Outlier Cut AI Data Contractor Pay by $5 an Hour

01What the Nova contract changed

Contractors opened the new contract and found the rate cut. The platform's explanation email said the lower rate came with "more stable multimedia task volume and a higher cap on working hours" — the rate went down, and workers were invited to make the difference back in extra hours.

Labor complaints around Outlier are not new. In September 2024, Fast Company published an investigation into contractors whose pay was delayed or randomly deducted. What is new is that the discount is now written into the contract itself.

The contrast with two years earlier is sharp. When large models first took off, people were earning more than $50,000 in six months from part-time labeling work. Outlier's own site claimed it had paid experts over $500 million and counted more than 700,000 registered experts.

02Why basic labeling demand collapsed

Early large models needed oceans of basic text and image data just to learn to produce coherent language. Reasoning models ended that era. With OpenAI's o-series and DeepSeek-R1, low-skill annotation is something models increasingly do themselves, or replace with synthetic data generated in bulk.

Demand did not disappear; it split. The intermediaries and production-line annotators once called "the humans behind artificial intelligence" are the group left behind. What buyers now want is expert knowledge engineering: complex logic problems and long-horizon coding tasks that can push large models to their limits. The premium pays for what a specialist knows, not for headcount.

03The same job title, a 20x pay gap

China's salary data shows the split in numbers. The 2025 AI Industry Talent Development Report puts the average advertised monthly salary for data labeling and AI trainer roles at about 6,479 yuan, and about 9,063 yuan in Beijing, the highest of any city.

The averages hide two different jobs. According to 36Kr, ordinary labelers start at 2,000-4,000 yuan a month — piece-rate work with room and board. A genuine "AI trainer" title starts above 10,000 yuan. Expert roles at top companies pay around 40,000 yuan a month. The two ends of the same job name differ by a factor of twenty.

For individual workers doing low-skill side gigs, the squeeze comes from three sides at once: synthetic data, model self-training, and expert networks. The durable options are real expertise or leaving the field.

04What data buyers should screen for

For model developers and data buyers, the old vendor metric — how many thousand part-time annotators a supplier can field — now measures nothing. The metric that matters is expert density: how many people on the vendor's side can write tasks that actually challenge frontier models.

DataMining is an example of the expert-first model. It was founded by a PhD team from the Software Institute of Peking University's School of Computer Science and has served ByteDance, Alibaba, and Tencent. It does not sell headcount; it builds original high-difficulty tasks targeting benchmarks such as Deep SWE and HLE.

What expert-built data does to a model: running GPT-5.6-sol on DataMining's data yields accuracy (ACC) 15 percentage points lower than on comparable vendors' data, and the model's reasoning traces stretch to about 20 rounds of interaction. That is the point of the purchase — data that pushes the model to its limit, not data that flatters it.

The contractors staring at Outlier's "more stable task volume" email are, in effect, settling the bill for the previous generation of models finishing school: the low-skill data they produced helped train the models that no longer need as much of it.

05Sources

• https://www.cnaiplus.com/a/news/8883321.html
• https://outlier.ai
• https://www.mercor.com