Data Annotation & Labeling

Data annotation, also called data labeling, is the work of teaching artificial intelligence to understand the world by drawing boxes around objects, tagging words, transcribing speech, ranking two chatbot answers, or flagging unsafe content so a model learns from human judgment. The person who does this work is a data annotator (also called a data labeler or AI data trainer), and data annotator jobs are the most accessible on-ramp into paid AI-training work. On this site today, data annotation roles advertise a typical Unavailable to Unavailable an hour (median Unavailable); micro-tasks pay far less and expert work pays more ( compare every kind of work), and you can start most of the legitimate platforms today with nothing more than a laptop, a stable internet connection, and (for the better-paying sites) the ability to pass a short unpaid assessment. This guide lays out exactly what the work is, which platforms actually hire, what each one pays by task, seniority and country, how to get started, and the very real traps to avoid. It is written for a person deciding whether this is worth their time, so it is specific about both the good money and the disappointing money.

The short version: this is real, legal, remote work that anyone can access, but it is not the "easy money from your couch" fantasy that ads imply. Effective pay is almost always lower than advertised pay because a meaningful slice of your time is spent hunting for tasks, sitting through unpaid assessments, and occasionally redoing rejected work. The workers who do well treat it like a skill (they qualify for higher tiers, get fast, and diversify across platforms), and the ones who quit disappointed expected a steady paycheck from a click-farm.

What Data Labeling and Annotation Actually Is

Modern AI models do not understand images, text, or audio on their own. They learn from millions of examples that a human has already interpreted. Data labeling (also called data annotation) is the act of creating those interpreted examples. If a self-driving car needs to recognize a pedestrian, a person first has to draw a box around thousands of pedestrians in thousands of photos. If a chatbot needs to give helpful answers, a person has to compare its responses and mark which one is better. You are, quite literally, the ground truth the machine is trained against.

The work splits into a few broad families, and which one you do determines both what tools you use and what you get paid.

Image and Video Annotation

This is the classic computer-vision work. You draw shapes over images and video frames so a model learns to detect and locate objects.

• Bounding boxes: drawing a rectangle tightly around an object (a car, a face, a product). In a tool like CVAT, you place two opposite corners, or use a four-point method anchoring the topmost, bottommost, leftmost and rightmost points of the object.
• Semantic and instance segmentation: coloring in every pixel that belongs to an object using polygons or masks. This is slower and more precise than boxes, and newer tools use the Segment Anything Model (SAM) to auto-generate a mask you then refine, which speeds things up considerably.
• Keypoint / landmark annotation: marking joints on a body, points on a face, or corners of a building for pose and shape models.
• Video tracking: following the same object across frames so the model learns motion.

Text and Language Annotation

Natural-language work covers everything from simple tagging to nuanced judgment.

• Named Entity Recognition (NER): highlighting people, places, organizations, dates and amounts in a body of text, often in a browser tool like Doccano.
• Sentiment and intent classification: labeling whether a review is positive or a message is a complaint versus a question.
• Text categorization and moderation: sorting content into topics or flagging policy violations.
• Prompt-and-response work and RLHF: writing prompts, writing ideal answers, and ranking two model outputs by which is more helpful, accurate, and safe. This is the fastest-growing and best-paid category, discussed below.

Audio and Speech Annotation

• Transcription: typing what is said in an audio clip.
• Speaker diarization: marking who spoke when.
• Audio classification: tagging sounds, accents, emotions, or background noise.

Search, Ads, and Social Media Evaluation

A related family (often called "search engine evaluation" or "AI rating") asks you to judge whether a search result, ad, or feed item is relevant and high-quality against a detailed rubric. This is the historical bread-and-butter of platforms like Appen and TELUS.

What a Real Workday Looks Like

You log into a platform dashboard, see which projects you are currently qualified for, open one, and read a guideline document (these can run to dozens of pages for a serious project). Then you work through a queue of tasks, one at a time, applying the guideline consistently. Your work is spot-checked by a quality-assurance layer or a more senior reviewer, and your "quality score" determines whether you keep the project and whether your rate holds. Task availability fluctuates: some days the queue is full, some days it is empty, which is the single biggest complaint across every platform.

How Much Data Labeling Actually Pays in 2026

Here is the honest range. On this site today, the data annotation roles that state a rate advertise a typical Unavailable to Unavailable an hour, with a median of Unavailable ( the roles it is measured from). That blends simple tagging with expert review and counts only roles that publish a rate, as aggregators like ZipRecruiter blend micro-task cents with salaried in-house annotators, so treat it as the advertised middle of the market, not what a beginner nets. The real distribution looks like this:

Two numbers anchor the extremes. On the low end, a widely cited research study ("Quantifying the Invisible Labor in Crowd Work") found MTurk workers earning a median of $2.83 an hour, about 39.5% of the US federal minimum wage, once you count the "invisible labor" of finding and managing work, per Unite.AI's summary of the research (published 5 October 2021). On the high end, DataAnnotation's own FAQ states coding work "starting at $40-$150+ per hour" and STEM and professional work at $40 to $125+ (read 10 November 2026), and Outlier/Scale AI's pages state ceilings of up to $150/hr for ML research, electrical engineering and quantitative finance experts (checked 8 Nov 2026).

The Advertised-vs-Effective Pay Gap (Read This)

Every rate above is an advertised rate. Your effective rate is lower for three structural reasons:

1. Unpaid assessments. To qualify for a project you often complete an unpaid test that can take one to two hours.
2. Task hunting. When your queue empties, you are refreshing and waiting, unpaid.
3. Rejections. Failed quality audits are usually unpaid, and low quality scores can get you dropped from a project.

The same 2021 study measured that invisible labor at about a third (33%) of a crowd worker's time ( Unite.AI), and the gap has not gone away: one 2026 review that tested DataAnnotation itself averaged about $15.30 an hour against the $14 to $20 the platform promised for its rating tasks, and concluded the $20 figure is "achievable only if you are very fast and consistently qualify for the highest-tier tasks" ( EarnifyHub, 28 March 2026). So plan around an effective rate well below any advertised one.

The Platforms That Hire, Ranked and Compared

Below is the practical rundown of the major legitimate platforms, what they pay, who they accept, and what real workers say. Pay is each platform's researched rate card on this site (it opens the card and the day it was checked) or its own stated rate; Trustpilot scores were read on 29 October 2026 and are a useful (if blunt) signal of how the worker experience is trending.

DataAnnotation.tech

Widely considered the best-paying accessible platform right now. Its own FAQ states general text, image, and audio work "starting at $25-$50 per hour", multilingual work at $25 to $40, and coding at $40 to $150+ an hour, with STEM and professional (law, finance, medicine) work at $40 to $125+, all with "higher rates based on strong performance" (read 10 November 2026); The catch: it takes workers only in the US, UK, Ireland, Canada, Australia, and New Zealand (the countries worker reports point to; it publishes no list, its page here), and you must pass a one-time Starter Assessment ("about an hour", per its FAQ, no retakes). Its Trustpilot is 3.9/5 (1,978 reviews, read 2 Nov 2026), and some workers report account suspensions and payout friction. It also has a long record of salary submissions on Glassdoor. Payment is via PayPal, typically weekly.

Outlier (Scale AI)

Outlier is Scale AI's main contributor platform, run beside Remotasks, which still takes sign-ups ( remotasks.com, read 10 November 2026). It states ceilings rather than rates (its pages say "up to"), and our rate card, checked against them on 8 Nov 2026, reads: Up to $7–$50/hr; ML, EE and quant finance up to $150/hr, with generalist and language evaluation at up to $7–$31/hr and coding at up to $34–$50/hr. Reviews put real earnings lower still: general writers without strong STEM credentials often fall to $12 to $18/hr ( Breaking Even, updated 14 September 2026; see also AIGigJobs). Its big advantage is country access: the 117 countries and territories its working guidelines list (effective 15 July 2026, read 29 October 2026). Its big caveats: rate compression is real and documented (the same Breaking Even review reports projects that paid $28 to $35/hr in early 2025 restructured to $18 to $22/hr by Q1 2026), and Scale AI/Outlier faced four legal actions between October 2024 and January 2025 (The Register, 24 January 2025) and a US Department of Labor wage investigation, which the department dropped in May 2025 (TechCrunch, 9 May 2025). Trustpilot: 3.9/5 (5,781 reviews, read 8 Nov 2026).

TELUS Digital AI (formerly Lionbridge / TELUS International AI)

A long-established, reputable platform for AI rating, search evaluation, and linguistics. Its own public list of openings, which our rate card read on 31 Oct 2026, pays $18–$50/hr experts; linguists and evaluators $3.50–$25/hr, and its AI content evaluator postings, the closest to annotation, run $4.50–$25/hr; ZipRecruiter tracks its US annotation listings. Work tends to be steadier than Outlier but often part-time-capped (for example, the classic "Rater" roles limit hours per week). See its worker portal and a detailed 2026 review.

Appen

Once the biggest name in the space, Appen now has 1.5/5 (371 reviews, read 1 Nov 2026) on Trustpilot, 82% of them one-star ( Trustpilot, read 29 October 2026). It still offers a huge variety of tasks and languages, but its rate card here reads $2.40–$100/hr; CrowdGen experts up to $450/hr, and its "Fair Pay" policy sets rates a little above the local minimum wage, so the same role pays very differently by country: its bilingual AI evaluators earn up to $95/hr, by language and country depending on where they live. Its search-evaluation qualification exams are notoriously hard. Treat Appen as a volume fallback, not a first choice.

Clickworker

A micro-task platform with a low barrier to entry. Standard tasks pay cents each; the real money is in UHRS (a Microsoft-owned judgment platform Clickworker routes you to), where per-HIT pay runs ~$0.05–$0.30/HIT; our rate card for Clickworker overall reads: Per task, cents to tens of euros; no hourly rate, and most workers report the low end. Trustpilot: 2.3/5 across 3,198 reviews (read 31 October 2026), with recurring complaints about payment and support. Good for filling gaps, not for a primary income.

Amazon Mechanical Turk (MTurk)

The original crowdsourcing marketplace. Each requester set its own pay per HIT, so there was never one wage: batch work (categorization, sentiment, audio) paid fast workers who used scripts and tools more than one-off surveys did. The research-backed median, counting unpaid time, is the sobering $2.83/hr figure above, and our rate card puts MTurk at ~$2–$3/hr effective (2018 study); closed. None of that is available to anyone arriving now: Amazon closed MTurk to new customers on 30 July 2026 and shut it down for everyone, its workers included, on 30 September 2026 (mturk.com, read 13 October 2026: "Amazon Mechanical Turk permanently closed on September 30, 2026"), and there is no Amazon successor, because AWS also put SageMaker Ground Truth and Augmented AI into maintenance, closed to new customers (announced 30 June 2026). It is gone for everyone, so if you are starting today it is not an option, and the numbers above are here only because MTurk is the reference point the research on effective pay was built on.

iMerit, Sama, and DataForce

These lean more toward structured employment or agency models than open marketplaces.

• iMerit hires annotators (often in India, plus the US) for structured computer-vision and NLP projects. Its own postings pay by country and language, and our rate card reads $2.50–$40/hr by country and language; clinical experts $120/hr, a spread that is largely geography. See iMerit careers.
• Sama is an impact-sourcing company that employs workers in Kenya and Uganda. It markets a "living wage" and says new team members see "an average of 3.6x increase in their earnings" ( Sama's own numbers, read 29 October 2026), and our rate card reads: Salaried; not posted (Kenya, 2023: $1.32–$3.74/hr), but TIME reported on 18 January 2023 that workers on its OpenAI content-moderation contract took home about $1.32 to $2/hr for grueling work ( HPCwire's summary). This is the low end of the global market, and a stark illustration of how geography sets the price.
• DataForce (by TransPerfect) runs linguistics, localization, and newer annotation projects, generally positioned as a higher-paying AI-project vendor relative to beginner micro-task sites ( platform roundup).

Pay by Country: Geography Sets the Price

The single biggest determinant of your rate, after skill tier, is where you live. Global platforms pay a US worker and an Indian worker very different rates for the same task, and the top-paying platforms geo-restrict entirely.

Sources: Second Talent's Philippines rate card (updated 14 September 2026), TIME (18 January 2023), and the rate cards above. The practical takeaway: if you are in the US, UK, Ireland, Canada, Australia, or New Zealand, start with DataAnnotation and TELUS; if you are elsewhere, Outlier and Clickworker have the widest access, and iMerit is worth pursuing as structured employment.

Skills and Qualifications You Actually Need

The honest answer on credentials: for generalist work you do not need a degree, but the best-paying platforms increasingly want one or an equivalent, and the expert tiers require real domain expertise.

• Baseline for everyone: reliable internet, a computer (not just a phone for most serious work), strong reading comprehension, and the discipline to follow a long guideline exactly. Attention to detail is the single most valuable trait, because your quality score is everything.
• For generalist DataAnnotation/Outlier work: strong written English is a genuine differentiator, since so much modern work is prompt-writing and RLHF where you evaluate and write clear prose. DataAnnotation's baseline is stated as a bachelor's degree "or equivalent real-world experience."
• For expert tiers (coding roles here advertise a typical Unavailable to Unavailable): an advanced degree or provable expertise in coding, mathematics, a science, law, finance, or medicine. This is where the money is, and it is gated behind specialist assessments.
• For language projects: native or near-native fluency in a specific language is a real, paid skill (DataAnnotation's FAQ states multilingual work at $25 to $40 an hour, read 10 November 2026, and multilingual roles across this site advertise a typical Unavailable to Unavailable).

The tools are learned on the job and are mostly browser-based: CVAT and Labelbox for vision, Doccano for text, and each platform's own custom interface for evaluation and RLHF. You do not need to know these before applying.

How to Get Started (and Pass the Assessment)

The path is the same across the good platforms: apply, pass an unpaid assessment, get placed in a track, then qualify for individual projects.

1. Apply on the official site only. Fill out the short application (DataAnnotation gives immediate confirmation, no resume wait). Never apply through a link from a Discord, Telegram, or social-media DM.
2. Take the assessment seriously. DataAnnotation's Starter Assessment takes about an hour; specialist assessments (coding, math, chemistry, biology, physics, finance, law, medicine, a specific language) take one to two hours. On DataAnnotation you get one attempt, no retakes (its FAQ, read 10 November 2026), so read every instruction and review before submitting.
3. Treat it as a quality filter, not a trivia test. The assessment is designed to find workers whose judgment is consistently better than the model's, per SmartOne's explainer. Graders reward: exact adherence to the guideline, internal consistency, thoroughness (do not cut corners), and clear written explanations. For writing-heavy assessments, well-structured, error-free, specific prose beats fast-and-sloppy every time.
4. Prepare deliberately. Practice against sample rubrics, make sure you are well-rested, and keep your workspace and time managed ( OperationsArmy assessment guide).
5. Once in, protect your quality score and diversify. Keep your quality high on your first project (it determines whether you keep it and unlock more), then apply to specialist assessments to climb into higher-paying tiers. Sign up for two or three platforms so you always have a queue with tasks.

A key mindset note: never cheat on an assessment. Using AI to write your answers on an AI-evaluation test is both easy for graders to detect and self-defeating, since the whole point is to measure your human judgment. Disqualification is permanent on no-retake platforms.

Getting Paid: Taxes and Payment Logistics

In the US and most Western countries, platform annotators are independent contractors (1099), not employees. That has real consequences:

• No taxes are withheld. You receive gross pay and owe taxes yourself. DataAnnotation, for example, pays via PayPal and issues a 1099 once you earn enough.
• Self-employment tax applies. On top of income tax you owe 15.3% self-employment tax (12.4% Social Security + 2.9% Medicare) on net earnings ( IRS, read 29 October 2026).
• Set aside enough of what you earn for federal, state, and self-employment taxes together, and pay quarterly estimated taxes (Form 1040-ES) if you expect to owe $1,000 or more for the year, the IRS rule (read 29 October 2026).
• 1099 thresholds. Payment processors report payments to the IRS, and platforms issue 1099-NEC forms (due to you by January 31). Even if you do not receive a form, the income is still taxable and you must report it.
• Payment methods are usually PayPal or direct deposit, typically weekly or biweekly. Confirm the method and cadence before you invest time in a platform.

Deductible expenses can lower your tax bill: a portion of your internet, a home-office deduction if you qualify, and equipment used for the work. Keep records.

The Honest Pros and Cons

Pros

• True flexibility. Work whenever you want, no set schedule, no boss on a call.
• Zero commute and low startup cost. A laptop and internet is the whole setup.
• No degree needed to start the generalist tier, and a real ladder upward for people with expertise.
• Genuinely interesting on the good projects. RLHF and prompt work can be intellectually engaging, not mindless.
• Legitimately scalable for experts. Coding roles on this site advertise a typical Unavailable to Unavailable, against Unavailable to Unavailable for annotation.

Cons

• Income is unpredictable. Task availability swings day to day; there is no guaranteed queue.
• Advertised pay overstates real pay because of unpaid assessments, task hunting, and rejections (the 2021 crowd-work study put the unpaid share at about a third).
• Rate compression is real. Rates on major platforms have fallen through 2025 into 2026.
• No benefits, no job security. You are a contractor; you can be dropped from a project without notice.
• Quality-score anxiety. One bad audit streak can cost you a project.
• The floor is grim. Micro-task and impact-sourcing work at the bottom of the market pays below minimum wage.

Red Flags and Scams to Avoid

The demand for this work has spawned a large ecosystem of scams. Legitimate platforms are free to join and pay you; scams try to extract money or data. Watch for these, per reporting from TIME and the FTC:

• Any upfront fee. No legitimate platform charges for "registration," "training materials," or "account activation." If they ask for money, it is a scam. Full stop.
• "Top up your account to unlock earnings." A classic task-scam pattern (especially via Telegram/WhatsApp): you are asked to deposit your own money to access higher-paying "tasks." You never get it back.
• Recruitment via DM. Real hiring does not happen through unsolicited Discord, Telegram, or WhatsApp messages from a "hiring manager."
• Gmail addresses and unverifiable companies. Legitimate platforms use company-domain emails and are searchable on Glassdoor, Trustpilot, and Reddit.
• Guaranteed earnings. "Earn $50/hr guaranteed" with no explanation of how is a lie; real platforms never guarantee your rate or task volume.
• Endless unpaid "assessments" that never lead to a job. Some operations harvest free labor (or your personal data) under the guise of an application test. A one to two hour assessment is normal; being asked to label hundreds of items "to qualify" with no clear path to paid work is not.
• Phishing for personal data. Never enter passwords for unrelated services, and be cautious about handing over ID documents before you have verified the company is real.

Verify before you invest time: search the exact platform name plus "reddit" and "trustpilot," confirm the URL is the official one, and apply only through it.

The 2026 Outlook: Is This Work Going Away?

A reasonable worry is whether AI will automate away the very work of training AI. The market data says the opposite, at least for now: the field is growing fast, but its center of gravity is shifting from cheap volume labeling toward skilled human judgment.

• The data collection and labeling market was valued at about $3.8B in 2024 and is projected to reach $17.1B by 2030, a 28.4% CAGR ( Grand View Research, read 29 October 2026). Technavio expects the data labeling and annotation tools market alone to grow by $2.75B over 2026 to 2030, also at a 28.4% CAGR (read 29 October 2026).
• RLHF is the growth engine. As models get better, demand shifts to skilled reviewers who rate helpfulness, spot contradictions, and identify safety issues, work that "commands premium rates and longer workforce ramps" per the market analysts. Open-ended comparative ranking and safety work is precisely the high-paying tier.
• Automation is eating the bottom, not the top. Tools like SAM auto-generate segmentation masks and box suggestions, which reduces demand for the simplest, cheapest labeling. The work that remains and pays is human judgment: nuance, edge cases, expertise, and taste.

The strategic implication for you is clear. Basic bounding-box-for-cents work is a shrinking, low-paid corner. The durable, well-paid opportunity is in the expert and RLHF tiers, where your domain knowledge, language skill, or writing ability is the product. If you are going to invest time in this, invest it in climbing toward those tiers rather than grinding micro-tasks.

Bottom Line

Data labeling and annotation is legitimate, accessible, flexible remote work that genuinely pays, and for the right person (a coder, a subject-matter expert, a strong writer in the US/UK/CA/AU/NZ) it can pay very well: coding roles on this site advertise a typical Unavailable to Unavailable, and DataAnnotation states coding work from $40 to $150+ an hour (its FAQ, read 29 October 2026). For everyone else it is a real but modest supplemental income, well below the advertised rate once the unpaid parts are counted, and a lot less at the micro-task bottom. Start with the reputable, well-rated platforms (DataAnnotation and TELUS if you are eligible, Outlier for broad access), take the assessments seriously, protect your quality score, diversify across two or three sites, set aside money for taxes, and never, ever pay to work. Do that, and you will get a clear, honest read on whether this is worth your time within a few weeks.