Home โ€บ Blog โ€บ Human data
Inside the Data Layer ยท Part 5 ยท Human data

The Human Layer: Labellers, Experts and the Multibillion-Dollar Business of Teaching AI

Behind every helpful, harmless-sounding answer are people: writing example responses, ranking pairs of outputs, flagging harmful content, designing hard problems and checking a model's work. Human data is a tiny share of training tokens and a large share of what makes a model behave. It has grown into an industry worth tens of billions of dollars on paper, shifting from low-paid crowd workers to PhDs and professionals. This part explains what human data is used for, how it is collected and quality-checked, what it costs, and who does the work.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 27 Oct 2026) Sources are the InstructGPT and Llama 2 papers, TIME's 2023 investigation into Kenyan labellers, and reporting by Reuters, Bloomberg and TechCrunch on Scale AI, Surge AI and Mercor. Company revenues, valuations and pay rates are press reports or estimates (โš ๏ธ) unless stated. The labelling game, agreement and cost calculators are our own illustrations. Disclosure: Claude, made by Anthropic, helps write this site; Anthropic, like other labs, buys human data and does not disclose its vendors or spend.

1. Case file

๐Ÿ‘ฉโ€๐Ÿซ
Human data, October 2026Stage 6 ("enrich") of the data supply chain
Llama 2 (Meta, 2023)1M+ human preference comparisons
Kenyan labellers for OpenAI via Sama~$1.32โ€“$2 per hour (TIME, 2023)
Meta โ†’ Scale AI (Jun 2025)~$14.3B for 49%; Scale cut ~14% of staff a month later
Surge AI (2025)Talks at $15B+, later $25B+ (reported) โš ๏ธ
Mercor (Oct 2025)$350M at a $10B valuation
Expert workPhDs, doctors, lawyers, coders writing and grading tasks

2. Timeline

Human data for AI, 2005 โ†’ 2025

Tap a milestone. Colour: crowdwork era, RLHF era, industry, labour.

Tap a milestone.

3. What humans make

4. Try labelling

The workhorse of post-training is the pairwise preference: show a person two answers to the same prompt and ask which is better. Llama 2 collected over a million of these to train its reward models. Try a few; your choices are compared with a guideline-following "expert" label.

Which answer is better?

Five invented prompts, written to show common judgement calls. Pick A or B.

5. Quality and agreement

Human labels are noisy. People disagree, get tired, misread guidelines or hurry to hit quotas. Vendors measure inter-annotator agreement, seed tasks with known "gold" answers, and route items to several annotators and take a majority vote or send disagreements to an expert. More labels per item cost more but give more reliable data.

How many annotators per item?

Majority-vote accuracy if each annotator is independently right with probability p. Real errors are correlated, so this is an upper bound. Our arithmetic.

6. The workforce

From crowd to expert

Tiers of human-data work. Pay ranges are indicative and vary hugely by country and vendor; only the crowd figure is from a published investigation.

Tap a tier.

The centre of gravity has moved up the pyramid. Early RLHF needed generalists who could judge fluency and safety; reasoning models need people who can check a proof, a legal argument or a pull request. InstructGPT's labels came from a small team of contractors hired through Upwork and Scale AI โš ๏ธ; frontier labs now recruit specialists in dozens of fields, often through platforms such as Mercor, Surge AI, Scale's Outlier and Turing, and sometimes as direct hires.

7. What it costs

Human-data cost calculator

Our own model. Pay rates are inputs; vendors also charge a margin.

For scale: one third-party estimate put Meta's spend on Llama 2's preference data at about $8 million โš ๏ธ, small next to the compute bill. For expert data the economics flip: an hour of a doctor's or senior engineer's time can cost more than the GPUs needed to train on what they write.

8. The industry

The big three, 2025

Reported valuations. Hatched = in talks or press estimate, not a confirmed round.

9. Labour and ethics

In January 2023 TIME reported that OpenAI's contractor Sama paid Kenyan workers roughly $1.32 to $2 an hour to label graphic descriptions of violence and abuse for a toxicity filter, and that some workers described lasting psychological harm; Sama disputed parts of the account, and OpenAI confirmed Sama's workers had contributed to the tool. The story pushed the industry toward wellness programmes, pay floors and more transparency, but the structure remains: contract work, opaque pay, sudden project cancellations and strict confidentiality. At the expert end the issues differ: misclassification as contractors, changing rates, and professionals training systems that may automate parts of their jobs.

10. Humans vs AI feedback

Where humans still leadDefining what "good" means (guidelines, rubrics, constitutions), judging taste and helpfulness, red-teaming for new harms, building hard tasks at the edge of a model's ability, and auditing AI judges.
Where AI feedback took overBulk preference labelling and grading (RLAIF and model judges), synthetic instructions (Part 4), and verifier-checked tasks. Humans increasingly label the hardest few percent and check the machines.

11. Choosing a human-data strategy

Build, buy or skip?

An editorial guide.

VerdictHuman data has become smaller in volume and higher in value. The bulk of labelling is now done by models; what labs buy from people is judgement: expert answers, hard problems, rubrics and checks on the AI judges. That has turned a low-wage crowd business into a market of multibillion-dollar intermediaries recruiting professionals, and made the neutrality and confidentiality of those intermediaries a strategic concern, as the Scale AI episode showed.

Inside the Data Layer

12. Reading the numbers

Primary / confirmed
Llama 2: over 1 million binary human comparisons (paper). Mercor: $350M at $10B, Oct 2025 (confirmed to TechCrunch). Scale layoffs: ~14%, 200 staff and 500 contractors, Jul 2025 (Bloomberg via press). TIME: $1.32โ€“$2/hour via Sama, Jan 2023.
Reported โš ๏ธ
Metaโ€“Scale ~$14.3B (some reports $14.8B). Surge: talks at $15B+ then $25B+; 2024 revenue >$1B. Customers leaving Scale. InstructGPT's ~40 contractors via Upwork and Scale (secondary). Llama 2 preference data ~$8M (third-party estimate). Expert pay ranges.
Our illustrations
Labelling game, majority-vote calculator (independent errors), workforce tiers and pay bands, cost calculator and chooser.