Guide

Get Paid to Train AI: The Complete Guide to AI Training Jobs

How to get paid to train AI models in 2026. Platform comparisons, realistic pay rates, country eligibility, scam warnings and tax rules — updated July 2026 with the new expert marketplaces (Mercor, Alignerr).

December 6, 2025 · Updated August 16, 2026 · The AI Rankings Team

Quick answer: Getting paid to train AI is legitimate, but by mid-2026 it pays less for generalists and much more for specialists than it did a year ago — and the gap keeps widening. Realistic rates are roughly $12–20/hr for general annotation and $35–60/hr for coding, STEM, medical and legal experts, with the new expert marketplaces pushing the specialist ceiling to $80–250/hr. The best-paying entry point for most generalists is still DataAnnotation.tech, which advertises general projects from $25–30/hr and coding and STEM from $50–100/hr, against contributor reports nearer $14–20/hr general and $25–45/hr coding; it is reported to accept workers from six countries only — the United States, Canada, the United Kingdom, Ireland, Australia and New Zealand — though the company publishes no eligibility list. The biggest 2026 shift is Mercor, an AI-screened marketplace that places vetted professionals (coding, medicine, law, finance) into AI-lab contracts at $80–250+/hr — but only if you clear its screening. Outside the top tier, Appen, TELUS Digital and Prolific offer the widest access at lower rates. Treat all of it as variable supplemental income, not a salary: inconsistent task availability is the number-one complaint on every platform.

Every AI assistant you’ve used — ChatGPT, Claude, Gemini, Grok — was shaped by thousands of ordinary people evaluating responses, comparing outputs and teaching machines what “helpful” actually means. That work is real, it pays, and it isn’t going away. But the market has shifted hard in 2026, and most older guides are now wrong on pay, platforms and tax.

This guide covers the current state: the legitimate opportunities, the realistic earnings, the scams to avoid, and exactly how to get started based on where you live. If you want the higher-level hiring map instead — the job types, the recruitment platforms and salaried roles, and how to get hired — see our companion pieces, AI Training Jobs in 2026 and, for labelling work specifically, Data Annotation Jobs.

What AI training work actually is

AI training work — variously called RLHF (Reinforcement Learning from Human Feedback), data annotation, or AI feedback work — involves teaching AI systems what good responses look like.

The core tasks

Response evaluation. You see an AI-generated response and rate it on scales (typically 1–7) for accuracy, helpfulness, tone and safety. Does the response actually answer the question? Is it factually correct? Would a normal person find it useful?

Comparison tasks. You’re shown two AI responses to the same prompt, choose which is better, and explain why. These head-to-head comparisons are the foundation of how models improve.

Prompt creation. You write prompts designed to test AI capabilities — edge cases, tricky questions, scenarios where AI might struggle. The goal is finding weaknesses.

Content writing and editing. You write high-quality responses that become training examples, or you edit AI-generated text to fix errors, improve clarity or adjust tone.

Specialised annotation. Depending on your expertise: reviewing code for correctness and efficiency, evaluating medical or legal content for accuracy, rating creative writing, or assessing translations.

Image and audio tasks. Drawing bounding boxes around objects in images, labelling content, transcribing audio, or evaluating AI-generated images and audio.

A typical work session

You log into a platform, check for available tasks (this is the frustrating part — availability varies wildly), and work through a queue. Each task might take 30 seconds to 15 minutes depending on complexity. Most tasks involve reading carefully, making a judgement, and providing a brief written justification. The work requires focus but not advanced technical skills for general tasks. Specialist work — coding, medical, legal — requires demonstrable expertise and pays significantly more.

Why this work exists (and why it can’t be fully automated away)

Here’s the technical reality that keeps this work alive even as AI improves.

The model-collapse problem

When AI models train on AI-generated content, they suffer what researchers call “model collapse”: outputs become increasingly wrong and homogeneous, like photocopying a photocopy until the image degrades into noise. A 2024 study in Nature (Shumailov et al., vol. 631, pp. 755–759) demonstrated that indiscriminate training on recursively generated data causes irreversible defects, with the tails of the original distribution disappearing (Nature). The practical fix is continuous injection of human-generated feedback. This isn’t a temporary gap that better AI will close — it’s a fundamental limitation of how these systems learn.

Why human feedback specifically

RLHF works in three steps: humans see AI outputs and select which is better; those preferences train a “reward model” that predicts what humans prefer; the AI is then refined to score higher on that reward model. Remove step one — actual humans making actual judgements — and the loop breaks. AI can generate content, but it can’t reliably judge what makes content good for humans.

The automation question

Companies are increasingly using RLAIF (Reinforcement Learning from AI Feedback), where AI evaluates AI. It’s far cheaper than paying humans, and it’s eating the easy, high-volume, generalist end of the market. But it creates a circular dependency — you still need humans to train the AI that does the grading — and in high-stakes domains (healthcare, finance, autonomous vehicles, law), human judgement remains the standard because the liability of AI-only evaluation is unacceptable.

The clearest signal of where this is heading came from xAI. In September 2025 it laid off about 500 of its roughly 1,500 data annotators — a third of the team — and announced it would “surge our specialist AI tutor team by 10x,” hiring across STEM, finance, medicine and safety while scaling back generalist roles (TechCrunch).

The takeaway: this work isn’t going away, but it’s bifurcating. Generalists face automation and downward rate pressure; specialists face growing demand.

Realistic earnings expectations

Platform marketing doesn’t tell the full story, and 2026 rates are softer than the headline numbers suggest.

Advertised rates vs reality (mid-2026)

The trend since 2025 is consistent across reviews of the major platforms: the generalist floor has dropped, while the specialist ceiling has held but become harder to reach as more credentialed workers compete for a fixed pool of expert tasks.

The availability problem

This is the number-one complaint across every platform. Workers routinely report a feast-or-famine pattern: “Initially I was earning $1,500–2,000/week full-time. Over months, available work declined significantly.” Platforms surge-hire for specific client projects, those projects complete, and task volume drops until new contracts come in. You cannot rely on consistent full-time hours.

Realistic income scenarios

Supplemental income ($20–50/day): general annotation at $12–20/hr for 1–3 hours daily when tasks are available, or expert work at $35–60/hr for 30–75 minutes.

Part-time income ($500–1,500/month): multiple platforms, checking availability regularly, completing all qualification tests to unlock more project types, and maintaining high quality scores for priority access.

Reported high earners: software developers on DataAnnotation report $1,000–1,600+/week during busy periods working close to full-time; credentialed specialists (chemistry, law, medicine) command the highest hourly rates on Outlier. The pattern is consistent — the top earners hold graduate degrees in high-demand fields and treat this as serious work, not passive income.

Platform comparison table

Rates, ratings and country eligibility change frequently. Trustpilot and Glassdoor scores were read from the live pages on 16 August 2026; treat everything else as a July 2026 snapshot, and verify on each platform before applying. The clearest 2026 trend across the whole table: credentialed expert work is where the money moved, and the new marketplaces (Mercor, Alignerr) exist to route it there.

Top-tier platforms (best pay, strictest requirements)

These platforms offer the highest rates but carry geographic restrictions and qualification hurdles.

Mercor (the 2026 story)

• Pay: roughly $40–55/hr for RLHF/preference work, $85–110/hr for coding evaluators, and $90–250+/hr for domain experts in law, medicine, finance and science. The premium end is real but gated behind hard screening.
• Reputation: legitimate and heavily funded — Mercor raised a $350M Series C in early 2026 at a reported ~$10bn valuation, and contractors are paid on real, ongoing AI-lab contracts.
• Countries: global in principle, but access is decided by screening rather than geography.

Mercor is the biggest change to this market in 2026. Rather than a task queue, it’s an AI-screened talent marketplace: you build a profile, complete a ~20-minute AI interview (retakeable up to three times), and its matching system places you into ongoing contracts with AI labs — strong profiles get offers without applying. It targets expert knowledge work, not entry-level labelling, which is why the rates sit well above the rest of this table. The catch is the funnel: if you don’t get matched within about four to six weeks of finishing the process, you most likely didn’t clear the screening threshold, and there’s no task-by-task fallback to fall into.

Requirements: a strong, verifiable professional profile (software engineering, or a credential in medicine, law, finance or science); the AI interview; and, for the top rates, demonstrable production or graduate-level expertise.

Best for: experienced engineers and credentialed professionals who want the highest hourly rates and can pass a rigorous screen. Not a fit for absolute beginners.

Apply to Mercor →

DataAnnotation.tech

• Pay: DataAnnotation’s FAQ advertises general projects “starting at $25-$30+ per hour”, coding, STEM and professional projects at “$50-$100+ per hour” and multilingual work at “$20+ per hour”. Contributor reports on review sites that publish no sample size or method are consistently lower, at roughly $14–20/hr general and $25–45/hr coding. Treat the advertised figure as a ceiling and the reported one as a working expectation.
• Worker rating: Trustpilot 4.0/5 from 1,955 reviews, with reviews praising prompt, reliable payment and criticising inconsistent task availability. Full detail in our DataAnnotation review. Trustpilot notes the company has not recently invited reviews, so the sample is self-selected.
• Countries: DataAnnotation publishes no eligibility list on its own site. Third-party reviews consistently report six — the United States, Canada, the United Kingdom, Ireland, Australia and New Zealand — and that you can create an account from elsewhere and be notified if eligibility expands. Treat it as unconfirmed and check at signup.

DataAnnotation remains the most-recommended starting point for workers in the six countries it is reported to accept, mainly for its payment reliability (via PayPal) and the breadth of coding and STEM work. Pay has drifted down from the rates quoted in 2024–25 guides, so calibrate expectations accordingly.

Requirements: a bachelor’s degree or equivalent experience; a Starter Assessment (prepare carefully — retakes are not generally offered); valid ID matching an approved country; and an expertise assessment for specialist tracks.

Best for: software developers, STEM graduates, and credentialed professionals in medicine, law or finance who live in the six approved countries.

Apply to DataAnnotation →

Outlier (Scale AI)

• Pay: about $18–28/hr for standard RLHF work, rising to $35–60/hr for credentialed specialists (chemistry, law, medicine, STEM). Real-world rates span roughly $12–45/hr depending on task and qualifications.
• Worker rating: Glassdoor around 3.2/5 from 699 reviews, with the most common theme being declining work volume rather than non-payment. Outlier also scores 3.8 on Trustpilot from 4,544 reviews, though many of those carry Trustpilot’s “Invited” label — see our Outlier review for why the two numbers differ.
• Countries: global, but expert verification is required for the better-paid tracks.

Outlier is the contractor-facing brand of Scale AI, and that matters more than ever in 2026. In June 2025, Meta invested about $14.3 billion for a 49% stake in Scale, valuing it near $29 billion, and founder Alexandr Wang left to lead Meta’s superintelligence effort (TechCrunch). Several major customers — reportedly including Google (Scale’s largest) and OpenAI — cut their reliance on Scale afterwards over conflict-of-interest concerns. In the weeks that followed, Scale laid off about 200 employees and cut ties with roughly 500 contractors, with interim CEO Jason Droege saying the data-labelling business had scaled “too quickly” (Tom’s Hardware). The Outlier of 2026 is therefore leaner, more technical and much lower-volume: generalist work has thinned out as the platform pivots toward expert tasks in fields like medicine, robotics and finance, while Scale itself chases enterprise and government contracts. For contractors, that means real uncertainty around task availability — Outlier still recruits, but the steady generalist queues that once made it worthwhile are largely gone.

Requirements: valid government ID; résumé and LinkedIn profile; undergraduate-level expertise minimum (graduate degree preferred for top rates); and expertise verification in your claimed field.

Best for: graduate students and academics, especially in STEM. Those with PhD-level expertise in chemistry, physics, mathematics or medicine see the highest rates — when expert tasks are available.

Apply to Outlier →

Prolific

• Pay: an enforced minimum of $8/hr, with Prolific recommending researchers pay at least $12/hr; most studies land in the $8–15/hr range.
• Reputation: strong standing in the academic community; reliable payment and robust participant protections.
• Focus: academic research studies from institutions such as Oxford, MIT, UCL and Stanford.

Prolific connects researchers with paid participants under ethical-review standards. It always pays cash via PayPal — never gift cards — which is itself a useful scam benchmark. Availability is more consistent than AI-specific platforms, but maximum earnings are lower.

Requirements: valid ID; demographic screening (studies target specific populations); and attention checks (failing them reduces access).

Best for: anyone seeking consistent, ethical work with stable payment — more reliable than AI-specific platforms, with a lower ceiling.

Join Prolific →

Mid-tier platforms (wider access, lower pay)

These platforms accept workers from far more countries but typically pay less than the top tier.

Alignerr (Labelbox)

• Pay: roughly $20–50/hr for expert tasks, with credentialed specialists (e.g. a chemistry doctorate) reaching ~$150–200/hr on the rare top projects. Everyday project rates are lower.
• Availability and payment: contributors report inconsistent project availability and variable payment timing; verify current terms before committing time.
• Countries: restricted; access is not guaranteed even where the site is available.

Alignerr is the contributor arm of the data-labelling company Labelbox, and along with DataAnnotation and Outlier it’s now one of the three largest consumer-facing annotation names. Verified workers are paid weekly via PayPal. The premium headline rates are real but rare; most contributors see standard expert rates and the familiar feast-or-famine availability. Treat it as a credible second or third platform rather than a reliable primary income.

Requirements: profile and skills screening; credential verification for specialist tracks.

Best for: credentialed specialists who want another expert queue alongside DataAnnotation or Mercor.

Apply to Alignerr →

Appen

• Pay: around $10–20/hr on average, with some specialist projects higher.
• Countries: 170+ countries, 180+ languages.
• Track record: 25+ years operating; publicly listed (ASX: APX).

Appen offers the widest geographic access of any major platform and is strongest in multilingual text and audio. Note the recent turbulence: Google terminated its search-rater contract with Appen in early 2024 — Google had accounted for roughly 30% of Appen’s revenue — forcing cost cuts and layoffs. The company has been rebuilding through 2025–26 around AI-lab data work. Payment reliability (monthly via PayPal, Payoneer or bank transfer) is generally solid; the trade-off is lower rates than US-focused platforms.

Requirements: valid ID; language-proficiency tests; project-specific qualifications.

Best for: workers in countries excluded from top-tier platforms, and anyone with rare-language skills (low-resource languages command premium rates).

Apply to Appen →

TELUS Digital AI (formerly Lionbridge AI)

• Pay: about $10–20/hr; US raters in search/social-media evaluation typically earn $14–17/hr.
• Worker rating: Trustpilot 1.2/5 from 293 reviews, though projects often run for years.
• Countries: 100+ countries, 40+ languages; a community of more than one million.

TELUS offers more stability than task-based platforms, with multi-year search-quality and social-media evaluation projects. Note that the broader search-rater market contracted in 2024–25 as Google cut thousands of third-party rater roles, so this work is less plentiful than it once was.

Requirements: valid ID; project-specific qualifications; often local-language fluency plus English.

Best for: workers who want long-term, stable projects rather than variable task work.

Apply to TELUS Digital →

Clickworker

• Pay: $5–15/hr; many micro-tasks pay only cents.
• Reputation: mixed; fine for casual earning.
• Countries: global. Headquartered in Germany.

Clickworker provides access to Microsoft UHRS tasks and is best suited to students or casual earners rather than primary-income seekers. Many tasks pay very little, but volume is often available.

Requirements: minimal — ID and basic assessments; UHRS access requires separate qualification.

Best for: students, workers in countries without access to higher-paying platforms, and casual earners after flexible micro-tasks.

Join Clickworker →

Specialised platforms

Voice and audio work

Voices.com offers premium compensation for AI/TTS voice licensing; unlike per-task platforms, voice work can generate ongoing royalties from licensed recordings, though it requires a professional recording setup. RWS TrainAI hires globally for voice recordings and audio annotation, and is particularly valuable for speakers of less common languages and dialects.

Translation and localisation

Appen multilingual projects cover 70+ dialects across 30+ languages; low-resource languages can command 2–3x standard pay. Argos Multilingual runs 150+ languages for major LLM providers, with professional translators in legal, medical or technical fields earning the highest rates.

Coding-specific

The DataAnnotation coding track pays roughly $25–45/hr for code review, debugging and writing programming challenges, gated behind a technical assessment. The Outlier STEM track pays $35–60/hr for developers and computer scientists, with a graduate degree preferred. If coding is your specialism, it’s worth reading our guide to the best AI for coding to understand the tools and models you’ll be evaluating.

Lower-paying and limited-availability platforms

Scale AI / Remotasks

• Pay: $10–30/hr advertised.
• Parent company: Scale AI, valued near $29 billion after Meta’s 2025 investment.

Scale’s labelling business has been reshaped by Meta’s 2025 investment and subsequent customer departures, and Remotasks volume is low as a result. Task availability is the main practical issue for contributors.

Recommendation: if you use Remotasks, withdraw earnings immediately rather than letting balances accumulate. This is sensible practice on any gig platform.

Toloka

• Pay: $0.01–1.00 per task for crowd work, often translating to under $3/hr; $20 minimum withdrawal.
• Reputation: repositioned in 2025, but crowd-contributor pay remains low.
• History: originally a Yandex subsidiary, now Amsterdam-based.

Toloka has changed materially. In 2025 it raised a $72 million round led by Jeff Bezos’s Bezos Expeditions (with Shopify’s Mikhail Parakhin becoming executive chairman) and pivoted from cheap microtasks toward PhD-level expert data and AI-agent evaluation, building a network of 200,000+ vetted experts across 50+ domains. That’s a more credible, better-funded company than the old Toloka — but the well-paid expert work is largely invitation-based, and open crowd tasks still pay very little.

Recommendation: not a viable primary income for general crowd workers. The expert track is worth pursuing if you have verifiable domain credentials.

Handshake AI

• Pay: varies by project; contributors reported rate reductions following a January 2026 restructuring.
• Origins: the student-careers platform Handshake, which only moved into AI data work in 2025.

Handshake AI scaled fast — it told the Financial Times its annualised revenue rate hit about $1 billion by April 2026. A January 2026 restructuring of its flagship “Project HH” cut pay and offboarded workers, with some told the work was narrowing to specific professions (finance, health) and that their field no longer qualified. If you try it, treat it as speculative income and withdraw promptly.

Recommendation: project scope has narrowed to specific professions. Verify current terms and eligibility before investing time, and treat it as variable income.

Amazon Mechanical Turk (MTurk)

• Pay: $3–10/hr realistic; rates essentially unchanged for years.
• Geographic focus: primarily US and Indian workers.

MTurk is saturated. Good-paying HITs (Human Intelligence Tasks) are claimed within seconds, requesters can reject work without payment, and there’s little quality control on the requester side. Other platforms offer better pay and conditions.

Recommendation: not recommended for AI-training income.

Geographic availability: what’s open where you live

Most guides skip this, and it’s the single biggest factor in what you can earn.

Tier 1 (full access to the best-paying platforms)

United States, Canada, United Kingdom, Ireland, Australia, New Zealand. These six are the countries third-party reviews report DataAnnotation.tech accepts — the company does not publish the list — and they also have the full range of Outlier projects. If you live here, start with the top-tier platforms. Realistic earnings: $12–20/hr for general work, $35–60/hr for specialists.

Europe (EU/EEA)

Available: Appen, TELUS Digital, Prolific, Clickworker, Toloka. DataAnnotation and most high-paying Outlier projects are not available. Realistic earnings: $8–20/hr for basic work, $40–75/hr for expert work where available.

India

Available: Appen, TELUS Digital, Clickworker, plus local platforms (FlexiBench, Indika AI, Macgence, INFOLKS). The market is large but rates are modest. Realistic earnings: $3–15/hr depending on platform and task.

Philippines

Available: Appen, TELUS Digital, Innodata. Realistic earnings: $3–18/hr.

Africa

Scale AI ceased operations in Kenya, Nigeria and Pakistan in March 2024, so those markets are no longer served. Regional alternatives: Sama operates in Kenya and Uganda as an employer rather than a gig platform; South African platforms such as Enlabeler and Sebenz.ai pay well above local minimum wage. Realistic earnings are highly variable, and local platforms are often more reliable than international ones.

Latin America

Available: Appen, TELUS Digital, Clickworker; some Outlier projects accept Latin American workers, particularly for Spanish-language tasks. Realistic earnings: $5–20/hr depending on language pair and expertise.

Rest of world

If none of the above applies, your best bets are Appen (170+ countries), TELUS Digital (100+ countries), Prolific (global) and Clickworker (global). Expect lower rates than Tier 1, and focus on rare-language skills or specialist expertise to command better pay.

Requirements and equipment

Hardware

For voice/audio tasks: a quiet environment, a USB microphone (not a built-in laptop mic), and a webcam for some platforms.

Account requirements

A government-issued ID matching the country you apply from (platforms verify this); a PayPal account (most platforms pay via PayPal, some via Payoneer or bank transfer); a LinkedIn profile (required by Outlier, recommended elsewhere); and a résumé for specialist applications.

Skills and qualifications

General annotation (no degree required): strong reading comprehension, attention to detail, the ability to follow complex instructions, and consistent availability. Specialist tracks (credentials required): demonstrable programming ability for coding (a GitHub portfolio helps), a graduate degree for STEM, professional credentials for medical/legal, and native or certified proficiency for language work.

How to maximise your earnings

1. Specialise in high-demand fields. Generalist annotation is crowded and automating; specialists in software development, mathematics and physics, medicine, law, and chemistry/biology command 2–4x generalist rates. The xAI shift toward specialist tutors is the clearest signal of where demand is heading.

2. Maintain quality scores. Most platforms use quality scores that gate access to higher-paying projects, determine priority when tasks are scarce, and influence suspension risk. Scores are hard to recover once they drop, so get tasks right rather than rushing volume.

3. Complete every qualification. Each qualification test you pass opens a new task queue. Workers consistently report that completing all available qualifications significantly increases their available work.

4. Run multiple platforms. Given inconsistent availability, don’t rely on one. A typical Tier-1 combination is DataAnnotation (highest pay when available), Outlier (specialist work), Prolific (consistent baseline) and one backup (Appen or TELUS). Elsewhere, lead with Appen, TELUS Digital, Prolific and Clickworker.

5. Time your checks. New projects often launch Monday–Wednesday; weekends are quieter. Check platforms multiple times a day — desirable tasks are claimed within minutes.

Scam warning signs

The growth of AI-training work has attracted scammers. Protect yourself.

Red flags — never proceed

Upfront fees. Legitimate platforms never charge for registration, training or equipment. Cryptocurrency-only payment. Real platforms pay via PayPal, Payoneer or bank transfer; crypto-only is untraceable and favoured by scammers. WhatsApp or Telegram recruitment. Legitimate platforms use official websites and email; unsolicited chat-app messages are scams. Unrealistic earnings claims. “Earn $6,000/week with no experience” is a scam. Pressure to buy accounts. A black market exists for accounts from approved countries; participating risks permanent bans. Requests for sensitive information. Platforms need ID verification, not your bank login, passwords, or full government ID number beyond tax purposes.

Verification steps

Before applying anywhere: search “[platform name] reviews reddit” for real worker experiences; check Trustpilot (scores below 2.5 warrant caution); verify the exact website URL (scammers register look-alikes such as “data-annotation.tech” instead of “dataannotation.tech”); look for verifiable company history and press coverage; and confirm payment is cash via PayPal or bank transfer, not gift cards or crypto.

As Prolific states plainly: it will always pay cash via PayPal and will never pay with gift cards. Anyone claiming to represent an AI-training platform while offering gift-card payment is running a scam.

Tax obligations

AI-training income is self-employment income in virtually every jurisdiction. The rules changed in several countries for 2026 — here’s the current picture.

United States

All earnings are self-employment income. The big change: under the One Big Beautiful Bill Act (signed July 2025), the Form 1099-NEC reporting threshold rose from $600 to $2,000 from 1 January 2026, indexed to inflation from 2027 (OnPay summary). That means a platform may not issue you a 1099 unless you earn $2,000+ in a year — but you are still legally required to report every dollar of income, 1099 or not.

Obligations: self-employment tax of 15.3% (Social Security + Medicare); regular income tax at your bracket; and quarterly estimated payments if you expect to owe $1,000+ for the year. Deductions available include a home office, the work portion of internet and equipment, and related professional development. Set aside 25–30% of each payment for tax.

United Kingdom

This is self-employment income. The £1,000 trading allowance remains the only tax-free amount, and you must register for Self Assessment and file by 31 January if your trading income exceeds £1,000. A higher £3,000 Self Assessment reporting threshold has been announced but does not take effect until around 2029/30 — and it’s a reporting simplification, not a tax cut; you’ll still owe tax on anything over £1,000 (GOV.UK).

National Insurance has changed too: mandatory Class 2 NICs were abolished from April 2024 (you can still pay voluntarily at £3.65/week to protect your State Pension), and Class 4 NICs are now 6% on profits between £12,570 and £50,270, then 2% above (GOV.UK). Income tax applies at 20% (basic), 40% (higher) and 45% (additional). Since 2024, digital platforms report worker earnings directly to HMRC, so assume your income is already visible.

Australia

This is sole-trader income, reported on your individual tax return. You’ll likely need an ABN; you must register for GST only if turnover exceeds $75,000 (gross, not profit). Income is taxed at marginal rates — from 0% up to $18,200 to 45% above $190,000 — plus the 2% Medicare levy. From your second year, the ATO may enrol you in quarterly PAYG instalments if you owed more than $1,000 in tax on business income. Keep detailed records and lodge by 31 October (or later through a registered tax agent).

Other countries

The general principles hold globally: AI-training income is typically self-employment/freelance income, must be reported, may require registration as a sole trader or equivalent, and warrants setting aside 20–30% for tax. Consult a local tax professional for your specifics.

Get paid to learn AI (not just train it)

“Get paid to train AI” and “get paid to learn AI” are different searches with different answers. Training work (everything above) pays you to evaluate and label data. Learning programs pay you a wage or stipend while you build AI skills — and in 2026 there are more of them than ever, because governments and big employers are trying to close the AI-skills gap.

• Registered AI apprenticeships (US). In April 2026 the US Department of Labor launched a national push to fold AI skills into Registered Apprenticeships, backed by roughly $243 million in reskilling funding and a new AI Literacy Framework published in February 2026 (DOL). These are genuine earn-while-you-learn roles — you draw a wage from day one — not courses you pay for.
• Employer AI apprenticeships. Companies including IBM, Microsoft and Accenture run AI apprenticeships that pay roughly $50,000–$70,000 while you train, and several explicitly accept applicants without a degree. Programs like Apprenti pair on-the-job training with a recognised certificate (e.g. via NC State’s AI Academy).
• Paid AI fellowships. A growing number of AI labs and nonprofits run paid fellowships that place people in AI-focused roles while they upskill. Terms vary widely, so verify the sponsor, the stipend and the commitment before applying.

The key distinction from training gigs: apprenticeships and fellowships are usually employment or stipend income (often with benefits and a fixed term), not variable task-by-task pay. If a program asks you to pay to “learn AI and get hired,” treat it as a course — or a scam — not a paid opportunity.

Get paid to train AI from home

Almost every platform in this guide is remote by default — DataAnnotation, Outlier, Mercor, Alignerr, Prolific, Appen, TELUS and Clickworker are all work-from-home by design, with no office and no fixed hours. So “get paid to train AI from home” isn’t a separate category of work; it’s the normal way this work is done. What actually determines your from-home earnings is three things covered above: where you live (Tier-1 countries unlock the best platforms), your expertise (specialists earn multiples of generalists), and your setup.

For a reliable home setup you need a computer less than five years old, 8GB+ RAM (16GB recommended), stable 10+ Mbps internet, a current Chrome browser and a PayPal account — plus a quiet room and a USB microphone for any voice work (see Requirements and equipment). Two from-home cautions specific to remote seekers: never accept a “work from home training AI” offer that arrives unsolicited via WhatsApp or Telegram, and never pay an upfront fee for equipment or “training” — legitimate remote platforms cost nothing to join (see Scam warning signs).

AI testing for pay

“AI testing” covers two very different paid activities, and it’s worth knowing which one you’re suited to.

Model evaluation and preference testing. This is the mainstream, lower-barrier version — rating responses, comparing outputs, running preference tasks. It’s exactly the work the platforms above pay for: Prolific for structured studies, DataAnnotation and Mercor for evaluation tasks, and Appen/TELUS for search and social-media evaluation. No security background required; pay tracks the general and specialist rates already listed.

Adversarial red-teaming (AI bug bounties). The higher-paying, skill-gated version is trying to break models — finding jailbreaks, unsafe outputs or security flaws — through public bug-bounty programs:

• Anthropic runs a public program on HackerOne paying up to $35,000 for a new universal jailbreak technique, mirrored against its latest production model.
• OpenAI hosts bounties on Bugcrowd; xAI and Google (via its Vulnerability Reward Program) run their own AI-security intake.
• Gray Swan Arena runs public red-teaming competitions with prize pools historically $40,000+, and invites top finishers into a private, paid red-teaming network.

Red-teaming pays well but rewards genuine security, ML or domain expertise — it’s closer to freelance security research than to annotation. Note too that the most sensitive programs (for example OpenAI’s bio red-teaming) are invite-only and vetted, not open sign-ups. Everything here is legitimate paid testing; be wary of any “AI tester” listing that promises high pay for no skills and asks for money up front.

Where the work is heading

Market trajectory

The data-annotation tools market is worth roughly $3.1 billion in 2026 and is forecast to grow at 26–28% a year, reaching somewhere between $5.3 billion and $9 billion by 2030 depending on how the market is defined (Grand View Research; Technavio). Demand is strong; the question for workers is who captures it.

The specialist shift

The clearest evidence of a shift toward specialists is what the labs have done. In September 2025 xAI cut 500 generalist annotators — about a third of the team — and said it would prioritise specialist AI tutors, planning to expand that team tenfold across STEM, finance, medicine and safety (TechCrunch). Scale AI cut 200 employees and 500 global contractors in July 2025 (Tom’s Hardware). On the demand side, expert marketplaces have grown quickly against the same backdrop.

Automation impact

AI-graded feedback (RLAIF) is materially cheaper than human review, and labs have applied it to high-volume, low-complexity work. What the published record does not show is a lab removing human feedback entirely. Read the layoffs above as the best available evidence of where demand is moving; anything beyond that is forecasting, and we do not have a source for it.

Getting started: step-by-step

Step 1 — Assess your qualifications (Day 1). Honestly evaluate your location (are you in a Tier-1 country?), education, specialist expertise, languages, and weekly time availability.

Step 2 — Prioritise platforms (Day 1). Tier-1 + specialist expertise: DataAnnotation (specialist track), Outlier (STEM/expert), Prolific (baseline). Tier-1 + general background: DataAnnotation (general), Prolific, Appen. Outside Tier-1: Appen, TELUS Digital, Prolific, Clickworker.

Step 3 — Prepare your profiles (Days 1–2). Update LinkedIn to emphasise relevant expertise, prepare a résumé, have your government ID ready, set up PayPal, and test your internet speed (10+ Mbps).

Step 4 — Apply to your top 3–4 platforms (Days 2–3). Expect ID verification, a skills assessment, expertise-specific tests for specialist tracks, and writing samples for some platforms. Take assessments seriously — many (including DataAnnotation) don’t allow retakes.

Step 5 — Complete every qualification (Weeks 1–2). Each one opens new task types; treat it as investment in future earnings.

Step 6 — Establish a routine (Week 2+). Check platforms 2–3 times daily, track hours and earnings, set aside tax money immediately, and protect your quality scores.

Step 7 — Optimise and expand (Month 2+). Analyse which task types pay best for your time, add platforms if income is insufficient, and pursue specialist qualifications if your background supports it.

Frequently asked questions

How much can I realistically earn training AI in 2026?

For supplemental income ($20–50/day), expect 1–3 hours daily of general annotation at $12–20/hr when tasks are available. For $500–1,500/month, plan on 15–25 hours weekly across multiple platforms. Specialists with graduate-level credentials in coding, STEM, medicine or law earn $35–60/hr and report the highest weekly totals — but only when expert task volume is there, which isn’t guaranteed.

Do I need a degree to get paid to train AI?

Not for general annotation. But top-tier platforms prefer degree-holders, and the specialist tracks that pay $35–60/hr require demonstrable expertise — a graduate degree in STEM, medicine or law unlocks the highest rates.

Which AI training platform pays the most?

In 2026 the highest realistic rates come from the expert marketplace Mercor — roughly $85–110/hr for coding evaluators and $90–250+/hr for credentialed medical, legal and finance experts — if you clear its screening. Below that, DataAnnotation.tech and Outlier pay about $25–60/hr for coding and specialist work in the Tier-1 countries (US, Canada, UK, Ireland, Australia, New Zealand — reported rather than published in DataAnnotation’s case), and Alignerr occasionally reaches ~$150–200/hr for rare credentialed projects. Outside Tier-1, the best pay comes from specialist or rare-language projects on Outlier and Appen. General annotation, wherever you are, sits at $12–20/hr.

Get paid to learn AI — is that a real thing?

Yes, but it’s different from training work. Registered AI apprenticeships (backed by a 2026 US Department of Labor initiative) and employer programs at firms like IBM, Microsoft and Accenture pay a wage — often $50,000–$70,000 — while you learn AI skills, frequently with no degree required. That’s earn-while-you-learn employment, not gig work. If a “learn AI and get paid” offer asks you to pay first, it’s a course or a scam, not a job.

Can I get paid to test AI models?

Yes, in two ways. Lower-barrier evaluation and preference testing is the same work platforms like Prolific, DataAnnotation and Mercor already pay for. Higher-paying adversarial red-teaming — trying to break models — runs through public AI bug-bounty programs (Anthropic on HackerOne pays up to $35,000 per new universal jailbreak; OpenAI runs bounties on Bugcrowd; Gray Swan Arena competitions have $40,000+ prize pools). Red-teaming rewards real security or ML expertise, not general availability.

Is AI training work available in my country?

The best-paying platforms — DataAnnotation and most Outlier expert projects — are reported to accept only the US, Canada, UK, Ireland, Australia and New Zealand, though DataAnnotation publishes no eligibility list, so check at signup. Everyone else can use Appen (170+ countries), TELUS Digital (100+ countries), Prolific (global) and Clickworker (global), at lower rates.

Why is task availability so inconsistent?

AI companies hire annotators for specific projects. When a project finishes, those tasks vanish until new contracts create new volume. That project-based cycle produces the feast-or-famine availability every platform is criticised for.

Is money from training AI taxable?

Yes — it’s self-employment income in nearly every country. In the US, expect ~15.3% self-employment tax plus income tax (and note the 1099-NEC reporting threshold rose to $2,000 in 2026, though you must still report all income). In the UK, register for Self Assessment above £1,000. In Australia, report it as sole-trader income. Set aside 25–30% for tax.

Can I do AI training work alongside a full-time job?

Yes — it’s flexible and asynchronous, and many people work evenings or weekends. The catch is that quality scoring rewards focus, so tired late-night work after a long day can hurt the scores that gate your access to better-paid tasks.

How do I avoid AI training scams?

Never pay upfront fees, and never accept payment via gift cards or cryptocurrency only. Verify platform URLs carefully (scammers register look-alikes), and treat any recruitment via WhatsApp or Telegram as a scam. Legitimate platforms pay cash via PayPal or bank transfer.

Will AI automate this work away?

Partially. Generalist annotation is already being automated (via RLAIF) or shifted to lower-cost regions — xAI cut 500 such roles in 2025. But specialist expertise, high-stakes domains and the model-collapse problem keep human feedback essential. The work is concentrating toward expertise, not disappearing.

How long does onboarding take?

Roughly 1–7 days for application approval, 1–2 weeks to complete qualifications, and 1–3 weeks to first payment after completing work. Plan for 2–4 weeks from first application to meaningful earnings.

What equipment do I need to train AI?

A computer less than five years old, 8GB+ RAM (16GB recommended), stable 10+ Mbps internet, and a current Chrome browser. For audio/voice tasks, add a quiet space and a USB microphone. You’ll also need a PayPal account for payment.

Summary: is AI training work worth it?

Yes, if you have realistic expectations (supplemental income, not a salary); you’re in a Tier-1 country or hold specialist expertise; you can tolerate inconsistent availability; you treat it as serious work; and you prepare for tax obligations.

Approach with caution if you need predictable income; you’re counting on specific hours or earnings; you’re outside Tier-1 countries without specialist skills; or you expect easy money.

The bottom line: AI-training work is legitimate and still pays reasonably — especially for specialists — but 2026 has compressed generalist rates and reshuffled the platform landscape. Scale’s Outlier has shrunk to a leaner expert operation, while new AI-screened marketplaces like Mercor now route credentialed professionals into $80–250/hr contracts. The model-collapse problem means human feedback isn’t going away; it’s just rewarding expertise more and generalist volume less. Go in with the right expectations and the right platforms for where you live, and it can be a solid source of supplemental income.

For context on the AI products this work improves, see our rankings of the best AI models and best AI apps.

Platform rates, worker ratings and country eligibility were checked on 16 August 2026 and change without notice. There is no audited independent survey of pay in this category, so every hourly figure here is either a platform’s own advertised rate or a self-reported contributor estimate collected by review sites that publish no sample size or method — labelled as one or the other in the text. Company funding, ownership, layoff and revenue figures are sourced to the outlets linked inline. Tax thresholds are from the official government pages linked in that section and are general information, not tax advice.

← All guides