement. He thinks about Prompt Engineer in 2023. AI Trainer in 2024. AI Solutions Architect in 2025. Each title seems to have a shorter half-life than the one before it. He doesn't reply to the email. He doesn't delete it either. He files it in a folder called "maybe later." There are twenty-seven emails like it in there already.

Cole's senior electrician is named Dale — fifty-two years old, thirty years in the trade. Last year he had his right knee replaced. Six months of recovery, and now he can't climb anymore; he does site QA and apprentice mentoring instead. Dale walks with a slight hitch in the right leg, but his eye for a bad termination is sharper than anyone under forty on the crew. One lunch break, Dale sits on a toolbox eating a sandwich and tells Cole he figures eight more years, then retirement. "Eight years from now you'll be forty-five," Dale says. "You figured out what you're doing then?" Cole says he hasn't thought about it. Dale nods, takes a bite of the sandwich, doesn't push. That afternoon Cole climbs a twelve-foot ladder to land smoke-detector wiring on the ceiling deck. At the top he stops for a second and looks down at the slab below. Then he reaches for the junction box.

The Daily Reality of AI-Era Workers: Three Unfinished Moments

At 11:40 p.m., Erin finishes another long day of labeling. She closes the laptop, rolls her neck, walks to the kitchen to put on coffee. The kettle is loud in the quiet apartment. She pours a cup and drinks two sips standing at the window. Outside, March snow in Buffalo, the streetlight turning the flakes orange. There are still 113 items in the queue; it refreshes at eight in the morning. She sets the cup down, goes back to her chair, and opens tomorrow's to-do list.

At 12:15 a.m., Miles's CI pipeline finishes the last integration suite. Pass rate: 94.7% — two points under what he expected. He opens the failure list. Three timeouts, two assertion mismatches. He doesn't fix them now. He copies the failure logs into a markdown file, saves it, closes the IDE. He walks to the window. On a clear night in Denver you can see a few stars. He stands there about a minute, then kills the lights and gets into bed. The phone goes on the nightstand, unplugged. In the morning he'll look at the five failures first, then decide whether to touch the retry interval.

At 5:30 a.m., Cole's belt buckle clinks as he fastens it. Boots on. He checks his pockets for tools, goes out to the truck. Today is rooftop fiber terminations; the forecast says thunderstorms after two in the afternoon, so they have to be off the roof before then. He starts the F-150, and the headlights cut across an empty road. The radio is on low — country. Twenty minutes later he reaches the site, signs in, pulls on a hard hat, and walks toward the gray data center building. A safety notice is taped at the entrance: today's high-risk operations, work at height, energized circuits. He signs it. Inside the locker room, he takes his harness off the hook.

References

U.S. Bureau of Labor Statistics — Electricians https://www.bls.gov/ooh/construction-and-extraction/electricians.htm U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Electricians. Updated August 2026. Supports the article's discussion of electrician wages, apprenticeships, overtime, and the growing demand for electrical workers. The BLS reports a 2025 median annual wage of $63,190 and projects 9% employment growth from 2025 to 2035.
U.S. Bureau of Labor Statistics — Software Developers, Quality Assurance Analysts, and Testers https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm U.S. Bureau of Labor Statistics. Occupational Outlook Handbook: Software Developers, Quality Assurance Analysts, and Testers. Updated August 2026. Provides current wage and employment context for software engineering. The BLS reports a 2025 median annual wage of $135,980 for software developers and projects 10% employment growth from 2025 to 2035, with continued demand tied in part to AI software development.
LinkedIn Economic Graph — Labor Market Report 2026 https://economicgraph.linkedin.com/research/labor-market-report-2026 LinkedIn Economic Graph. Building a Future of Work That Works. 2026. Provides broader labor-market context for the article's discussion of AI-era job titles and skills. LinkedIn reports that U.S. jobs requiring AI literacy skills, including prompt engineering, grew 70% year over year, while 1.3 million new AI-enabled jobs emerged globally over the previous two years.
DataAnnotation — Generalist Roles https://www.dataannotation.tech/jobby-roles/generalist DataAnnotation. Generalist. 2026. Provides current platform-level information on AI evaluation and annotation work, including reviewing and rating AI responses, writing improved responses, independent-contractor status, flexible hours, and advertised compensation of $25–$50 per hour for generalist work.
Outlier AI — FAQ https://outlier.ai/faq Outlier AI. Frequently Asked Questions. 2026. Provides current information on AI training and evaluation work on the Outlier platform. Outlier states that rates vary by expertise, project complexity, and location, that contributors see the tasking rate before starting a project, and that payments are processed weekly.

Menu