The state of AI training work, late 2026

The business of paying people to train AI has been reshaped over the past eighteen months. Rival labs stepped back from Scale AI after Meta's investment in June 2025, two expert marketplaces grew into companies valued in the billions, and the work began shifting from ranking chatbot answers toward building tasks for AI agents. This report, current to late October 2026, covers who buys, who sells, how the work is changing, what has gone wrong and what it means for contributors in Africa.

A short list of buyers#

Demand comes mostly from a few buyers: the frontier AI labs and the largest technology companies. According to the privacy group noyb, Outlier's customers include Meta, OpenAI, Google and the US government. Forbes's sources said two frontier labs were among the investors in the micro1 round it reported in September 2026.

A short list of buyers makes for a volatile market. In June 2025, Meta agreed to invest about $14.3 billion in Scale AI for a 49 percent stake, and Scale's founder, Alexandr Wang, left to work at Meta. Within days, Reuters reported that Google, which had planned to pay Scale about $200 million that year, was preparing to end the relationship. OpenAI said it had already been winding down its work with Scale in favor of more specialized data providers.

The vendors, eighteen months on#

Scale AI has spent the period rebuilding. In July 2025 it cut about 200 full-time staff and ended work with about 500 contractors; interim chief executive Jason Droege said it had ramped up its generative-AI capacity too quickly and that hiring would focus on its enterprise and public-sector businesses. Francis deSouza, most recently chief operating officer of Google Cloud, became chief executive on August 10, 2026, and Wang chairs the board.

Surge AI has never announced outside funding and publishes no financial statements; court filings name it as the company behind DataAnnotation. Talks to raise about $1 billion were reported in mid-2025; we found no report of a closed round by late October 2026.

Mercor grew quickly. TechCrunch put its annualized revenue near $450 million in September 2025; by mid-2026 it was paying more than $2 million a day to over 30,000 contractors active each week, according to Dealroom. Bloomberg reported early talks at about a $20 billion valuation in July 2026; we found no report of a closed round.

micro1, which began as an AI recruiting tool, went from a $500 million valuation in September 2025 to a reported $4 billion a year later. Forbes's report of that round cited unnamed sources; micro1 has since said it priced the round at $4 billion and that its run rate passed $1 billion, without saying whether that is gross or net.

Others. Handshake, the US campus-careers network, began paying students and graduates to test AI models in mid-2025 (see Handshake AI). Labelbox (through Alignerr), Toloka (through Mindrift), Turing and Invisible also supply expert data to AI labs. We could not confirm their finances.

From generalists to experts#

The clearest change is in who gets paid. As models write fluent answers on their own, the useful feedback comes from people who can see what a fluent answer gets wrong: clinicians, lawyers, accountants, engineers and scientists.

Published rates show the gap. micro1's own guide, dated May 2026, describes three bands: roughly $20 to $60 an hour for generalist feedback, $50 to $100 for domain review in fields such as medicine, law and code, and $100 to $200 or more for frontier experts.

Generalist work has not disappeared, but it is crowded and exposed to price cuts. In late 2025, Business Insider reported that Mercor had ended Musen, a Meta project that involved more than 5,000 people at its peak, and offered contractors similar work on a new project at $16 an hour, $5 less than before. Mercor called the report inaccurate without saying which details it disputed.

The practical reading: lead with the credential that is hardest to find, not the skill that is easiest to test. Where to apply first matches profiles to platforms.

RL environments: from grading answers to building tasks#

The second change is in what contributors make. In reinforcement learning (RL), a model attempts a task many times and learns from a score. That requires environments: a realistic task, a sandbox to attempt it in and an automatic check of whether the attempt worked. Code with tests is the classic case; coding for AI labs shows how such tasks are built.

The vendors are moving toward this work. Surge's work includes RL environments as well as human feedback. In July 2026 Mercor announced a deal to buy Deeptune, a startup focused on training AI agents, and in October said it was building a research team to measure models on real work and post-train them where they fail. In September, Fortune reported that micro1 had offered $12.5 million for records from the bankrupt Spirit Airlines, topping Google's $10 million bid; unions and privacy advocates objected. In October, micro1 said it would spend $1 billion over 12 months, with financing from Citi and Hercules Capital, licensing operational data from companies to build RL environments.

For contributors, the job shifts from judging an answer to designing a problem: a task a professional would recognize, an instruction with no hidden requirements and a check that cannot be gamed. The adjacent skill is writing criteria that two graders apply the same way; see rubrics that hold up.

Lawsuits, breaches and surveillance#

Classification. Since December 2024, California contributors have sued Scale, arguing they were misclassified as independent contractors. Scale, an affiliate and the staffing firm HireArt have agreed to pay $12.5 million to settle a group of those suits for people who lived in California between December 2020 and February 2026, while denying wrongdoing; settlement trackers list a final approval hearing for October 30, 2026. The US Department of Labor dropped an investigation into Scale's pay practices in May 2025, TechCrunch reported. Later that month, Clarkson Law Firm, which filed one of the Scale suits, brought a similar proposed class action against Surge Labs, DataAnnotation's operator, alleging unpaid training time and task time limits that docked pay. We found no ruling or settlement there by October 2026.

Security. In June 2025, Business Insider found that Scale had left project documents open to anyone with the link, some listing contractors' names, private email addresses and performance notes; Scale said it had disabled public sharing. Surge was reported to have left internal documents publicly accessible the same year. The most serious incident came in March 2026, when Mercor was hit through a supply-chain attack involving LiteLLM, an open-source software project. Contractors' lawsuits allege that recorded interviews, facial biometric data, screenshots of their computers and background-check files were exposed. Mercor denies wrongdoing and published an update on the incident on June 25, 2026.

Surveillance. On October 27, 2026, noyb filed five complaints with data-protection regulators in Austria, France, Italy and the Netherlands, alleging that Outlier runs third-party trackers during work sessions and makes workers repeat biometric identity checks run by the verification company Persona. Scale disputes the claims.

Contributors hand over more of themselves than most freelancers, from identity documents to face scans and screen activity, and the rules for that data and that labor are being tested in California courts and European regulators' offices.

What it means if you work from Africa#

Access is set by lists that change. Remotasks, Scale's earlier platform, stopped operating in Kenya on March 8, 2024; Rest of World reported that it shut down in Nigeria and Pakistan around the same time. As of October 2026:

• micro1 did not list Nigeria, Kenya, Ghana, South Africa or Egypt among the countries it cannot hire from.
• Mercor listed Nigeria, Ghana, South Africa and Egypt as supported; Kenya was on its unsupported list, though one Mercor page suggested otherwise.
• Outlier listed Egypt and South Africa; Nigeria, Kenya and Ghana were reportedly absent.
• DataAnnotation reportedly accepted applicants from six countries, none in Africa.
• Handshake AI limited its main fellowship to people based in the US.

Expertise travels. The shift toward experts favors a tax lawyer in Lagos, a clinician in Nairobi or an engineer in Accra whose knowledge is scarce wherever it sits. The limit is residence: many expert listings are open only to residents of named countries, often the US, UK or Canada, and Outlier, among others, says pay varies with location. Read the eligibility line before the pay line.

Real workflows are an asset. RL environments need realistic tasks, and realistic tasks come from people who have run the process: reconciling accounts, triaging patients, scheduling freight. That experience counts wherever it was earned.

Concentration is the risk. Scale lost major clients after Meta bought in, Mercor's work for Meta was paused after a breach, and Remotasks left whole countries; none of it turned on any one contributor's work. Keep two or three platforms active, keep income that depends on none of them, and treat ASV's own paid projects as one more line, not the only one.

Recourse is uneven. The Scale settlement covers people who lived in California, and noyb's complaints concern workers in Europe. Elsewhere, contributors rely mostly on their contracts, their records and local law.

Of the platforms covered here, micro1 and Mercor were the most open to applicants in Nigeria, Ghana, South Africa and Egypt; for Kenya, micro1 was the clearer option. Current pay ranges, payouts and country notes are on each platform profile, and where we have a referral link, Apply buttons on ASV use it, as our disclosure page explains.

Sources#

• After data breach, $10B-valued startup Mercor is having a month (TechCrunch)
• Scale AI taps former Google Cloud executive as new CEO (Axios)
• Mercor doubles to $2B gross revenue run rate as AI labs buy expert data (Dealroom)
• This 25-year-old raised US$100 million for his AI startup at a US$4 billion valuation (Forbes Australia)
• micro1 commits $1bn to company data for AI agents, financed by Citi (The Next Web)
• An AI startup powering Meta and OpenAI cut thousands of workers, then offered them a similar project for less money (Business Insider, via AOL)
• AI training platform Outlier illegally surveils European platform workers (noyb)
• Online gig site Remotasks exits Kenya (Business Daily Africa)