AI Training Jobs and Platforms Compared: Pay, Assessments and Reputation

AI training jobs are paid, usually remote contract tasks in which people write, rank, rate, check or record data that AI labs use to train and test their models, offered through contributor platforms such as Outlier, DataAnnotation, Alignerr and Mindrift. This page compares 14 such platforms, alphabetically, on the legal entity behind each brand, the pay each states on its own pages (checked 2026-11-09), screening, payment, contributor terms and court records, among them the final approval of the McKinney v. Scale AI settlement on 2026-10-30.

Contributor work is the human route within AI training data licensing: instead of licensing files that already exist, labs pay people through these platforms to create new data. Every figure below is the platform's own claim on the check date, never a confirmed earning, and no platform pays for its place on the page.

What AI training jobs are and who pays for the work

An AI training job is contract work in which a person produces or judges data for an AI model, and the money comes from AI labs that buy that data through the platform. The platform recruits, screens and pays the contributor, while the lab sets the task and receives the result. "Data annotation job" is the older name for the same work, from the time when most tasks were labels on images or text.

The platforms describe five kinds of work on their own pages (checked 2026-11-09):

• Writing: prompts and model answers, as Outlier lists under prompt writing.
• Ranking and rating: ordering or scoring model responses, as in Outlier's response evaluation and ranking and its Aether voice and image rating.
• Expert review: checking answers in a field, as DataAnnotation lists for coding, STEM, medical, legal and finance work.
• Adversarial testing: trying to make a model fail or produce harmful output, supplied to labs as red-teaming data.
• Recording: speech, video or first-person footage, as in Mindrift's egocentric video collection projects.

The method predates today's platforms. The InstructGPT paper of March 2022 describes "about 40 contractors on Upwork and through ScaleAI" who ranked 4 to 9 model responses per prompt. That ranking work feeds reinforcement learning from human feedback, in which a model learns from people's stated preferences between its outputs.

Fourteen contributor platforms side by side

The table compares 14 contributor platforms, alphabetically, on who the contributor contracts with, the work listed, pay as the platform states it, screening, payment and where it accepts contributors. A platform enters the table when its operator's legal entity and its contributor brand are on the public record; brands whose operator could not be confirmed are left out. Four rules govern how to read it:

• Claim: the platform's own figure on its own page on the check date, never an earning.
• Not recorded: no value was confirmed on the check date, which does not prove that none exists.
• Two figures: where a platform's own pages disagree, both appear side by side and neither is picked.
• Platform name: leads to the platform's or its operator's profile once that page is published.

The table compares the 14 platforms on operator, work, stated pay or scale, screening, payment and eligibility, as of 2026-11-09 (Mindrift re-read 2026-11-10); every pay or scale cell is the named platform's claim.

Who runs each platform and who the contributor actually contracts with

Most contributor platforms are brands, and the contract, the tax form and any lawsuit name a different legal entity behind the brand. Outlier's Terms of Use, effective 2026-05-11, are an agreement with Scale AI, Inc., a Delaware corporation, and its affiliates, including Smart Ecosystem, Inc., doing business as Outlier AI. The same page's footer reads "Last revised on: March 13, 2026", so the terms carry two dates. Alignerr contracts through ALIGNERR LLC and calls itself "powered by Labelbox". DataAnnotation is a trade name of Surge Labs, Inc., and the Handshake AI Fellowship runs through Stryder Corp. (d/b/a Handshake) and its affiliate Handshake AI Solutions, LLC.

Several brands belong to data vendors that sell finished data to labs. Toloka AI BV runs Mindrift, Invisible Technologies, Inc. runs Meridial, Appen Limited runs CrowdGen, Centific Global Solutions, Inc. runs OneForma, and TELUS International (Cda) Inc. runs the TELUS Digital AI Community. Shipd names two entities on its own pages: Datacurve AI, Inc. in its Contributor Terms and Datacurve, Inc. in its privacy policy. Prolific's pages read on 2026-11-09 name no legal entity at all.

The entity in the terms is the one a contributor can bring a claim against, and the court record follows it. McKinney v. Scale AI, the wage-and-hour settlement over Outlier work, names Scale AI, Smart Ecosystem and HireArt as defendants. Boggs names Stryder Corp. and Handshake AI Solutions, LLC, and Cavalier names Surge Labs, Inc. dba DataAnnotation. A brand name on a job ad says little until the terms show which company owes the pay.

How to read the pay figures platforms publish

Every pay figure on this page is the platform's own statement on its own page on the check date, and no contributor-confirmed earnings dataset exists to test it. "Up to" figures describe the best-paid listing: DataAnnotation states expert rates of up to USD 75 to 125+ per hour, checked 2026-11-09, and the Handshake AI Fellowship's page description reads "Earn up to $100/hr". An average has no stated population, as with Alignerr's USD 80 per hour, checked on the same date. A total paid out describes the platform, not any one contributor.

The platforms' own pages also disagree with each other, all checked 2026-11-09:

• Outlier paid out: USD 500M+ on the home page vs USD 600M+ on the about page.
• Outlier countries: 50 on the home page vs 52 on the about page, while its working guidelines list 117 locations.
• Alignerr experts: 100,000+ on alignerr.com vs 2.6M+ on labelbox.com vs more than 4 million in Labelbox's Series E post of 2026-10-14.
• Mindrift trainers: 20K in a stats block vs 10,000+ in the page description.
• Alignerr pay after deactivation: no payment for completed work after a rejected appeal vs "All Alignerrs are paid for completed work" except in extreme fraud.

Stated rates rise with the field, from Mindrift's entry rate of USD 15 to 30 per hour to Handshake listings of up to USD 150 per hour for software engineers, both claims checked 2026-11-09. Rates by field, monthly maxima and the gap between a stated rate and take-home pay are dated on the page on pay for AI training work, by platform and field. This page cites those claims only to compare platforms.

How each platform screens applicants before the first paid task

Platforms screen before any paid work: DataAnnotation uses a Starter Assessment that can be taken once, Outlier runs 30 to 90 minutes of onboarding with skill screenings, and Alignerr adds ID verification and, for some jobs, an interview or skills assessment. DataAnnotation states that its Starter Assessment takes about an hour, that specialized assessments take 1 to 2 hours and that results arrive within a few days (checked 2026-11-09). Outlier adds ID and location verification and prohibits VPN use, and its Aether voice and image projects use 15 to 60 minutes of onboarding plus an Outlier Interview of about 20 minutes.

Alignerr's process runs from Google sign-in and a resume upload to Persona ID verification, then for some jobs a "Zara interview or a skills assessment", then a calibration phase, after which work is scored on a composite of quality and speed. Mindrift's egocentric video projects are the exception: Mindrift states that they need no assessment and that setup time is unpaid (blog post of 2026-09-17). The other platforms' pages do not state whether assessment time is paid, and Cavalier v. Surge Labs alleges unpaid training and qualification time, an allegation pending as of 2026-11-09.

No screening steps are on record for the Handshake AI Fellowship, Mercor or micro1. What the tests contain and how platforms onboard contributors is covered on the page on AI trainer assessments.

Where the platforms accept remote contributors: countries, documents and languages

The platforms describe the work as remote, but each limits who may sign up: Outlier lists 117 working locations, the Handshake AI Fellowship is for US-based fellows, and Alignerr accepts applicants worldwide while its FAQ says only US identity documents are accepted. Outlier's Flexible Working Guidelines, effective 2026-07-15, do not list Kenya, Nigeria or Pakistan, and Outlier asks for a valid ID and a mobile phone from the country of residence. Handshake states that CPT and OPT may qualify and that STEM OPT is not supported, and Alignerr states that it offers no visa sponsorship.

Scale claims show how far some networks reach. TELUS Digital's annual report on Form 10-K for 2024 describes payments to one million AI Community contributors across more than 118 countries, and Shipd states that its contributors come "from 40+ countries" (checked 2026-11-09). DataAnnotation states no eligibility rule on its public pages. Mindrift's general eligibility page was not read; for egocentric projects it requires residence in a listed country.

Language is the other filter. DataAnnotation lists 43 language roles among the 133 on its job board, all stated at up to USD 40 per hour (checked 2026-11-09). Demand and pay by language are compared on the page on AI training work by language.

Payment schedules, payout methods and referral formulas

Platforms pay contractors per hour, per task or per bounty, on schedules they set: Outlier pays weekly on Tuesdays, Mindrift every two weeks, and DataAnnotation through PayPal a few days after a contributor requests a deposit. Outlier pays through PayPal, Airtm or ACH bank transfer, either hourly or per task, and Mindrift states that the amount per task is shown before a contributor starts. Shipd shows bounties of USD 80 to 220 per item but publishes no payment terms (all checked 2026-11-09). Prolific's pricing page of 2026-11-09 sets cash-out from USD 6 or GBP 6 through PayPal.

Referral formulas are published by two platforms. Handshake's Fellowship referral terms of 2026-09-24 pay 10 % of the referred fellow's hourly rate per paid hour for up to 6 months after account creation, capped at 40 times the bonus rate, which equals 4 times the hourly rate. Mercor states 20 % of the referral's earnings, capped at 4 times the referral's hourly rate, claimed generally within 7 days of account creation and paid only once the referral is hired and starts billable work (checked 2026-11-09). Outlier accepts referrals from the contributor dashboard for candidates the referrer can "personally vouch for" and publishes no bonus amount.

Section 3 of Outlier's terms withholds no tax except as required by law, so a stated rate is a gross figure. Tax forms, self-employment tax and filing abroad are covered in the guide to taxes for AI training contributors.

Contributor terms to read before signing

The terms that matter most before a first task are who owns the work, how disputes are decided, what the platform may record, and whether pay survives a deactivation. The table compares the five platforms whose contributor terms are on record, as of 2026-11-09; each row cites the terms, FAQ or privacy page it comes from.

Ownership is settled in the platform's favor on every page that addresses it. Outlier (s.4(b)) and Shipd (s.3.1) take an irrevocable assignment of all rights in the work product, and Alignerr takes a perpetual right to use interview material without any duty to pay for the interview. Outlier's section 10 sends disputes to individual AAA arbitration with a class waiver, which a US resident can opt out of within 30 days, and its confidentiality duty survives termination (s.5).

Monitoring is the third term to read. Alignerr tracks time with Hubstaff and periodic screenshots after a Persona ID check, and DataAnnotation, which publishes no public terms of service, has a privacy notice that allows collection of facial recognition data from contractors. Mindrift states plainly that there is "no employment contract and no fixed hours". ID checks, monitoring software and breach notices are covered on the page on contributor privacy on human data platforms.

Reputation on the record: court cases, complaints and breaches by platform

The court and regulatory record attaches to four operators: Scale AI for Outlier, in US courts and before EU data protection authorities, Surge Labs for DataAnnotation, Stryder for the Handshake AI Fellowship, and Mercor. The table lists each proceeding with court, docket, subject and status as of 2026-11-10; complaints that no court has decided are marked as allegations.

McKinney is the only proceeding with a public settlement amount. On 2026-10-30 the Superior Court of California, County of San Francisco granted final approval to a gross settlement of USD 12,500,000 for Outlier contributors residing in California, with payments reported on Form 1099 rather than as wages. Schuster v. Scale AI was voluntarily dismissed, not settled, and the Ramey settlement terms are not public. Cavalier, Boggs and White remain allegations; none has been decided.

The EU complaints come from noyb, which states that it and three partner groups filed five GDPR complaints against Scale AI over Outlier on 2026-10-27 and that the platform has more than 900,000 workers. The Mercor case concerns contributor data, not pay. Online reviews and forum threads are not treated as evidence here, and full case narratives sit on the page on contributor lawsuits against human data platforms.

Benefits, risks, downsides and alternatives of AI training work

AI training work pays some people well for flexible remote hours, but it is contract work on terms the platform sets and can change, with screening that may go unpaid and no guarantee of tasks.

Benefits

• Flexible hours: Mindrift states "no employment contract and no fixed hours", and DataAnnotation sets no minimum hours (checked 2026-11-09).
• Expert rates as claims: on 2026-11-09 DataAnnotation states up to USD 75 to 125+ per hour for experts, and Alignerr lists USD 40 to 120 per hour.
• Breadth of fields: DataAnnotation's 133 roles span languages, medicine, finance, engineering and law.

Risks

• Rights: Outlier and Shipd take an irrevocable assignment of the work, and Outlier's terms require individual arbitration with a class waiver.
• Monitoring: Hubstaff screenshots at Alignerr, facial recognition wording in DataAnnotation's privacy notice, and the noyb complaints over Persona checks at Outlier.
• Deactivation: Alignerr's FAQ says a rejected appeal ends eligibility for payment for completed work.
• Court record: misclassification allegations in Boggs and Cavalier, and the McKinney wage-and-hour settlement.

Downsides

• Unpaid time: Mindrift states that egocentric setup time is unpaid, and no platform states that its assessments are paid.
• Self-employment tax: in 2026, US contractors owe 15.3 % on net earnings of USD 400 or more (26 U.S.C. 1401).
• No task guarantee: per-task and per-bounty pay stops when projects end, and no platform publishes the hours available per contributor.

Alternatives

• Participant panels: Prolific researchers set pay above a stated minimum of GBP 6 or USD 8 per hour (help page of 2026-07-17).
• Employed roles, which carry withholding and benefits that contractor work lacks.
• Expert work in the contributor's own field outside AI training.
• Not taking the work at all, which is a valid outcome.

The terms table and the court record say more about a platform than its headline rate, which is a claim the platform can change at any time.

How this comparison is compiled: alphabetical order, check dates, no fees

Platforms appear alphabetically and whether or not any of them pays referral fees; no row is ordered, filtered or described by pay. Each claim is dated 2026-11-09 unless a re-read is noted, as for Mindrift on 2026-11-10, and the method is set out in how we source, date and check every figure.

Job boards, forum threads and other comparison sites are never used as sources. The page is information, not legal or tax advice.

Where AI training work meets the rest of human data

This page compares platforms; the pages below cover the companies behind them, the kinds of work, how contributors are classified and how AI companies get training data overall.

• human data for AI: the companies behind the platforms
• how AI companies get training data, with human work as one route
• company data partnership programs run by some of the same operators
• how DataPacts is funded, and why no platform pays for a place
• is AI training work legit: scams and how to check a platform
• contractor or employee: how contributors are classified
• expert contributor work by field for doctors, lawyers, coders and scientists

Frequently asked questions

Are data annotation jobs legit?

The work is real and paid by named companies, and the court record shows both a settlement and open allegations. McKinney v. Scale AI reached final approval on 2026-10-30, while Cavalier v. Surge Labs is pending as of 2026-11-09. Fake offers that imitate platforms exist; the legitimacy page listed above covers how to check one.

What qualifications do you need for AI training jobs?

No platform on record sets a general degree rule; screening decides. DataAnnotation lists 133 roles on 2026-11-09, among them 43 language, 19 medical and 14 finance roles, and tests applicants with a one-time Starter Assessment. The Handshake AI Fellowship requires US-based fellows.

Which AI training platform pays the most?

No platform can be named, because every figure is a claim that differs by field and changes over time. Outlier's own pages state both USD 500M+ and USD 600M+ paid out on 2026-11-09. The pay page linked above compares stated rates by field.

Are AI trainers employees or contractors, and how are they taxed?

The contributor terms on record treat them as independent contractors. In the US, the Form 1099-NEC threshold is USD 2,000 for payments made after 2025-12-31, and self-employment tax is 15.3 %. The UK trading allowance is GBP 1,000 per tax year, and EU states must transpose Directive (EU) 2024/2831 by 2026-12-02.

Can you rely on Reddit threads and job boards like Indeed for AI training jobs?

Not as evidence: forums and job boards repeat listings and claims without primary sources. DataPacts uses only the platforms' own pages and court records, and lists no jobs. A listing can be traced back to the platform's own site and to the entity named in its terms.

What is the difference between an AI trainer and an AI rater?

A rater is one kind of AI trainer: the person who scores or ranks model outputs. Outlier lists response evaluation and ranking next to prompt writing and expert tasks. What raters do and who hires them is covered on AI raters and evaluation data.

Does Mercor hire AI trainers?

Mercor.io Corporation runs an expert marketplace for AI labs and stated a network of "nearly five million" experts on 2026-06-25. Its expert referral pays 20 % of a referral's earnings up to a cap (checked 2026-11-09). White v. Mercor, which pleads a 26 U.S.C. 7434 subclass, is pending as of 2026-11-09.