A PhD sociologist named James Maisiri got a recruiter's message offering him 600 rand an hour — about $37 — to teach an AI system how to design assessments, lecture undergraduates, and mark essays. That's the job he'd spent a decade learning to do himself. He turned it down, and his account of why, published by Rest of World on September 3, 2026, is the reason this question keeps landing in more inboxes every month.
Short answer: If you're offered money to train an AI on your own professional skill, decide with numbers, not panic: check what the platform actually pays (Mercor lists $12–$200+/hr by tier, DataAnnotation/Surge $25–$150+/hr), what exactly you'd be teaching the model, and whether that skill is your main income source. There's no universal right answer — only an informed one.
I test AI tools for a living, so when this story went around, I did what I do with any tool claim: I checked the actual numbers instead of reacting to the headline. In my testing of the three platforms Rest of World's reporting names most — Outlier, Mercor, and Surge AI (through its worker-facing brand, DataAnnotation) — the pay data, requirements, and what you're actually asked to hand over vary more than the "gig economy for experts" framing suggests. This isn't a step-by-step tutorial in the usual sense. It's the decision framework I'd want if that recruiter message landed in my own inbox tomorrow.
What you'll need
You don't need to have been contacted yet — the framework below works whether you're evaluating a real offer or just deciding whether to sign up in advance. Gather three things: your current hourly rate or salary, so you have a real number to compare against; a clear list of what the gig actually asks you to produce (grading rubrics, worked examples, correction notes — anything that captures your judgment, not just your typing); and 20 minutes to read the platform's own pay page rather than a third-party "top AI side hustles" roundup, most of which recycle numbers nobody verified. Maisiri's case is useful context, not a script — South Africa's youth unemployment sat at 47.4% in Q2 2026, which shapes his calculus in ways that won't match yours.
Step-by-step: deciding whether to train the AI that could replace you
1. Get the platform's own pay numbers, not a blog's
Mercor's own freelance AI training guide, updated June 23, 2026, lists entry-level trainer pay at $12–$25/hr, mid-level at $25–$53/hr, and expert specialists (medicine, law, finance) at $75–$200+/hr, with a stated national average of $31/hr. DataAnnotation, the consumer-facing side of Surge AI, advertises $25–$50/hr for general work and $75–$150+/hr for specialized coding, legal, medical, or finance tasks. Outlier, owned by Scale AI, doesn't publish specific rates at all — its site says only "competitive pay + quality-based rewards, paid weekly" and "$500M+ paid to experts" to date. If a recruiter quotes you a rate, check it against the platform's own page before you compare it to your day job.
2. Ask exactly what you'd be teaching the model
Maisiri's offer wasn't data entry — it was reproducing the judgment calls of a decade of teaching: how to weight a weak argument in an essay, how to design a fair exam question, how to explain a concept to a confused student. That's a materially different ask than labeling images or checking factual accuracy. Before you accept anything, get the task description in writing and ask whether you're rating existing AI output (lower stakes) or generating the reference answers the model trains on directly (higher stakes, and closer to what Maisiri walked away from).
3. Weigh the money against your actual market rate, not against zero
$37 an hour sounds generous next to South Africa's national minimum wage of roughly 30.23 rand ($2/hour), which is likely why the offer felt aggressive in that context. It looks different next to a PhD's normal consulting or teaching rate. Run the comparison against your own field's going rate, not against a platform's marketing or a country's minimum wage — that's the number that actually tells you whether you're being paid fairly for expertise or paid to undercut it.
4. Check what happens to your work after you submit it
Training data typically isn't licensed back to you, and most of these platforms' contributor agreements assign the AI company full rights to whatever you produce. Read that clause specifically — it's usually buried in the contributor terms, not the recruiting email — before you decide the money is worth it.
5. Decide on principle deliberately, not by default
Maisiri's own words capture the real tension: "When the opportunity to earn a livelihood collides with your principle on helping a technology that can compete with you, what do you do?" He didn't finish that sentence with a universal rule, and neither will you. The honest move is picking a side of that trade-off on purpose, with the numbers above in hand, instead of drifting into either "I'd never do that" or "I need the money" without actually having looked.
Example prompts you can copy
Use these to interrogate a real offer before you accept it, or to research a platform before you apply:
- "Here's a gig description for [platform]. Based on it, am I rating existing AI outputs or generating reference answers the model will train on directly?"
- "Compare [platform]'s advertised pay range to the typical hourly or project rate for a [your profession] in [your country]. Is this offer above, at, or below that rate?"
- "Summarize this contributor agreement's data-rights clause in plain English — who owns the work I submit?"
- "I do [your job]. Which of my specific tasks would be easiest for a general AI model to already do without my help, based on current tools?"
The last one is the most useful gut-check: it tells you whether the gig is teaching a model something new about your field, or just formalizing something a chatbot already does adequately.
Common mistakes to avoid
The mistake I see most in the "AI side hustle" coverage is treating all three major platforms as interchangeable, when their pay transparency alone is wildly different — Mercor and DataAnnotation publish tiered ranges, Outlier doesn't publish rates at all. Second is comparing an offer to a country's minimum wage instead of your own field's rate, which makes almost any gig look generous by comparison. Third is skipping the contributor agreement and assuming "freelance" means you keep rights to your work; in this industry, you usually don't. Fourth is treating the decision as permanent — Maisiri turned down one recruiter for one role; that's not the same as swearing off every AI-adjacent gig forever, and it's worth revisiting case by case.
Outlier vs. Mercor vs. Surge AI (DataAnnotation): what each actually discloses
| Platform | Stated pay range | Entry process | What they disclose |
|---|---|---|---|
| Mercor | $12–$25/hr entry, $25–$53/hr mid, $75–$200+/hr expert; ~$31/hr national average (Jun 2026) | Structured evaluation/assessment | Full tiered pay breakdown on its own site |
| Surge AI (DataAnnotation) | $25–$50/hr general, $75–$150+/hr specialized | Skills-based qualification, no interview | Tiered ranges by task type |
| Outlier (Scale AI) | Not publicly listed ("competitive pay + quality-based rewards") | Profile + skills verification, 30–90 min onboarding | Aggregate figure only ("$500M+ paid to experts") |
Tools that make this decision easier
If the real worry underneath the recruiter email is your job's exposure to AI generally, my breakdown of what’s actually happening to jobs and why the AI jobs apocalypse probably isn’t coming anytime soon both dig into the occupation-level data rather than the anecdotes. If you're specifically in a teaching, writing, or editing role like Maisiri, the safest job from AI may be writing covers why judgment-heavy work holds up differently than the task-level exposure scores suggest. Before trusting any productivity claim a recruiter or platform makes, my review of AI productivity gains is a useful reality check on what these tools actually deliver versus what they're marketed as. And if you decide the gig isn't for you but you're job-hunting anyway, my best AI tool for job searching guide and honest AI tool reviews are the more useful next stop than another training-data application.
My take
I don't think there's a clean verdict here, and I'd be suspicious of anyone who hands you one. The people most likely to be asked to train the AI that could replace them are exactly the people with the deepest expertise — that's the whole point of these platforms, and it's also what makes the trade-off uncomfortable. What I'd actually do: read the platform's own pay page instead of the recruiter's pitch, get the task description in writing before I get the offer letter, and decide against my real market rate, not a headline number. Maisiri walked away. That's a defensible call for a PhD with other paths forward. It might not be the right call for someone with fewer options and the same recruiter email sitting in their inbox — and pretending otherwise doesn't help anyone actually facing the decision.
Frequently Asked Questions
Is it worth it to train the AI that could replace me?
It depends on the specific pay, task, and your alternatives — there's no universal answer. Compare the platform's own published rate (not a recruiter's verbal quote) against your actual market rate, and check whether you're generating reference answers (higher stakes) or just rating existing outputs (lower stakes).
What did James Maisiri actually decide, and why?
Maisiri, a PhD in industrial sociology, declined a 600-rand-an-hour ($37) offer to train an AI to design assessments and grade essays, after a 45-minute AI-run interview. He described the core tension as principle colliding with the need to earn a living, and chose not to help build a system that could do his job.
How much do Outlier, Mercor, and Surge AI actually pay?
Mercor publishes $12–$25/hr for entry-level trainers up to $75–$200+/hr for expert specialists, with a $31/hr national average as of mid-2026. Surge AI's DataAnnotation lists $25–$50/hr general and $75–$150+/hr specialized rates. Outlier doesn't publish specific rates, only aggregate figures and "competitive pay" language.
Does refusing one AI-training gig mean I should avoid all of them?
No. Maisiri's decision was about one specific role that asked him to reproduce his own professional judgment directly. Rating or reviewing AI outputs is a different, lower-stakes task than generating the reference material a model trains on — worth evaluating separately, not lumped together.
What should I check before accepting any AI-training gig?
Three things: the platform's own published pay range (not a recruiter's number), the exact task you'd be doing (generating vs. reviewing), and the contributor agreement's data-rights clause, since most platforms keep full rights to whatever you submit.