Data Annotation vs. Expert Evaluation: Which Tier Is Your Work In?
2026-09-17 · Expert Match AI team
Search for "data annotation jobs" and you will find $20/hr transcription work. Search for "AI training jobs" and you will find $400/hr attorney queues. Both searches lead to the same platforms, and both jobs are described with the same vocabulary - "help train the next generation of AI," "shape how models learn." That shared vocabulary is the problem: it lets a $10/hr audio-transcription listing borrow the prestige of a $250/hr physician pool, and it leaves applicants unable to tell, from the listing alone, which market they are applying into.
There are two markets. This post draws the line between them using the live board.
The split, in numbers
Of the 222 listings on our board today with a disclosed hourly rate, here is where the midpoints land:
| Band (listed midpoint) | Roles | Share | Who lists here |
|---|---|---|---|
| Under $20/hr | 18 | 8% | micro1, Alignerr |
| $20–35/hr | 28 | 13% | Alignerr, micro1, Outlier, DataAnnotation, Braintrust, Stellar |
| $35–50/hr | 17 | 8% | micro1, Braintrust, Outlier, Mindrift |
| $50–100/hr | 77 | 35% | micro1, Mercor, Braintrust, DataAnnotation, Stellar, Mindrift |
| $100–200/hr | 59 | 27% | micro1, Mercor, Braintrust, Stellar |
| $200/hr and up | 23 | 10% | Surge, micro1, Mercor, Braintrust |
Read the middle column and the two markets fall out. Roughly a fifth of the board (46 roles) sits at or below $35/hr: that is the collection and annotation tier. About 72% (159 roles) sits at $50/hr and above: that is the expert evaluation tier. The $35–50 band between them is thin - 17 roles - because there is not much work that is "somewhat expert." Either the task needs your credential or it does not, and the pay steps rather than slopes.
The full salary report breaks this down by field and platform every six hours.
What the two tiers actually are
Collection and annotation: you are the sensor. The task is to produce or label raw material - record 360° video of yourself walking through a city, transcribe forty hours of Brazilian Portuguese audio, tag objects in frames, rate two chatbot replies on a five-point scale. The listing needs a human because a human has a camera, a voice or a native language, not because of anything they know. Today's examples: Alignerr's audio transcription wave in French, German, Hindi, Italian, Spanish and Chinese at $10–35/hr; micro1's video recorder roles at $14–20; DataAnnotation's Writing & Rating generalist queue at $20–30; Outlier's Expert Writing & Editing at $15–35 despite the word "expert" in the title.
Characteristics: no credential required, fast onboarding, per-task pay in practice, high volume, queues that empty abruptly when the batch closes, and heavy competition because the qualification is being a person.
Expert evaluation: you are the judge. The task is to apply professional judgment a model does not have - grade a model's contract clause against what a BigLaw associate would draft, write a physics problem the model cannot solve, decide whether a clinical answer is dangerously wrong, review agent code for the failure a senior engineer would catch. Today's examples: Surge's physician tier at $250–450 and attorney tier at $500–1,000; micro1's BigLaw attorney roles at $140–400 and its PhD & Academic Expert pool at $245–280; Mercor's law experts at $110–150; Braintrust's Staff ML Engineer at $180–260 and RLHF Healthcare Expert at $100–180.
Characteristics: a credential or verifiable seniority is the entry ticket, onboarding is slow and selective, pay is usually clocked hourly, and the pools are standing rather than per-batch - the queues that never quite close.
How to tell which tier a listing is in
Titles lie in both directions - "Expert" appears on $15/hr listings and "Contributor" on $140/hr ones. Read for these instead:
- What does the listing require you to have? Equipment, a language, a phone, availability: annotation tier. A licence, a degree, a bar admission, years at a named class of employer: expert tier.
- What is the verb? Record, transcribe, label, collect, rate: annotation. Evaluate, author, review, grade, red-team, write problems: expert.
- Is the rate a range or a point? A tight point rate ($20/hr, $14–15/hr) is almost always per-task work quoted as hourly. A wide range ($140–400) means tiers inside the pool set by credential.
- Who is the customer? "A customer's project" collecting footage or audio is a dataset commission. "Frontier lab," "model evaluation," "reasoning benchmarks" is evaluation.
- Where does the platform sit on the board? Alignerr's 15 priced roles average $22/hr; Surge's four average $406. A platform's average tells you which market it is mostly in. micro1 is the exception - it runs both tiers on one board, which is why its $92/hr average describes nothing and why we broke it into three boards in the micro1 analysis.
Why the distinction matters before you apply
Because your effective hourly rate depends on it. Annotation-tier listings lose a large share of their listed rate to per-task timing and empty queues; expert-tier listings lose most of theirs to the months of waiting on the way in. Those are different bets and they suit different people. Someone with no credential and free evenings should take the annotation tier for what it is - a real, modest income with a fast start, as the first-90-days plan lays out - and not expect it to become the expert tier through effort. Someone with a licence should skip the annotation tier entirely: every hour there is an hour not spent getting into a standing pool that pays five to ten times as much.
Moving between tiers
The path from annotation to expert evaluation runs through a credential, not through volume. Ten thousand labelled frames do not make you a physician. But there are two real on-ramps in the middle band: language evaluation (Cantonese Language Evaluator at $30–40 on micro1 sits a tier above Cantonese transcription at $20–35, for the same speaker) and domain rating (DataAnnotation's Medical & Health projects at $40–75 accept clinical background short of a licence). If you have a specialty without a formal credential, look for the listing where your specialty is the requirement rather than the equipment.
Browse the board sorted by pay and the two markets are visible in a single scroll: the top third and the bottom third are different industries that happen to share a word. Uploading a resume on the homepage tells you which one you are actually qualified for today.
Published by the Expert Match AI team. Band counts are from our tracker as of September 15, 2026 (222 priced roles of 316 tracked, 10 platforms), using the midpoint of each listing's range. All figures are listed rates, not accepted offers. Some outbound application links on this site carry disclosed referral codes; rankings and recommendations are never influenced by referral payout.