Get Paid to Train AI: Outlier, DataAnnotation, Mindrift, Alignerr and Others — What They Pay and the Red Flags

Yes, you can get paid to train AI, and the big platforms are real companies that really pay. Outlier, DataAnnotation, Mindrift, Alignerr and Prolific pay people by the hour or by the task to write prompts, rate chatbot answers, catch mistakes and solve problems in their field. The catch is everything around the pay: unpaid tests, work that vanishes overnight, accounts closed without a clear reason, and a flood of fake "AI trainer" recruiters who want your money instead of your time.

I've been making money online since 2010, starting as a freelancer and copywriter, and I've seen this movie before with content mills, micro-task sites and survey apps. The pattern is always the same: a new kind of online work appears, the first wave of people earns well, then the platforms get pickier, the rates split into "experts" and "everyone else", and scammers move in. AI training is somewhere in the middle of that story right now.

So this isn't a "make $50 an hour from your couch" piece. Below are the platforms and who owns them, the pay they state on their own pages and job listings, how they pay you, the onboarding tests, the complaints that keep coming up (including lawsuits against Scale AI), the scam patterns and the tax side. I haven't worked on these platforms myself, so there are no "I earned" claims here, just the facts and my read of them.

What "training AI" actually means

Training AI as a side job means producing the human judgment that AI companies can't generate on their own. Models learn from examples and from feedback, and somebody has to write those examples and grade those answers. That somebody is a contractor working from home.

The work usually falls into a few buckets:

• Rating and ranking. You see two or more chatbot answers to the same prompt and pick the better one, explaining why. This is the core of RLHF (reinforcement learning from human feedback), the technique that taught chatbots to sound helpful.
• Writing prompts and ideal answers. You write a tricky question and a model answer, or you write a prompt designed to make the model fail ("stumping" it).
• Fact-checking and editing. You go through an AI answer line by line, flag errors and rewrite the weak parts.
• Coding tasks. You write coding problems and solutions, or test AI-generated code for correctness and speed.
• Expert reasoning. Doctors, lawyers, mathematicians and scientists write graduate-level questions and check whether the model's reasoning holds up.
• Classic data labeling. Tagging images, transcribing audio, drawing boxes around cars. This is the older, cheaper end of the market.

Notice the trend: the old labeling work paid pennies, while the new generative-AI work pays for expertise. That's why job listings now ask for degrees, coding tests and professional experience. A general "I'm good with computers" profile still gets work on some platforms, but it competes with a lot of people.

The main platforms and who's behind them

Most AI training platforms are fronts for bigger data companies, and knowing the parent company helps you judge them. A brand-new site with no known owner deserves far more suspicion than a platform backed by an established firm.

Outlier (Scale AI)

Outlier is the contributor platform of Scale AI, one of the largest AI data companies. In June 2025 Meta agreed to buy a 49% non-voting stake in Scale, and Scale's founder Alexandr Wang moved to Meta. After the deal, Google and OpenAI reportedly pulled back work from Scale, and in July 2025 Scale cut about 14% of its staff, so Outlier today is leaner and more focused on expert work than in its boom days. Outlier hires generalists (writers, editors, students from writing-heavy fields) and experts in medicine, law, coding, math and machine learning. Scale also runs an older platform, Remotasks, which handled image labeling and similar work; more on what happened to it below.

DataAnnotation

DataAnnotation (dataannotation.tech) is one of the most visible names in the space, largely because its job ads run everywhere. It focuses on chatbot training: writing, rating and coding tasks. Its listings are short on company details, which makes people nervous, but its job ads clearly state pay, eligible countries and that it never asks applicants for money.

Mindrift (Toloka)

Mindrift is operated by Toloka AI, a data company that has worked in AI data since 2014. It targets domain experts: people with real knowledge in a field who can write and check AI content at a professional level.

Alignerr (Labelbox)

Alignerr is operated by Labelbox, a data platform for AI teams. It hires freelancers in dozens of fields, from software engineering and biology to language specialists, and runs its own assessments.

Prolific

Prolific started as a platform for academic research studies and now also supplies human feedback for AI projects. It's different from the others: you take part in studies posted by researchers and companies, with a pay floor set by the platform. It's less "job" and more "paid participation", but it's one of the easiest entry points.

Others worth knowing

• Appen and TELUS Digital. The veterans of search-engine rating and data labeling. They still recruit raters, but the market has shifted: Google ended its multi-million dollar contract with Appen in early 2024, a reminder that even big vendors lose work overnight.
• Mercor. A fast-growing AI recruiting platform that places experts on AI projects. It made headlines in November 2025 when Forbes and Business Insider reported that it ended a large project early and offered workers a new one at a lower hourly rate; Mercor called the claims inaccurate and said that's normal for project-based work.

What they pay: the comparison table

Stated pay on AI training platforms ranges from about $8 an hour on the low end to $150 an hour or more for rare expertise, but the number in the ad is a ceiling, not a promise. The table below uses only the platforms' own pages, help centers and job listings. Where I couldn't confirm something on a primary page, I left it out.

A few things jump out. First, the Outlier generalist rate isn't one number: the official wording is "up to $12.50–$34.50" per hour, based on quality, accuracy and project, with lower rates for onboarding. Second, Mindrift pays per approved task, not per hour logged, so the hourly figure depends on how fast you work and how many tasks pass review. Third, DataAnnotation's often-quoted "around $20 an hour" for general work comes from third-party reviews, not from a page I could verify, so treat it as reported, not official.

Fees matter too. Mindrift says it takes no commission, though Payoneer and Tipalti may charge their own transfer fees. Prolific pays out to PayPal, and PayPal may take a cut on currency conversion. If you get paid in dollars and live elsewhere, that conversion spread quietly eats a few percent of every payout.

How the onboarding tests work (and why they're unpaid)

Almost every legit AI training platform makes you pass a skills assessment before you see paid work, and on most of them that test is unpaid. Mindrift says so directly: the screening assessment is unpaid because it's a skills check and nothing you write is delivered to a client.

The typical path looks like this:

1. Sign up and verify your identity. Most platforms run an identity check before you see paid work. Expect to show a government ID from the country you live in.
2. Take a general assessment. Writing, reasoning, fact-checking or, for coders, LeetCode-style problems you must also explain. DataAnnotation's coding listings mention exactly that kind of test.
3. Wait. DataAnnotation warns that applications can take up to two weeks to process because of volume. Some people hear back fast; some never hear back at all.
4. Take project-specific qualifications. Each new project often has its own guidelines and test. On Outlier, onboarding and similar "non-core" phases can pay less than the core rate.
5. Get tasks, if a project needs your profile. Passing the test doesn't guarantee work. It puts you in the pool.

This is where a lot of frustration comes from. You can spend a few evenings on assessments and still end up with an empty dashboard. That's not necessarily a scam; it's how project-based work functions when a platform recruits a big pool and calls on people as clients need them. But it does mean the real cost of entry is your time, and you should budget it like money.

The complaints that keep coming up

The three biggest complaints about AI training platforms are unpaid time, work that disappears without warning, and sudden account deactivations. None of them is proof that a platform is fake, but together they explain why so many reviews read like breakups.

Projects drop overnight

Outlier itself says projects end when a client has all the data it needs, and that there will be periods with no work available. In practice, that means a project you've been working on for weeks can disappear from your dashboard between one evening and the next. The Mercor story above is the same pattern at a bigger scale.

Whole countries get cut off

In March 2024 Rest of World reported that Scale's Remotasks shut down entirely in Kenya, Nigeria and Pakistan and stopped taking new sign-ups in several other countries, with workers getting little or no explanation. Scale blamed "enhanced security protocols" and apologized that Kenyan contractors got no advance notice. If you live outside the main English-speaking markets, that's the risk to keep in mind: access can depend on where you live, and it can change.

Lawsuits over pay and classification

Scale AI faced three worker lawsuits in about a month, according to TechCrunch: two over alleged wage theft and misclassifying workers as contractors, and a January 2025 class action by Outlier contributors who said they were made to write disturbing prompts without adequate psychological support. Scale said it would defend itself and pointed to its safeguards, including opt-outs and advance notice of sensitive content. A US Department of Labor investigation into Scale's compliance with federal wage law was reported in March 2025 and dropped in May 2025. In May 2025 Surge AI, another AI data firm, was hit with a similar misclassification suit in California.

What does this mean for you? Mostly that you're a contractor, not an employee, and the whole industry is still arguing about whether that's fair. As a contractor you don't get minimum-wage protection, overtime or paid breaks, and the platform can end the relationship any time. Plan around that instead of hoping it won't happen.

Deactivations with money still in the account

Forums and review sites are full of people whose accounts were closed for "quality" or "policy" reasons, sometimes with an unpaid balance. I can't verify individual stories, and some deactivations are clearly deserved (using AI to write answers on a project that bans it is the classic one). But it happens often enough that you should act as if it could happen to you tomorrow.

Is it worth it? An example of real hourly pay

Your real hourly rate on an AI training platform is almost always lower than the stated rate, because unpaid time and taxes come off the top. Here's an example to show the gap. These numbers are made up for illustration, not a result or a benchmark.

Say a project pays $25 an hour and you log 10 paid hours a week. That's $250. But you also spent about 3 unpaid hours that week reading new guidelines, taking a qualification and refreshing an empty dashboard. Spread $250 over 13 hours and you're at about $19.23 an hour.

Now taxes. If you're a US contractor, self-employment tax applies on top of income tax. A rough estimate on that $250 (net earnings are 92.35% of the total, taxed at 15.3%) is about $35.32, leaving roughly $214.68, or about $16.51 per hour actually spent. And that's before income tax.

Per-task pay works the same way. A task worth $8 that's estimated at 20 minutes pays $24 an hour if you hit the estimate. If it takes you 35 minutes because the guidelines are long and you're careful, the same task pays about $13.71 an hour. Speed matters, but rushing gets you flagged for quality, so the real skill is getting fast without getting sloppy.

And then there's volatility. If you work three good weeks at $250 and the fourth week the project dries up, you made $750 for the month, which averages about $187.50 a week. That's why I'd never build a budget on this kind of income. Use it as a flexible extra, the way I'd treat any gig work. To play with your own numbers, the earnings calculator helps you see how hours, rates and gaps change the monthly total.

Fake "AI trainer" recruiters: the red flags

The most dangerous thing in this niche isn't low pay; it's fake recruiters using the "AI training" hype to steal money. The rule that catches nearly all of them is simple: a real platform pays you, you never pay it.

The FTC has tracked this for years. Reports of so-called task scams went from almost nothing in 2020 to about 20,000 in just the first half of 2024, and reported job scam losses topped $220 million in those six months. The script is almost always the same: a vague text or WhatsApp message offers easy online work, you get small payouts for simple "tasks" to build trust, and then you're asked to deposit your own money, usually crypto, to "unlock" the next level. Once you send it, it's gone.

Red flags to watch for:

• Unsolicited messages on WhatsApp, Telegram or text. Legit platforms recruit through their own sites and job boards. The FTC says real employers don't recruit by random text.
• Any fee. A "starter kit", a "certification", a "deposit to unlock tasks", a "training fee". Mindrift and DataAnnotation both state they never charge you. Nobody legit does.
• Crypto everywhere. Payments in crypto, deposits in crypto, "your balance is in USDT". The FTC calls crypto the payment method of choice for task scams.
• Pay quoted per day with no job details. Real listings describe the task, the skills, the countries and the pay range.
• Look-alike domains and personal emails. Legit platforms write from their own domain (Mindrift's is mindrift.ai), never from a free email account. Check the address letter by letter.
• Being paid to like, rate or review products. The FTC says it's illegal and honest companies don't do it. Rating chatbot answers inside a platform is fine; boosting product reviews for a stranger is a scam.
• A check you're asked to deposit and partly send back. That's a fake check scam, no matter how "AI" the job title sounds.

If something smells off, stop replying and report it. In the US that's ReportFraud.ftc.gov; most countries have a similar consumer agency or police fraud line.

The fear that stops people: "What if I waste my time?"

The fear that keeps most people away from AI training work is wasting hours on tests and getting nothing back. It's a reasonable fear, because it really happens. But the answer isn't to avoid the work; it's to cap how much time you risk.

Psychologists call our tendency to feel losses more strongly than equal gains "loss aversion". Two hours on an unpaid test feels like a big loss, while two hours scrolling feels like nothing, even though the time is the same. The trick is to treat the test as a fixed, limited bet:

• Give each platform one serious attempt. Set aside one evening per platform, do the assessment properly, then move on. No endless re-reading of guidelines before you even start.
• Apply to three or four platforms in parallel. Not to juggle them all, but because one empty dashboard hurts much less when another one has work. Diversification works for gig income the same way it works for multiple streams of income in general.
• Set a stop rule. For example: if no paid task shows up within a month after passing, stop checking daily. Check once a week and put your energy elsewhere.

The opposite trap is real too: the sunk-cost feeling of "I've spent 30 hours on this platform, I can't quit now." If a project has dried up and the rate on the next one is worse, the hours you already spent don't oblige you to stay. Treat each week on its own merits.

Tricks that give you an edge

The people who earn steadily on AI training platforms usually aren't smarter; they understand how the platforms decide who gets work. A few things most guides don't mention:

• Your profile is your rate card. Expert tracks pay several times more than generalist ones on the same platform. On Outlier that's the difference between up to $34.50 and up to $150 an hour. If you have a degree, a professional license or real coding skill, put it front and center and apply to expert listings directly.
• Language is an underrated edge. Platforms recruit native speakers of many languages, and these pools are often less crowded than English generalist work. If you're fluent in a second language, look for those listings specifically.
• Read the guidelines like a contract. Most quality flags come from breaking a rule buried on page 12, not from bad writing. Ten extra minutes with the guidelines protects weeks of income.
• Never use AI to do the AI-training task unless the project allows it. Platforms run detectors and reviewers look for it. It's the fastest way to a deactivation.
• Watch for incentives. Outlier says contributors earn on average 7.5% more through its "Missions" bonuses, and the top quarter averages 11%. DataAnnotation's coding listings mention bonuses for high quality and volume. These are often the difference between an average and a good month.
• Referral programs exist, so use them honestly. Some platforms, Outlier among them, run referral rewards that typically pay only once the person you refer starts doing real, approved work. Refer people who'd actually be good at it; referring yourself or fake accounts gets everyone banned.

Taxes when you're paid as a contractor

AI training platforms pay you as an independent contractor, which means nobody withholds tax for you and you're responsible for reporting and paying it yourself. That's true in almost every country, even if the details differ.

In the US, Outlier and Alignerr describe their roles as 1099 freelance contracts. A 1099-NEC is the form that reports nonemployee pay, and for payments from 2026 a platform only has to send one once it has paid you $2,000 or more in the year (the old threshold was $600). Your income is taxable whether or not you get one. If your net self-employment earnings reach $400 or more in a year, you owe self-employment tax of 15.3% (12.4% for Social Security and 2.9% for Medicare) on top of income tax, and you may need to make quarterly estimated payments to avoid a penalty.

Outside the US, the general rule is the same: contractor income from a foreign platform is still your income where you live. The UK, EU countries, Canada and Australia all have their own self-employment rules and thresholds, so check with your local tax office or an accountant. A simple habit works everywhere: move a fixed share of every payout into a separate account the day it lands, and you won't be surprised at tax time.

A practical note on payouts: Prolific requires the name, residential address and primary email on your PayPal to match your Prolific account, and you have to be in your listed country of residence when you cash out. If you don't have one yet, open a PayPal account in your own name before you apply anywhere, since most of these platforms pay through it.

How AI training compares with other online work

AI training is one of the fastest ways to earn online without an audience or a startup budget, but it's also one of the least secure. Here's how I'd place it next to the other options I write about.

• Versus freelance writing. AI training pays you to write and edit too, but the platform owns the client relationship. Freelance copywriting is harder to start but builds clients and a reputation you keep. My guide on how to get paid for copywriting covers that route.
• Versus other AI side hustles. Training is the "sell your time" end of the spectrum. Clipping, UGC and automation services can grow beyond your hours, with more risk. I compared ten of them in AI side hustles that actually pay.
• Versus a remote job. A salaried remote role gives you stability and protections that contractor work doesn't. If that's what you're after, read the rise of remote work for the real numbers and where those jobs are.
• Versus gig marketplaces. On Fiverr you set your own prices and keep your buyers, but you have to sell. If you'd rather start there, see how to make money on Fiverr without skills.

My honest take: AI training is a good bridge. It pays real money for skills you already have, it needs no audience, and it teaches you a lot about how AI models fail, which is useful for anything else you build online. Just don't mistake a bridge for a house.

Where to start this week

Start by picking the two or three platforms that match your profile and country, and set a time limit on the testing phase. A sensible order:

1. Check eligibility first. Look at each platform's country list and payout options before you spend time on tests.
2. Set up a payout account in your own name. PayPal covers most of them; Mindrift may need Payoneer depending on your country.
3. Apply through the official site only. Type the address yourself. Ignore anyone who contacts you first on WhatsApp or Telegram.
4. Do each assessment once, properly. One evening per platform, then move on.
5. Track your real hourly rate and withdraw every payout. Those two habits protect you from most of the bad stories you'll read online.

If it works, great: you've got a flexible income stream for evenings and weekends. If it doesn't, you've lost a few evenings and learned which skills the AI industry is paying for right now, and that's useful for whatever you try next.