Key findings

1
Within the same kind of work, roles listed less than 14 days earlier drew 6.6 times the apply clicks per open role of roles listed before that, the strongest association measured (11 kinds of work compared).
Figure 1
2
Roles with 'freelance' in the title drew 54% fewer apply clicks per open role than roles without it in the same kind of work, the weakest association measured.
Figure 1
3
Pooled across kinds of work, roles whose title names a profession or discipline drew 16% fewer apply clicks per open role than roles whose title names none, but within the same kind of work 40% more clicks per open role: they are concentrated in kinds of work that draw fewer applicants per role.
Figure 1
4
Roles in the top fifth of their kind of work's advertised rates drew 1.32 times the clicks per open role of their kinds' average, against 0.91 for the bottom fifth and 0.82 for roles stating no rate.
Figure 2
5
In their first week, roles drew 4.66 times the clicks per open role of their kind of work's average; listed more than three months earlier, 0.48 times, a gap that partly reflects the site's own recency ranking.
Figure 3
6
Roles stating an hourly rate drew more apply clicks per open role than roles stating none in 7 of the 9 kinds of work compared; among the kinds with enough clicks and roles on both sides to read alone, the ratio ran from 1.26x in Law, Medicine & Finance to 0.90x in Data Annotation & Labeling.
Figure 5
7
Roles open worldwide drew 1.23 times the clicks per open role of their kind of work's average, and roles open only to US residents 0.78 times.
Figure 4

Figure 1

What goes with more applicants, compared within each kind of work

Which traits of an AI training role draw more apply clicks per open role than similar roles without them?

Trait	Within kind (with / without)	All kinds pooled	Roles with the trait	Kinds of work compared	Compared with
Listed less than 14 days earlier	6.56x	5.81x	1,219	11	Roles listed before that
Open worldwide	2.14x	1.96x	2,067	11	Roles open only to listed countries
States an hourly rate	1.44x	1.25x	3,303	9	Roles that state none
Title names an expert field	1.40x	0.84x	1,941	9	Roles whose title names none
Pay in the top third of its kind	1.22x	1.16x	1,103	10	Roles in the bottom third
Title of 9 words or more	1.06x	1.04x	236	6	Roles with 4 words or fewer
Title names a language	0.90x	0.69x	1,678	8	Roles whose title names none
Open only to US residents	0.61x	0.59x	500	7	Roles open worldwide
Title says freelance	0.46x	0.54x	571	5	Roles without it

Sample. 2,261 apply clicks on 4,296 roles listed at some point from 1 Sep 2026 to 30 Sep 2026, over 10,910 open role-weeks.

Rows are marked, and not ranked in the findings, when either side has fewer than 30 clicks, fewer than 10 roles or fewer than 3 roles open on an average day; the same rule marks the rows of every figure in this report.

Within kind: in each kind of work with at least 10 roles and 30 open role-weeks on both sides, apply clicks per open role a week with the trait over the same rate without it; then the average across those kinds, weighted by the open role-weeks with the trait. 1.00x is no difference; a trait needs two such kinds of work. Pooled: the same ratio with every kind of work together. Where the two disagree, the trait sits in kinds of work that draw more (or fewer) applicants per role.

Pay thirds are formed within each kind of work, from the roles listed in the window that state an hourly rate (ties share a third). A listing's age is taken at the middle of the days it was open in the window; roles with no known listing date are left out of that comparison. Expert fields and languages are read from the title with the same vocabularies as the expert-fields and languages reports.

Title words often follow a platform's posting template ('freelance', for example, appears in the titles of a handful of platforms), so within a kind of work a title trait also compares the platforms that use it with those that do not.

Part of every association here is this site's own ranking: the default order puts recently listed roles, roles that state their pay and roles at partner networks higher, and shows each visitor the roles open in their country first. A ratio describes the people who use this site, not a cause.

Clicks on a platform's general sign-up link (1,787 in the window) have no role and are left out, as are 43 clicks on roles with no open time in the window (seen only once, or removed before it).

Source. AITraining.jobs Research, What Draws Applicants to AI Training Roles, September 2026 edition.

Figure 2

Applicants by pay level, within each kind of work

Do better-paid roles draw more applicants than other roles of the same kind?

Pay level within its kind of work	Within-kind index	Clicks per role a week	Median rate, per hour	Roles	Open role-weeks	Clicks
Lowest fifth of its kind's rates	0.91x	0.18	$21	657	1,608	295
Second fifth	0.96x	0.19	$50	701	1,931	367
Middle fifth	1.36x	0.29	$63	643	1,622	466
Fourth fifth	0.65x	0.15	$90	560	1,258	187
Highest fifth	1.32x	0.26	$115	734	1,905	486
No hourly rate stated	0.82x	0.17	n/a	976	2,531	437

Sample. Roles listed from 1 Sep 2026 to 30 Sep 2026 in the 11 kinds of work with at least 25 roles stating an hourly rate (4,271 roles, 2,238 apply clicks).

Within-kind index: the clicks the group drew over the clicks it would have drawn at the average rate of its own kinds of work. 1.00x is that average; the rows of this figure add up to the same clicks observed and expected.

Fifths are formed within each kind of work from the advertised rates of its roles listed in the window, each role once, ties sharing a fifth; the median rate is across kinds, for scale.

Source. AITraining.jobs Research, What Draws Applicants to AI Training Roles, September 2026 edition.

Figure 3

Applicants by listing age, within each kind of work

How much more do new listings draw than older roles of the same kind?

Listing age	Within-kind index	Clicks per role a week	Roles	Open role-weeks	Clicks
First week	4.66x	1.18	693	564	664
Second week	1.85x	0.44	526	1,083	478
Weeks 3 and 4	0.93x	0.20	630	1,769	359
1 to 3 months	0.54x	0.11	1,224	3,334	364
More than 3 months	0.48x	0.09	1,182	4,062	370
Listing date unknownSmall sample	1.48x	0.26	41	99	26

Sample. 2,261 apply clicks on 4,296 roles listed at some point from 1 Sep 2026 to 30 Sep 2026, over 10,910 open role-weeks.

A listing's age is taken at the middle of the days it was open in the window, from its listing date (the job board's posting date, or else its first day in the feed).

Within-kind index: clicks over the clicks expected at the average rate of the role's own kind of work; the rows add up to the same clicks observed and expected.

Recency is a lever of this site's own default order, which puts recently listed roles first, so this figure reads the site's ranking as much as applicants' preference. A newer role is also more likely to be one people have not applied to yet.

Source. AITraining.jobs Research, What Draws Applicants to AI Training Roles, September 2026 edition.

Figure 4

Applicants by who may apply, within each kind of work

Do roles open worldwide draw more applicants than country-limited roles of the same kind?

Location rule	Within-kind index	Clicks per role a week	Roles	Open role-weeks	Clicks
Open worldwide	1.23x	0.28	2,067	4,487	1,251
Worldwide except listed countries	1.71x	0.26	198	819	212
Only the United States	0.78x	0.17	500	1,257	208
Only other listed countries	0.69x	0.14	1,530	4,344	589
No rule statedSmall sample	0.69x	0.28	1	4	1

Sample. 2,261 apply clicks on 4,296 roles listed at some point from 1 Sep 2026 to 30 Sep 2026, over 10,910 open role-weeks.

Only the United States: a rule that names the United States and no other country. Only other listed countries: every other include-list, whatever the countries.

Within-kind index: clicks over the clicks expected at the average rate of the role's own kind of work.

Each visitor is shown the roles open in their own country first, so a country-limited role is ranked first only for visitors from its countries.

Source. AITraining.jobs Research, What Draws Applicants to AI Training Roles, September 2026 edition.

Figure 5

Stating the pay, kind by kind

In which kinds of work do roles that state an hourly rate draw more applicants than roles that do not?

Kind of work	With / without	Clicks per role a week, rate stated	No rate stated	Roles with / without	Clicks with / without
Agentic Tasks & RL EnvironmentsSmall sample	5.80x	1.13	0.19	65 / 35	107 / 15
Coding & Software EvaluationSmall sample	4.28x	0.27	0.06	282 / 100	183 / 17
Writing & ContentSmall sample	3.03x	0.51	0.17	58 / 20	53 / 8
Law, Medicine & Finance	1.26x	0.09	0.07	928 / 195	188 / 31
AI Evaluation & Rating (RLHF)	1.07x	0.40	0.37	478 / 40	380 / 33
Transcription, Audio & Speech	1.04x	0.28	0.26	320 / 104	263 / 82
Multilingual & Localization	1.01x	0.14	0.14	598 / 341	278 / 136
Data Annotation & Labeling	0.90x	0.50	0.55	154 / 70	156 / 80
Math & STEMSmall sample	0.88x	0.18	0.20	250 / 62	127 / 29

Sample. Roles listed from 1 Sep 2026 to 30 Sep 2026 in the 9 of 12 kinds of work with at least 10 roles and 30 open role-weeks both stating and not stating an hourly rate.

Rows with fewer than 30 clicks, 10 roles or 3 roles open on an average day on either side are marked: a few popular roles can move them.

Whether a role states its pay is mostly set by the platform's posting template, so within a kind of work this also compares platforms with one another.

Source. AITraining.jobs Research, What Draws Applicants to AI Training Roles, September 2026 edition.

For labs, data vendors and researchers

Buy the data behind this report

The figures above are a summary. Behind them are records we collect ourselves, every day:

9,572 roles tracked, each with its history
Every role on 29 job boards, read daily since 13 Jul 2026: its pay and place, and the first and last day it was open.
6,200 apply clicks in 28 days
Which roles people actually apply to, by field, pay and kind of work. It happens on this site, so nobody else can count it.
39 lab and vendor hiring boards
The job boards of 25 AI labs and 14 human-data vendors, read every day: who is hiring for human data, and who is hiring to buy it.

What you can buy

The records
The rows behind any figure, with their history, not only the totals.
A regular feed
The same figures every week or month, as files.
A custom cut or benchmark
Your countries, languages or fields, or your own platform against the market.

Method and definitions

Apply clicks made on AITraining.jobs, joined to the role clicked and the days each role was open, with each role's pay, location rule, listing date and title from the public job boards this site reads every day. Every figure is computed by one database query over the window shown, the same query for the live report and for every edition. Each figure states its sample beneath it.

Apply click: A click on an apply button on AITraining.jobs that opens a platform's application page for one role. Made by a person: automated traffic is set aside as it is recorded. Clicks on a platform's general sign-up link have no role and are left out.
Open role-weeks: For every role, the days of the window it was listed, divided by 7 and summed. Only roles listed at some point in the window count.
Clicks per open role a week: Apply clicks divided by open role-weeks.
Within-kind ratio: For each kind of work with at least 10 roles and 30 open role-weeks both with and without the trait (and a click without it), clicks per open role a week with the trait over the same rate without it; then the average across those kinds, weighted by the open role-weeks with the trait. A trait needs two such kinds. 1.00x is no difference.
Pooled ratio: The same ratio with every kind of work together. Where it differs from the within-kind ratio, the trait is concentrated in kinds of work that draw more or fewer applicants per role.
Within-kind index: For a group of roles, the apply clicks they drew over the clicks they would have drawn at the average rate of their own kinds of work. 1.00x is that average.
Pay thirds and fifths: Formed within each kind of work from the advertised hourly rates of its roles listed in the window, each role once, ties sharing a band. Fifths need at least 25 priced roles in the kind.
Listing age: Days from a role's listing date (the job board's posting date, or else its first day in the feed) to the middle of the days it was open in the window. A new listing is under 14 days old.
Location rule: Read from each posting's location text and, where that names no country, from what the posting's own description states: open worldwide, worldwide except listed countries, only listed countries (US-only when the United States is the one country), or no rule stated.
Expert field, language, freelance: Named in the title, matched as whole words against the vocabularies of the expert-fields, languages and job-titles reports.
Kind of work: Each role is classified from its title and description into one of the twelve kinds of AI training work this site tracks. Every kind links to its own page.
Window: Live: the last 28 complete days, in UTC. Edition: one calendar month, in UTC.

What this data cannot tell you

Part of every association is this site's own ranking: the default order puts recently listed roles, roles that state their pay and roles at partner networks higher, and each visitor sees the roles open in their country first. The listing-age, pay and location figures read that ranking as much as applicants' preferences.
A ratio is an association among the people who use AITraining.jobs, not a cause: traits travel together (a new role is often also worldwide), and the comparison holds only the kind of work fixed.
Title words often follow a platform's posting template, so within a kind of work a title trait also compares the platforms that use it with those that do not.
Clicks per role are few (well under one a week on average), so a ratio can move with a handful of popular roles from one window to the next; read a trait across editions before relying on it.
An apply click starts an application. It is not a submitted application, an acceptance or a hire, and one person can click several roles.
Platforms that recruit without a public job board are not counted.

How the job feeds are read and roles are classified: site methodology. How fresh, complete and consistent the data is, checked every day: data quality and coverage.

Editions

The live report covers the last 28 complete days and changes every day. Each complete month is also published as an edition, frozen two days after the month ends, so a figure you cite from an edition never moves.

Live dataLast 28 complete days, recomputed daily
September 2026 edition (this page)Frozen 10 Oct 2026
August 2026 editionFrozen 10 Oct 2026

Cite and reuse

Free to quote, chart and republish under CC BY 4.0 with a link to this page. This edition is frozen, so its figures will read the same later.

Citation

AITraining.jobs (2026). What Draws Applicants to AI Training Roles, September 2026 edition. https://aitraining.run/research/what-draws-applicants-to-ai-training-roles/2026-09

BibTeX

@misc{aitrainingjobs_what_draws_applicants_to_ai_training_roles_2026_09,
  author = {{AITraining.jobs}},
  title = {What Draws Applicants to AI Training Roles, September 2026 edition},
  year = {2026},
  howpublished = {\url{https://aitraining.run/research/what-draws-applicants-to-ai-training-roles/2026-09}},
}

Data: every figure as CSV (one line per value) and JSON, sent to your work email.Pay and open-roles data

More research

All 78 research reports