Skip to content
HumanSourcer

AI-training market data

Last updated

7755 active listings across 33 AI-training networks, rebuilt on every daily deploy. Every rate figure here is computed over 723 distinct advertised rates, not per listing - providers repeat one pay sentence across a whole batch, so counting listings would let a single advertisement vote thousands of times.

For the median rate itself and how it is derived, see the Pay Index.

Share this findingShare this findingShare this finding

How much do advertised rates vary?

The middle half of advertised AI-training rates falls between $22–$100/hr, but the full range runs from $1 to $450 per hour, with 9.8% of distinct advertised rates below $10/hr and 10.2% at $150/hr or above. Measured over 723 distinct advertised rates, not per listing.

$22–$100/hr · n=723

Share this findingXRedditLinkedIn

The rates split by how they were derived. 566 were stated outright, with a median of $65/hr; 157 came from text that hedged (“approximately”, “on average”), with a median of $24/hr. The two differ by more than double, so the cheap end of the distribution is substantially an artefact of how the figure was derived rather than what the market pays.

Why one pay sentence covers a thousand listings

6,440 listings state an hourly rate, but they carry only 723 distinct advertised rates - each one repeated across 8.9 listings on average. Providers post a batch of roles quoting one pay sentence, so counting per listing would let a single sentence vote thousands of times. Every rate figure on this page is computed over the distinct rates instead.

8.9× · n=723

Share this findingXRedditLinkedIn

Those rates are also unusually round: 71.0% are exact multiples of $5. These are advertised prices rather than negotiated ones, and advertised prices cluster.

How concentrated is this market?

Across 33 platforms with live listings, the largest single source accounts for 41.4% of indexed listings and the top three for 63.1%, giving a Herfindahl–Hirschman Index of 0.207 - moderately concentrated. This describes the listings HumanSourcer indexes, not the AI-training market itself: a platform that publishes no scrapeable board is absent from it entirely.

0.207 · n=33

Share this findingXRedditLinkedIn

How is pay actually quoted?

96.0% of listings that disclose pay quote an hourly rate. The other 270 quote per task, per unit, per project or as a salary, and are excluded from every hourly figure on this site - there is no honest way to convert a per-task price into an hourly one without knowing how long the task takes.

96.0% · n=6,740

Share this findingXRedditLinkedIn

How often is pay disclosed?

86.9% of active listings state some pay signal - but that is listing-weighted, and one large source posting thousands of near-identical roles moves it (Prolific Participants alone is the largest). Counted per platform instead, 25 of 33 (75.8%) publish pay on at least one open role.

86.9% · n=7,755

Share this findingXRedditLinkedIn

Where do listings say applicants must be?

Of the 3,161 listings whose title names a country (40.8% of 7,755), the most frequently named are USA (36.0%), United Kingdom (21.5%), Canada (6.0%). The remaining 4,594 state no country in their title, which is reported as unknown rather than assumed to be worldwide.

USA · n=3,161

Share this findingXRedditLinkedIn

How long do listings stay open?

Of 866 listings first indexed between 2026-10-02 and 2026-10-09, 18% had closed within 7 days. Only listings whose full 7-day window has elapsed are counted, so a listing posted yesterday cannot drag the figure down; the earlier bound is where closures stop being visible in the published data, so the cohort is a window rather than everything older.

18% · n=866

Share this findingXRedditLinkedIn

Which platforms advertise what

Platforms with at least 50 open roles. A median is shown only where the platform advertises at least 5 distinct rates - below that a single listing would set the figure, so the row shows the advertised range instead. An em-dash means the platform states no hourly rate.

PlatformOpen rolesMedianMiddle halfDistinct ratesPay disclosed
Mercor Experts105$100/hr$75–$12554100%
micro1 Experts Opportunities257$95/hr$50–$13010197%
AfterQuery Experts220$93/hr$64–$12243100%
Handshake AI221$78/hr$50–$11132100%
Alignerr549$75/hr$50–$11066100%
Sepal Expert Hub88$75/hr$60–$937698%
Turing AI Advancement Work267$65/hr$19–$113814%
Rise Data Labs116$48/hr$44–$831276%
DataForce jobs139$40/hr$30–$50955%
Mindrift850$31/hr$22–$4519100%
Prolific Participants3208$30/hr$20–$805793%
Meridial833$30/hr$13–$655999%
LXT AI59$24/hr$23–$271551%
TELUS International AI Community portal64$17/hr$7–$262092%
Welo Data392$13/hr$8–$2310446%
CrowdGen60$4/hr advertised, too few distinct rates for a median153%
Appen60$4–$40/hr advertised, too few distinct rates for a median380%
DataForce Community50——00%
Share this tableXRedditLinkedIn

How many listings go up each day

Listings per day since 1 Aug, by use case and by domain. 1,090 new listings were first indexed in the 7 days to 15 Oct, up 29% on the 7 days before. The window opens on 1 Aug because the scraper fleet was still being built through 31 July, and the days before it measure that rollout rather than demand.

7,719 listings open on 15 Oct

Open listings per day by use case

Open is the number of listings still up at the end of each day. Past days are an upper bound: a listing that disappears and returns reads as continuously open.

01,0002,0003,00017 Aug29 Aug10 Sep21 Sep3 Oct15 Oct
AI model evaluationCoding / software engineeringResearch / analysisTranslation / linguisticsData annotation / labelingAll other use cases
Show data table
Open listings per day by use case, one row per day
DayAI model evaluationCoding / software engineeringResearch / analysisTranslation / linguisticsData annotation / labelingAll other use cases
2026-10-152,4686881,5287144831,838
2026-10-142,4506951,1446844921,856
2026-10-132,4557099346855031,869
2026-10-122,4686679317005051,874
2026-10-112,4606699397105051,869
2026-10-102,4576719397125061,868
2026-10-092,4296739427235371,875
2026-10-082,4346649437765421,876
2026-10-072,4977009428375461,970
2026-10-062,4916759448375291,964
2026-10-052,4916509468415311,967
2026-10-042,4986559488405322,037
2026-10-032,4685518847754791,834
2026-10-022,4795598897804851,854
2026-10-012,5045608907824891,862
2026-09-302,4275488827604821,827
2026-09-292,4305458847614971,820
2026-09-282,4305358857554961,816
2026-09-272,4305368857554961,818
2026-09-262,4315348867574951,826
2026-09-252,4315348787544941,804
2026-09-242,4395358827585361,802
2026-09-232,4355368857625351,793
2026-09-222,4305308907645351,773
2026-09-212,4425208847645491,779
2026-09-202,4435198847655511,779
2026-09-192,4445218857675521,779
2026-09-182,4965258887755451,783
2026-09-172,5075228747865571,778
2026-09-162,5055218617415551,733
2026-09-152,5125458627325511,727
2026-09-142,5045388527335461,695
2026-09-132,5045388527125471,692
2026-09-122,5055398527125501,693
2026-09-112,5115498507065501,693
2026-09-102,5985678687085711,734
2026-09-092,5815538547085591,715
2026-09-082,5985448537115631,723
2026-09-072,5995548577245621,730
2026-09-062,5995548587255631,729
2026-09-052,6085598607275651,735
2026-09-042,5355288406525571,688
2026-09-032,5115308356525541,662
2026-09-022,4995228196525861,659
2026-09-012,5245298156575801,628
2026-08-312,5285488186855801,640
2026-08-302,5565738196865841,639
2026-08-292,5585748206885901,639
2026-08-282,5575758096885881,630
2026-08-272,5256318166905961,630
2026-08-262,5016518036956061,612
2026-08-252,4656438036906061,606
2026-08-242,4676408036906031,605
2026-08-232,4746418046896031,604
2026-08-222,4756368046896041,606
2026-08-212,4736608036895981,596
2026-08-202,4686578036895941,600
2026-08-192,4696537856905911,594
2026-08-182,4226527836765701,543
2026-08-172,4146337806695721,544

Open listings per day by domain

Domain comes from the job title or, where the title names none, from a sentence in the listing that a model had to quote and that shows the work is in that domain. A listing can name several, so these lines overlap and do not add up to the total. 36% of listings still name no domain and appear on none of them. Open is the number of listings still up at the end of each day. Past days are an upper bound: a listing that disappears and returns reads as continuously open.

05001,0001,5002,00017 Aug29 Aug10 Sep21 Sep3 Oct15 Oct
CodingVoice / SpeechScience / STEMLinguisticsHealthcareAll other domains
Show data table
Open listings per day by domain, one row per day
DayCodingVoice / SpeechScience / STEMLinguisticsHealthcareAll other domains
2026-10-155331,0191,1751,466532971
2026-10-145321,0381,1801,450531969
2026-10-135461,0421,1971,438539999
2026-10-125471,0401,1531,4555401,001
2026-10-115491,0331,1561,4595431,006
2026-10-105511,0331,1571,4605401,005
2026-10-095861,0361,1581,4645141,004
2026-10-085851,0401,1541,4655131,015
2026-10-076311,0401,1841,5835211,045
2026-10-066301,0221,1571,5875211,044
2026-10-056301,0171,1411,5915011,049
2026-10-046331,0881,1411,5905051,070
2026-10-035309541,0531,552484944
2026-10-025329661,0551,557486950
2026-10-015359711,0601,559486960
2026-09-305229701,0431,537473920
2026-09-295189541,0381,540468932
2026-09-285099531,0381,533465931
2026-09-275109561,0371,533465931
2026-09-265109581,0321,534466937
2026-09-255119371,0201,535466937
2026-09-245129371,0221,536467935
2026-09-235149271,0221,538468928
2026-09-225109001,0201,532469930
2026-09-215079061,0121,533472937
2026-09-205059121,0141,534472939
2026-09-195079121,0151,536471941
2026-09-185119011,0141,536473959
2026-09-175109021,0131,536474968
2026-09-165098861,0031,482470964
2026-09-155328791,0061,479472964
2026-09-145318699931,474469950
2026-09-135318669931,453469950
2026-09-125338679931,453469952
2026-09-115448679941,448470948
2026-09-105588691,0301,449477990
2026-09-095478611,0201,440472988
2026-09-085378601,0231,4414731,005
2026-09-075478641,0251,4464731,006
2026-09-065478641,0261,4464731,005
2026-09-055518641,0311,4464751,009
2026-09-045338519801,378464913
2026-09-035357989811,336458910
2026-09-025287919741,323462877
2026-09-015337379751,324463858
2026-08-315537359961,355466863
2026-08-305797339931,353466894
2026-08-295797339911,353467897
2026-08-285797329881,351467889
2026-08-276387409731,346466875
2026-08-266817349691,343448856
2026-08-256747249681,342448855
2026-08-246737239661,342449859
2026-08-236737239661,342449859
2026-08-226677239621,342449863
2026-08-216917289611,342447857
2026-08-206907269561,341440862
2026-08-196857079561,339442858
2026-08-186816609541,328442848
2026-08-176646609301,324440845

Explore the data

Filter by domain, use case, provider or date. Every figure recomputes against the filtered set, and the filter state is kept in the URL so a view can be shared.

Advertised rate distribution

Each bar counts distinct advertised rates, not listings - providers repeat one pay sentence across a whole batch, so counting listings would let a single sentence vote thousands of times. Grey shows rates whose source text hedged ("approximately $20/hr"); those run materially lower than stated rates, so the left tail is partly an artefact of how the figure was derived.

No listings match these filters.
Show data table
Advertised hourly rate distribution, counted over distinct advertised rates
Hourly rateDistinct ratesStatedHedged
$0–$10000
$10–$20000
$20–$30000
$30–$40000
$40–$60000
$60–$80000
$80–$100000
$100–$150000
$150–$200000
$200+000

Median and middle half by group

Median advertised rate with the middle half (25th–75th percentile) as the bar. Computed over distinct advertised rates; a group with fewer than 5 of them shows an em-dash rather than a figure a single listing could set. Click a row to filter.

No listings match these filters.

Pay bands by provider

Share of each provider's listings by published hourly rate, top 10 providers by volume. "Unspecified" means no hourly rate could be determined from the listing. Click a row to filter by that provider.

No listings match these filters.

Demand heatmap

Darker = more listings for that use case × provider combination. Click a cell to filter by both.

No listings match these filters.
Listing counts by use case and provider. Empty cells mean no listings for that combination under the current filters.

How these figures are derived

Listings are scraped daily from the public boards of 33 AI-training platforms. Pay is read from a structured field where one exists, parsed by rule where it does not, and otherwise extracted from the listing prose by a model that must quote the sentence it took the figure from.

Rates are deduplicated before any percentile is computed. The key is the provider plus the verbatim pay sentence, so a batch of listings sharing one advertisement counts once. This is the single most important thing to know about the numbers above: the per-listing figures are materially different and would overstate the sample by roughly seven times.

Only rates quoted per hour are comparable, so per-task, per-unit and salaried listings are excluded from every rate figure and counted separately under how pay is quoted.

No rate trend is published. The index begins at 2026-07-19, so there is not yet enough history for a pay trend that would mean anything, and pay figures are re-derived whenever the extractor improves. Listing volume is different: first-seen and last-seen dates are never rewritten, so the listings per day series is reconstructed from them for every listing including closed ones, from 2026-08-01. The closure figure above uses a 7-day horizon and counts only listings whose full window has elapsed.

Domains come from two sources, and the second is gated. A domain is taken from the job title where the title names one. Where it does not, a model reads the listing and may assign a domain only by quoting a sentence from the listing that describes the work itself, not a required background or a tool. It is told to return nothing when no domain fits, so a share of listings still has none and is left out of the domain views rather than guessed at. The domain pages under /roles/ still use titles only.

Concentration describes the listings indexed here, not the AI-training market as a whole - a platform that publishes no scrapeable board does not appear at all.

The aggregated figures on this page are also available as JSON. Reuse terms are the same as the Pay Index.

Common questions

How concentrated is the AI-training listing market?
Across 33 platforms with live listings, the largest single source accounts for 41.4% of indexed listings and the top three for 63.1%, giving a Herfindahl–Hirschman Index of 0.207 - moderately concentrated. This describes the listings HumanSourcer indexes, not the AI-training market itself: a platform that publishes no scrapeable board is absent from it entirely.
Why do so many AI-training listings quote the same pay rate?
6,440 listings state an hourly rate, but they carry only 723 distinct advertised rates - each one repeated across 8.9 listings on average. Providers post a batch of roles quoting one pay sentence, so counting per listing would let a single sentence vote thousands of times. Every rate figure on this page is computed over the distinct rates instead.
Do all AI-training jobs pay by the hour?
96.0% of listings that disclose pay quote an hourly rate. The other 270 quote per task, per unit, per project or as a salary, and are excluded from every hourly figure on this site - there is no honest way to convert a per-task price into an hourly one without knowing how long the task takes.
How often do AI-training platforms publish pay at all?
86.9% of active listings state some pay signal - but that is listing-weighted, and one large source posting thousands of near-identical roles moves it (Prolific Participants alone is the largest). Counted per platform instead, 25 of 33 (75.8%) publish pay on at least one open role.
How many new AI-training listings appear each day?
1,090 new listings were first indexed in the 7 days to 15 Oct, up 29% on the 7 days before. "New" means first seen by HumanSourcer, not the date the provider posted it, and the daily series starts on 2026-08-01.
Are advertised AI-training rates round numbers?
Yes - 71.0% of the 723 distinct advertised rates are exact multiples of $5. Rates are advertised, not negotiated, so they cluster on round figures rather than spreading evenly.