Human Data for AI: The Companies, Expert Contributor Work, Pay and Platforms

Human data for AI is training and evaluation data that people are paid to create, rank, check or test, supplied to AI labs mostly through human data companies. This hub lists 21 such companies as of 2026-11-09, with the legal entity behind each contributor brand, what each company states about pay or scale on its own pages, and the court records on file, including the final approval of the McKinney v. Scale AI contributor settlement on 2026-10-30.

The companies appear alphabetically, and no row is ranked or ordered by referral fees. Human data is one route by which data is acquired for AI training, next to crawling, archive licenses and company records, so it sits within AI training data licensing as the route that runs through paid people rather than existing files.

What is human data for AI?

Human data is data produced by paid people for a model to learn from or be measured against, as opposed to data crawled from the web or licensed from an archive. The method is older than today's companies: the InstructGPT paper of March 2022 describes "about 40 contractors on Upwork and through ScaleAI" who ranked 4 to 9 model responses per prompt. The Llama 2 paper of 2023-07-19 reports 27,540 supervised fine-tuning annotations collected from people.

Human data work falls into five kinds, each sold to labs as a finished dataset or as ongoing work:

• Demonstrations: written answers, code or reasoning that show a model the target behavior.
• Preference rankings: people order or pair model outputs by quality.
• Evaluations and ratings: people score outputs against a rubric.
• Red-teaming attacks: people try to make a model fail or produce harmful output.
• Recorded collection: speech, first-person video or other capture made to a brief.

This hub covers the companies that supply this work. Labs' own in-house teams are outside its scope, because no primary source describes them.

The human data companies on record

The table lists 21 companies that supply human data to AI developers, alphabetically, with the contributor platform each runs, what it states about pay or scale, and the court records on file. Four rules govern how to read it:

• Claim: the company's own figure on its own page on the check date, never an earning.
• No record: none verified on the check date, which does not prove that none exists.
• Not stated: no figure on the pages checked.
• Cases: allegations unless the court record says settled or dismissed.

The table compares the 21 companies by legal entity, contributor brand, stated pay or scale, company data program and court record, as of 2026-11-09; every pay or scale cell is the named company's claim, checked on that date unless another date is given.

How labs acquire human data: managed vendors, expert marketplaces and participant panels

Labs acquire human data in three ways: they contract managed vendors for finished datasets, they draw on expert marketplaces where individuals contract per project, and they recruit through participant panels where the researcher sets the pay. No primary source on this page states which lab buys from which company.

Managed vendors run the workforce themselves and deliver a dataset or a service. iMerit is one: EXL completed its acquisition of the company on 2026-07-31, according to EXL's quarterly filing. TELUS Digital, whose AI Community counts one million contributors across more than 118 countries in its FY2024 10-K, became wholly owned by TELUS Corporation when the take-private closed on 2025-10-31. Sama and CloudFactory also work this way.

Expert marketplaces contract individuals per project through a platform entity that is often not the brand on the website. Outlier contributors contract with Smart Ecosystem, Inc. together with Scale AI, Inc. under terms effective 2026-05-11. The Handshake AI Fellowship accepts only US-based fellows, contracting through Stryder Corp. d/b/a Handshake and Handshake AI Solutions, LLC.

Participant panels leave the rate to the researcher. Prolific sets a minimum participant pay of GBP 6 or USD 8 per hour and recommends GBP 9 or USD 12 (help page dated 2026-07-17), and adds a corporate platform fee of usually 42.8 % on top, checked 2026-11-09.

Kinds of human data work and where each is covered

The child pages split human data by the work itself, by the contributor's practical questions, and by platform, so each question has one page. The table below lists every one of them as of 2026-11-09, with the question each answers; the first five rows cover kinds of work, the next eleven cover contributor practice, and the last four cover single platforms and labeling vendors.

What human data companies state about contributor pay, and why a claim is not an earning

Every pay figure in this hub is the company's own statement on its own page on the check date; no seller- or contributor-confirmed earnings dataset exists. A stated rate is a ceiling or average chosen by the company, and a paid-out total has no audited scope.

The stated rates spread widely by field and by platform. DataAnnotation, the contributor brand of Surge Labs, Inc., states expert rates of up to USD 75 to 125+ per hour and general projects from USD 25 to 50 per hour, checked 2026-11-09. On the same date, Surge AI's own workforce page carried one contractor listing at USD 200 to 400 per hour, and Alignerr stated an average of USD 80 per hour. Handshake states "Earn up to $100/hr" in its page meta for the Fellowship, while its listings run from up to USD 17 to up to USD 150 per hour, checked 2026-11-09. Mindrift, the expert platform of Toloka AI BV, states entry rates of USD 15 to 30 per hour and expert rates of USD 60 to 100+ per hour, paid bi-weekly, checked 2026-11-09.

Several companies state different figures for the same thing on their own pages, kept side by side here because picking one would invent a fact:

• Outlier states USD 500M+ paid to experts on its home page and USD 600M+ paid out on its about page (checked 2026-11-09).
• Alignerr states 100,000+ experts, Labelbox's home page states 2.6M+, and Labelbox's Series E post of 2026-10-14 states more than 4 million.
• Mindrift's stats block reads 20K AI trainers, while its page meta reads 10,000+.
• Alignerr's FAQ says a contributor whose deactivation appeal is rejected is not eligible for payment for completed work, and elsewhere says all Alignerrs are paid for completed work except in extreme fraud.

A claim describes the company, not the contributor. No company in the table publishes how many contributors earned the stated rate, for how many hours, or after which unpaid assessments.

The legal frame around human data work: classification tests, tax forms and EU platform rules

Human data contributors are engaged as independent contractors on the platforms whose terms are cited here, and that classification is what the current lawsuits contest. In California the ABC test of Cal. Lab. Code 2775 et seq. (codified by AB 5, amended by AB 2257, in force since 2020-01-01) presumes a worker is an employee unless the hiring entity proves all three prongs, and it is pleaded in Cavalier and Boggs. Federally, the FLSA contractor rule at 89 FR 1638 has been effective since 2024-03-11; the proposed rule at 91 FR 9932, published 2026-02-27, is not final.

Contractor status shapes tax. For payments made after 2025-12-31, platforms report US contributors paid USD 2,000 or more in a year on Form 1099-NEC (26 U.S.C. 6041(a) as amended by Pub. L. 119-21 s.70433), with inflation adjustment from 2027. Self-employment tax of 15.3 % (12.4 % plus 2.9 %, 26 U.S.C. 1401) applies to net earnings of USD 400 or more, with a 2026 Social Security wage base of USD 184,500. The McKinney settlement allocates contributor payments 100 % to non-wages reported on Form 1099. In the UK, the trading allowance under ITTOIA 2005 Part 6A Chapter 1, available since 2017, exempts up to GBP 1,000 of trading income per tax year, and Self Assessment applies above it.

In the EU, Directive (EU) 2024/2831 on platform work entered into force on 2024-12-01 and must be transposed by member states by 2026-12-02 (Art. 29). It sets a rebuttable presumption of employment (Art. 5), and Recital 19 names microwork and crowdwork platforms. As a directive, it reaches contributors through national transposition laws.

Why the same companies also license company records

Several human data companies also acquire company operating records, because the same labs that buy expert work buy records of real work. The company data partnership programs these firms run license chat, email, tickets, code and CRM history from operating or closed companies. Among the companies in the table, these run such a program, in alphabetical order:

• AfterQuery: Data Partnerships (email, chat, tickets, CRM, finance records).
• Centific: Enterprise Data (office, CRM and ERP records).
• Handshake AI: Data Partnerships (industry operating records).
• Labelbox: Prism (reasoning traces and cross-system work data).
• Mercor: Mercor Data (records from connected business tools).
• micro1: Company Data Partnerships (operating data and agent trajectories).
• Scale AI: Scale Data Partnerships (warehouses, chat, email, documents).
• SuperAnnotate: SuperAnnotate Data Partnerships (threads, runbooks, tickets).
• Toloka: Physical AI Data Partner Program (egocentric video, teleoperation).
• Turing: Project Lazarus (chat, tickets, wikis, code, CRM).

Their stated terms differ more than their contributor pages. micro1 states a commitment of USD 1 billion over 12 months to enterprise data acquisition, announced 2026-10-09; that is a commitment, not money paid. Labelbox states that Prism licenses are exclusive by default, paid as an upfront fee at signing plus a revenue share, checked 2026-11-09. Turing states that Project Lazarus licenses are non-exclusive, paid as a revenue share with nothing upfront. Scale AI states an illustrative value of USD 10K to 1M+ per data partnership, checked 2026-11-09. Mercor states payment on delivery, typically 2 to 4 weeks after a signed agreement, with Mercor covering extraction, anonymization and transfer costs.

Handshake's Data Partnerships page shows two figures at once on 2026-11-09: the headline "Earn $100K to $4M+" and an estimator that reads "Your data could be worth up to $100,000". The program pages set these against each other, and the comparison of program terms compared, dated keeps every such figure with its check date.

Benefits, risks, downsides and alternatives for contributors and for data holders

Contributor work and data licensing through human data companies pay real money to some people and companies, but both transfer rights that are hard to take back and both rest on terms the company can change. The four lists below set out each side for contributors and for data holders.

Benefits

• Contributors: remote, flexible work paid per hour or per task, with stated expert rates well above general rates.
• Contributors: weekly or bi-weekly payment schedules on several platforms, as the companies state them.
• Data holders: a buyer for records that otherwise sit unused, with some programs covering extraction and de-identification costs.

Risks

• Classification disputes: the McKinney v. Scale AI settlement reached final approval on 2026-10-30 (terms in the table above); Boggs v. Stryder Corp. is stayed pending mediation; Cavalier v. Surge Labs pleads misclassification and unpaid training and qualification time as allegations, pending as of 2026-11-09.
• Rights and dispute terms: Outlier's terms effective 2026-05-11 have contributors irrevocably assign all rights in their work product (s.4(b)) and agree to individual AAA arbitration with a class waiver and a 30-day opt-out for US residents (s.10).
• Monitoring and identity checks: Alignerr uses Persona ID verification and Hubstaff time tracking with periodic screenshots, and noyb states that it and three partner groups filed five GDPR complaints against Scale AI over Outlier in France, Italy, the Netherlands and Austria on 2026-10-27.
• Data exposure: Mercor reported to the California Attorney General on 2026-06-25 a breach via compromised LiteLLM versions on 2026-03-24 and 2026-03-30, and the consolidated case is stayed pending mediation; contributor cases sit beside the AI copyright lawsuits over training data in the case tracker.

Downsides

• Contributors: stated rates are ceilings, unpaid assessments come first, and work depends on project availability.
• Data holders: a license to a lab cannot be reversed once a model is trained; Labelbox states that models already trained on the data cannot be unlearned.

Alternatives

• Contributors: not contributing at all, or direct freelance work outside AI training.
• Contributors: participant panels where the researcher sets the pay above a stated minimum.
• Data holders: not licensing at all, which keeps every right and every record in house.
• Data holders: licensing through a program that is non-exclusive, or keeping records for internal use only.

The seller's choice between these routes is a business decision with its own steps. The guide on how to sell or license company data to AI companies sets those out, including the decision not to sell.

How this hub is compiled: neutral order, check dates and no fees

Companies appear alphabetically and whether or not any of them pays referral fees; no row is ordered, filtered or described by pay.

Each pay or scale figure is the named company's claim on the date shown, and each court entry carries its docket and its status as of 2026-11-09. A figure that could not be re-read on its own page on the check date is not stated. The page is information, not legal or tax advice.

The DataPacts Editorial Team checked every figure here against its primary source, following how we source, date and check every figure.

Where human data meets buyers, RL environments and the law

This hub answers who supplies human data and on what stated terms; the pages below answer who else acquires data, what RL environments are, which law applies and how the site is funded.

• Every buyer of AI training data, by category: who buys AI training data.
• RL environments, built partly by the same companies: RL environments and who buys them.
• Copyright, privacy, employee and contract rules: AI training data law by jurisdiction.
• Size, prices, payouts and history: the AI training data market.
• Funding and what the site never does: how DataPacts is funded.
• People licensing their own data rather than doing paid work: individuals selling their own data.

Frequently asked questions

Which AI training platform pays the most?

No single platform pays the most, because each states its own rates by field and changes them. The companies' stated rates, checked 2026-11-09, run from entry rates near USD 15 per hour to single expert listings above USD 150 per hour. The platform comparison in the table above sets them out without ranking.

Are human data companies the same as data labeling companies?

Data labeling companies are one kind of human data company, not the whole group. Labeling vendors annotate data to a client's specification, while other human data companies supply demonstrations, rankings, evaluations and red-teaming. Labeling vendors as buyers of data have their own page, listed in the table above.

Do AI labs hire contributors directly?

The published record shows labs acquiring human data through contractor platforms and human data companies rather than by hiring contributors directly. The InstructGPT paper of March 2022 describes about 40 contractors hired on Upwork and through ScaleAI. No primary source describes the current in-house human data teams of individual labs.

How is human data different from synthetic data?

Human data is made by paid people, while synthetic data is generated by models. Labs use human data to show target behavior, rank outputs and test failures that models cannot yet judge reliably. The page on synthetic data for AI training covers what model-generated data is and its limits.

How are human data contributors paid and taxed in the US?

Most US contributors are paid as independent contractors and receive Form 1099-NEC once paid USD 2,000 or more in a year, for payments after 2025-12-31. Outlier states weekly payment on Tuesdays via PayPal, Airtm or ACH. Self-employment tax of 15.3 % applies to net earnings of USD 400 or more in 2026.

Has a human data company settled a contributor lawsuit?

Yes: the McKinney v. Scale AI class settlement received final approval on 2026-10-30 in the Superior Court of California, County of San Francisco (CGC-24-620481). The gross settlement is USD 12,500,000, and contributor payments are reported as non-wages on Form 1099. A distribution status conference is set for 2027-09-13.