⌕

THE HOURLY RATE IS THE SERVING SUGGESTION

NOV 14, 2026 · SPLASH · MONEY · THE PIECEWORK
AI training data work pay per task Outlier Scale AI Appen Google contract 2024 FTC task scams 20000 reports piecework

THE PIECEWORK. “Up to $150 an hour” is a per-task rate wearing an hourly costume. And the contract behind it can end on someone else's calendar.

"UP TO $150 AN HOUR TRAINING AI." That's the pitch, and it's technically true in the way a lottery ticket is technically worth the jackpot.

Here's the mechanic nobody puts in the thumbnail. Most of this work is paid per task, not per hour — so the hourly rate is an estimate, and the estimate assumes a speed you might not work at.

One platform publishes its own example: two people on the same project quoted at $30 an hour. One finishes ten tasks in an hour. The other finishes three. Same rate on paper. Radically different paycheck.

The hourly number is the serving suggestion on the cereal box.

AND THE SPREAD IS ENORMOUS. This is one industry with two different lives:

In the United States, contractors on platforms like Outlier — Scale AI's contributor arm — have described earning $50 an hour at peak, working 50-hour weeks on projects lasting months.

In Colombia, a worker runs five different data-labelling platforms from her home in Medellín, earning between five and twenty-five cents per task.

Same supply chain. Same clients. Same task type. The difference is where you're sitting when you do it.

THE FIRST RULE: NEVER PAY TO GET PAID. Because this corner of the internet is full of a specific scam that looks exactly like the real thing — an app, a queue of tasks, a balance climbing on screen, and a request to deposit your own money — usually crypto — to unlock the withdrawal.

The Federal Trade Commission's numbers are brutal: reports of these gamified job scams quadrupled each year since 2022, reaching roughly 20,000 reports in the first half of 2024, with about $220 million in reported losses in six months.

The FTC describes the design honestly: the balance going up is meant to feel like a game, and the deposit is the win condition for the criminal, not the worker.

NOW THE STRUCTURAL PART, WHICH IS THE REAL STORY. The platforms don't create the work. Their clients do.

In January 2024, Alphabet told the Australian data company Appen that all contracts between them would be terminated, effective March 19. That was an $82.8 million contract — and roughly 2,000 contractors who had been doing search-quality and AI training work lost it with weeks of notice.

Not laid off. Not fired. Their projects simply stopped existing.

That's the sentence to remember about this entire industry: an account is not a job. It's a seat in the waiting room.

WHY IT'S BUILT THIS WAY — AND IT'S NOT AN ACCIDENT. Pay per task and the platform carries none of the risk. There's no hourly commitment, no benefits, no unemployment insurance, no notice period, and no obligation to offer you the next batch.

And per-task pricing does something else that's easy to miss: it means the fastest worker sets the wage. If the task is worth a fixed amount, then the only way to earn more per hour is to go faster — which sets the benchmark for everyone else doing the same task.

Speed becomes the price. Not skill, not accuracy — speed. And the quality the client says it wants is accuracy.

WHAT'S ACTUALLY BEING SOLD IS JUDGMENT. The valuable task in this industry isn't labelling a photo. It's the one described as reading a beautifully written answer and saying, with total confidence, this is wrong.

That's the work a model cannot do for itself — and it's why this market exists at all, and why it pays more than the low-skill tier. The industry's most important job is catching its own mistakes.

Which is also the strangest part of the deal: the people doing it are being paid to make the system better at the thing that might replace them, on a contract that can end on a client's calendar.

SPLASH'S READ: This isn't a scam story, and it isn't a "side hustle" story. It's piecework, rebuilt at internet scale — the same structure as garment work and farm labour, with a better interface.

The rate on the page is not what you'll earn. The estimate assumes a speed you don't control, on tasks priced by whoever goes fastest, for a client who can leave in March.

If you do this work, the only numbers that matter are your own: tasks completed per hour, times the per-task rate, minus the unpaid time you spend qualifying, waiting, and reviewing instructions. That's your wage. Not the headline.

And never pay to get paid. Not once. Not for "verification." Not in crypto.

Sources & claims: SOURCES, CLAIM BY CLAIM — The pay mechanics: much of this work is priced per task rather than per hour, so advertised hourly figures are estimates that depend on completion speed; one platform publishes an example of two workers on the same project at a quoted $30 an hour completing ten and three tasks respectively over the same time. Rates reported publicly range from roughly $50 an hour at peak for US contractors on Scale AI's Outlier platform, working long weeks on multi-month projects, down to payments of five to twenty-five cents per task for a worker in Colombia running several labelling platforms simultaneously. The scam layer: the Federal Trade Commission has reported that gamified 'task scam' reports roughly quadrupled each year from 2022, reaching about 20,000 reports in the first half of 2024 with approximately $220 million in reported losses across six months; the pattern involves fake earnings balances and demands that victims deposit their own money, often in cryptocurrency, to unlock withdrawals. The structural risk: Alphabet notified the Australian data services company Appen in January 2024 that its contracts would terminate effective 19 March 2024, ending an $82.8 million relationship and affecting roughly 2,000 contractors who had been performing search-quality and AI training work; the affected work stopped without the contractors being laid off in the conventional sense. The judgement premium: the higher-paid tier of this work involves evaluating model output for correctness and quality rather than simple labelling, which is why specialist projects in fields such as mathematics, law and coding command materially higher rates than the general task pool. The analysis — that per-task pricing transfers speed risk to the worker, that it makes the fastest worker the price-setter for everyone else, that the structure is piecework rather than employment, and the guidance to compute an effective wage from tasks per hour rather than advertised rates — is this piece's own. Source video: Websites That Pay You to Train AI — No Coding Required (Chill Dudes Stories, 13 November 2026) — used as the prompt only. That video is a ranked list of platforms and was not written up as a list; this article instead covers the labour economics the list depends on, and every figure above was checked against FTC data, company statements and reporting rather than taken from the video's earnings claims.