AI Data Labelling Work on the Record: Pay, Lawsuits and a $12.5M Settlement

Scale AI has agreed to pay $12.5 million to settle a California class action brought by contractors who label and write training data for AI models. It is the first priced outcome in a run of AI labelling lawsuits that now reaches Surge AI and Mercor. The filings, settlement papers and reported pay rates put a number on the human layer under frontier models, and the number is smaller than the valuations sitting on top of it.

The short answer

AI labelling work pays anywhere from under $2 to over $85 an hour depending on country, platform and expertise. Scale AI, Surge AI and Mercor all face lawsuits alleging they misclassified data annotation workers as contractors. Only Scale has a priced outcome so far: a $12.5 million settlement, with no admission and final approval still pending.

What AI data labelling work is, and who does it

Data labelling is the human work of tagging, ranking, correcting and writing the examples a model learns from. In 2019 that meant drawing boxes around cars. In 2026 it more often means a lawyer grading a chatbot's legal answer or a physician writing a reference response.

The workforce sits outside the labs. OpenAI, Meta, Google and Anthropic buy this work from vendors, and the vendors source it from contractors. Scale AI runs its contractor pools through platforms called Outlier and Remotasks. Surge AI runs DataAnnotation. Mercor recruits credentialed "experts" and places them on lab projects.

That three-layer chain is the whole story. The lab owns the model, the vendor owns the relationship, and the worker owns a 1099 form. Each layer has a reason to describe the arrangement differently, and the lawsuits are where those descriptions get tested.

This is a large market, not a footnote. Reuters reported in July 2025 that Surge AI earned more than $1 billion in 2024 revenue without outside funding, ahead of Scale AI's roughly $870 million. Meta paid a reported $14.3 billion for 49% of Scale in June 2025. Mercor raised at a $10 billion valuation in October 2025.

Those revenues are mostly labour passed through. The vendors' economics turn on the spread between what a lab pays per hour of human judgement and what the human receives. That spread is exactly what the wage and classification suits put on the table, which is why the ledger below matters to anyone modelling what AI actually costs to produce.

The ledger: every AI labelling lawsuit and complaint on the record

Here is everything we could trace to a filing, a court docket or named reporting. Where we only found aggregator coverage, the item is left out. That is a deliberate trade: a shorter list you can check beats a longer one you cannot.

Status reflects public dockets and reporting we could open. Several of these cases may have moved without coverage. Treat "no outcome found" as a statement about the public record, not about the case.

Read the table by column, not by row. All but one of the US court claims is a classification claim at heart, and every one sits in California or Texas. The international rows are not lawsuits at all. That gap is itself a finding: the workers paid least are the ones with the least access to a court that could hear them.

Scale AI's $12.5 million settlement, line by line

The lead case is McKinney v. Scale AI, Inc., San Francisco Superior Court No. CGC-24-620481. Steve McKinney filed it on 10 December 2024 with Clarkson Law Firm. It alleged that Scale failed California's ABC test, the three-part standard a company must meet to treat a worker as an independent contractor.

The settlement class covers people who worked as contributors for Scale AI or its affiliate Smart Ecosystem, or who were placed by staffing partner HireArt. Members must have lived in California between 10 December 2020 and 28 February 2026. That includes Outlier and Remotasks accounts with a California address.

How the money is divided

The net fund of about $7.54 million is split pro rata by workweeks in the class period. No claim form is needed; checks go to the address on file. A separate $500,000 covers penalties under PAGA, California's Private Attorneys General Act, with $325,000 to the state and $175,000 to workers by pay periods.

No class size is published in the materials we opened, so there is no honest per-person number. A worker with a single week of tasking and a worker with 200 weeks will receive very different checks, and that is by design.

What Scale conceded, which is nothing

The defendants deny every allegation and maintain contributors are independent contractors. A settlement with no admission prices the litigation risk. It does not establish that the business model is unlawful, and Scale can keep running it.

My read is that $12.5 million is cheap. Scale was valued at $13.8 billion in January 2025, months before Meta paid $14.3 billion for 49% of it. A one-off payment below 0.1% of the January valuation, with no change to classification required, is a good result for the company. If I advised a rival platform, I would read it as a ceiling to beat, not a warning.

The cases that did not settle

The settlement does not cover two other 2025 suits. In January 2025, Amber Rogowicz filed a PAGA claim saying her Outlier pay effectively worked out to $15 an hour, below California's $16 minimum at the time. She said she worked about 10 hours a day and was paid for 5, partly because reading instructions and training went unpaid.

Scale said it paid her fairly and removed her for breaching community guidelines. Separately, six Outlier workers filed a federal class action in the Northern District of California on 17 January 2025. They alleged they were made to write prompts about violence and child abuse without adequate psychological support. Scale said workers could opt out of sensitive projects at any time.

Surge AI and Mercor: the newer data annotation lawsuits

The two fastest-growing rivals now face the same core claim, and Mercor faces a wider version of it.

Surge AI: unpaid training time

On 20 May 2025, Clarkson Law Firm filed a class action against Surge Labs in San Francisco Superior Court. The firm's announcement alleges that Surge misclassifies its data annotators and demands off-the-clock work for training and project familiarisation. It names OpenAI and Meta as Surge clients. Counsel described "tens of thousands" of Californians relying on the work; that is the firm's characterisation, not a court finding.

The unpaid onboarding claim is the one to watch. It appears in the Rogowicz case against Scale and the Surge complaint independently. If reading a 40-page task guideline is work, every platform that pays only per approved task has an exposure that scales with how often guidelines change.

Mercor: from pay cut to ERISA

Mercor's dispute began with a pay cut. In November 2025 it cancelled a project reviewing video for Instagram and Facebook, then offered workers a new project at $16 an hour, down from $21, according to Forbes as reported by PEOPLE. Mercor said its role descriptions "clearly state that the work is temporary and project-based."

The litigation followed in 2026. White v. Mercor.io Corp., No. 6:26-cv-00201, was filed in the Northern District of Texas on 8 May 2026. Per Littler's summary of the complaint, it seeks a class of 30,000 experts and alleges denied 401(k) and health benefits, fraudulent 1099 filings, and a mandatory screen-tracking app.

Mercor moved on 9 June 2026 to compel arbitration, strike the class allegations, and stay or dismiss the case. That motion is the real fight. If it succeeds, the class collapses into individual arbitrations, and the economics of the suit change entirely.

A separate class action filed in the Northern District of California on 21 April 2026 names Mercor and unnamed AI lab defendants. It combines contractor treatment with claims about opaque AI screening and a March 2026 security incident that allegedly exposed the equivalent of complete HR files. The filing firm states the complaint contains allegations only.

Naming the labs is the strategic move. Every earlier suit stopped at the vendor. Joint-employer theories that reach the buyer are what would turn a vendor's labour problem into the lab's labour problem, and that is the version the buyers of this work should model.

AI training data workers' pay, from $1.32 to $85 an hour

Pay disclosures for this work are scattered and self-reported. Put the documented figures on one axis and the spread is the first thing you see.

There are now two labour markets using the same words. The first is generalist tasking: ranking outputs, flagging content, simple annotation. It pays near or below local minimum wages and has moved repeatedly toward cheaper geographies. The second is expert data, where a doctor or attorney writes and grades answers. Mercor's claim of an average above $85 an hour belongs here.

Lumping these together produces nonsense in both directions. "AI labellers are exploited" and "AI trainers earn $85 an hour" are both true of somebody. Neither describes the median worker, and no platform publishes that median.

The hourly rate also understates the problem the lawsuits describe. Workers report projects ending without notice, hour caps, and unpaid onboarding. In March 2024, Scale shut Remotasks in Kenya, Nigeria and Pakistan. Rest of World reported that one Nigerian worker had less than a dollar to show for more than 20 hours across five projects.

A $21 rate that can become $16 overnight, or zero, is worth less than its face value. That is the economic substance of the classification argument: employees carry less of the demand risk than contractors do.

The margin between what labs pay and what labellers get

The single most useful number in this whole record is a spread, not a wage. It shows how much of each dollar a lab spends on human data reaches the human.

The Sama figure is take-home pay against the contracted hourly rate; Sama's overheads, benefits and taxes sit in the gap. The Mercor figure comes from a single unnamed source reported by Dealroom and is unaudited.

TIME reported that OpenAI's contracts with Sama set an hourly rate of $12.50, while workers took home $1.32 to $2. Sama employed those workers in Nairobi, with office costs, managers and statutory deductions. So the 84% to 89% gap includes real cost, not just profit. It is still the clearest documented split in the industry, and it is three years old.

Mercor's numbers do not reconcile

Mercor's disclosed figures are newer and stranger. Dealroom reported a $2 billion gross run rate in June 2026 with 60% to 70% paid to contractors. That implies $1.2 billion to $1.4 billion a year in payouts. The same report says Mercor pays "over $2 million a day," which annualises to about $730 million.

Both cannot be precise at once. The run rate may annualise a peak month, or the daily figure may be stale. I would not build a model on either until Mercor publishes audited numbers. What is clear is that Mercor passes far more of each dollar through than the 2022 BPO model did, which is the expert-data market's actual innovation. For the broader point about how lab revenue figures are reported, see the AI revenue leaderboard built on disclosed numbers only.

Where this ledger is weakest

This post argues that the human layer is cheaper, more contested and less disclosed than the valuations suggest. Here is where that argument is thin.

Allegations are not findings

Almost everything in the ledger is a complaint. Complaints are written by plaintiffs' counsel to sound as bad as possible. The only federal agency that looked, the Department of Labor, closed its Fair Labor Standards Act probe of Scale in May 2025 without published findings. Upwork and HireArt were dropped from the inquiry too. A fair reader should weigh that as evidence, even if the reason was never given.

Many workers want contractor status

Flexibility is real value. A physician grading model answers between shifts may not want an employer, a schedule or a timesheet. The 30,000-person Mercor class includes people who chose this work because it is not a job. Reclassification would help some and end the arrangement for others, and the lawsuits speak for the first group only.

The pay data is selected, not sampled

Every pay figure here surfaced because someone complained or a company wanted to boast. There is no survey of labelling pay with a stated sample. The true distribution could sit well above the complaints or well below the boasts, and nothing in the public record settles which.

What it means if you buy training data or labelling services

You are probably not OpenAI. You may still buy evaluation sets, RLHF rounds or annotation from these vendors, or from smaller shops on the same model. Three implications follow, and I would act on all of them.

First, labour cost is your vendor's largest input, and it is under legal pressure in two US states at once. A vendor that loses a classification case reprices. Your multi-year rate card is only as stable as their worker model. This is the same pass-through risk examined in the analysis of AI unit economics and margin traps.

Second, the April 2026 Mercor complaint names AI lab defendants. If joint-employer claims reach the buyer, your own contract terms become evidence. Ask whether your agreement lets you set hours, monitor screens or specify individual workers. Those are the control facts courts examine.

Third, provenance of human data is becoming a diligence question, much like provenance of scraped data. The legal exposure on that side is mapped in the tracker of AI copyright cases over training data. Labour terms will sit beside it in procurement reviews within a year, and your AI procurement checklist should carry both.

Frequently asked questions

Where to start

If you buy human data from any vendor, request one document this week: the vendor's worker agreement template, including its arbitration clause. How a vendor classifies, pays and disputes with its workers is now a material risk to your pricing, and the template tells you more than a sales call will.

Then put 30 October 2026 in your calendar. The final approval hearing in McKinney is the first moment a court signs off on a price for this risk. Whatever number survives that hearing becomes the benchmark every other labelling vendor's lawyers will cite.

For the design side of the same question, the post on human-in-the-loop architecture covers where human review should sit inside your own systems.

Related reading