Key findings

1
Half of the 2,009 AI training roles listed from 3 Aug 2026 to 30 Sep 2026 came down within 28.7 days of being listed (Kaplan-Meier median), and 51.2% were still listed four weeks after going up.
Figure 1
2
Counting only the roles already removed puts the median at 10.5 days; counting the 57.5% still listed too, it is 28.7 days, 173.0% longer.
Figure 4
3
Agentic Tasks & RL Environments turns over fastest (half its roles came down within 8.3 days, 25.5% still listed after four weeks); Transcription, Audio & Speech stays listed longest, with 78.4% of its roles still up after four weeks.
Figure 2
4
Mindrift takes its roles down fastest (half its roles came down within 7.3 days, 0.8% still listed after four weeks) and AfterQuery keeps them longest (99.2% still listed after four weeks); 54.3% of Mindrift's removed roles were re-listed under the same title within 14 days, so much of its turnover is re-posting.
Figure 3
5
12.6% of the removed roles were followed within 14 days by a listing with the same title on the same platform, so part of what reads as turnover is re-posting.
Figure 3
6
1,347 of the 2,726 roles open on 1 Oct 2026 (49.4%) had been listed for 60 days or more, the most in Multilingual & Localization (74.1% of its open roles).
Figure 5
7
In the median complete week, 9.7% of the open roles came down (from 6.4% to 26.7% across 8 weeks); the highest was the week of 7 to 13 Sep, when OpenTrain AI took down 530 of the 721 roles removed.
Figure 6

Figure 1

How many roles are still listed, week by week

What share of AI training roles is still listed one, two, four or eight weeks after it goes up?

Age of the listing	Still listed	Roles listed at least that long ago	Of them still listed
After 1 week	82.2%	1,585	1,322
After 2 weeks	67.6%	1,385	949
After 3 weeks	58.9%	1,179	704
After 4 weeks	51.2%	888	516
After 6 weeks	38.1%	569	233
After 8 weeks	30.1%	132	47

Sample. 2,009 roles listed from 3 Aug 2026 to 30 Sep 2026 with a known listing date, 854 of them removed by 1 Oct 2026; back catalogues left out.

Kaplan-Meier estimate: a role that has come down ends at its last sighting; a role still listed is counted for as long as it has been seen, then leaves the count without being treated as removed.

Age runs from the listing date the job board states (or the first day the role appeared in the feed, if earlier or not stated), in hours. A role first seen more than 2 days after that date surfaced in the feed late and is counted only from the age at which it was first seen; within 2 days it is the daily read catching up and the role counts from its listing day.

A share needs at least 10 roles listed at least that long before 1 Oct 2026; ages the cohort has not reached yet read n/a. Points resting on fewer than 30 such roles are marked as a small sample.

Source. AITraining.jobs Research, How Long AI Training Roles Stay Open, September 2026 edition.

Figure 2

How long roles stay listed, by kind of work

Which kinds of AI training work turn over fastest, and which stay listed longest?

Kind of work	Roles listed	Still listed after 2 weeks	After 4 weeks	After 8 weeks	Median lifetime (days)
All kinds of work	2,009	67.6%	51.2%	30.1%	28.7
Transcription, Audio & Speech	168	85.2%	78.4%	63.9%	n/a
Multilingual & Localization	232	86.4%	69.4%	n/a	n/a
Math & STEM	139	75.6%	60.3%	n/a	40.5
Robotics & Physical-AI Data	42	65.3%	54.6%	n/a	n/a
Coding & Software Evaluation	195	64.5%	52.0%	33.9%	28.3
Law, Medicine & Finance	573	66.7%	51.8%	22.4%	31.0
Writing & Content	48	64.1%	49.7%	n/a	19.0
AI Evaluation & Rating (RLHF)	371	65.3%	44.0%	20.1%	21.4
Data Annotation & Labeling	88	44.0%	37.2%	n/a	12.0
Agentic Tasks & RL Environments	81	29.4%	25.5%	n/a	8.3
Red-Teaming & AI Safety	59	67.8%	n/a	n/a	17.2
Research Studies & SurveysSmall sample	13	38.2%	n/a	n/a	9.8

Sample. 2,009 roles listed from 3 Aug 2026 to 30 Sep 2026 with a known listing date, by the kind of work they are classified as.

Shares still listed and the median lifetime are Kaplan-Meier estimates, so roles still listed count for as long as they have been seen.

Median lifetime n/a: more than half of the group's roles were still listed at the oldest age it can be read at, so its median is not reached yet (it is longer than the time observed), or fewer than 10 of its roles were listed long enough ago to read it. A share reads n/a when fewer than 10 of the group's roles were listed that long ago.

Kinds of work with fewer than 30 roles listed are marked as a small sample.

Source. AITraining.jobs Research, How Long AI Training Roles Stay Open, September 2026 edition.

Figure 3

How long roles stay listed, by platform

On which platforms do AI training roles come down fastest?

Platform	Roles listed	Still listed after 2 weeks	After 4 weeks	After 8 weeks	Median lifetime (days)	Removed roles re-listed within 14 days
micro1	511	80.2%	66.2%	42.9%	39.0	25.9%
OpenTrain AI	488	45.1%	24.4%	n/a	8.8	6.5%
Mercor	457	65.1%	37.1%	11.6%	21.4	0.7%
AfterQuery	133	100.0%	99.2%	n/a	n/a	n/a
Braintrust	92	60.8%	53.6%	n/a	35.0	0.0%
Mindrift	81	6.2%	0.8%	0.0%	7.3	54.3%
TransPerfect DataForce	69	87.1%	58.5%	n/a	34.7	0.0%
Terac	43	57.3%	32.5%	n/a	23.0	0.0%
xAI (SpaceXAI)Small sample	29	87.3%	61.1%	n/a	37.0	0.0%
AlignerrSmall sample	24	91.7%	91.7%	n/a	n/a	n/a
Welo DataSmall sample	24	n/a	n/a	n/a	n/a	n/a
AppenSmall sample	23	75.0%	59.3%	n/a	n/a	n/a
Sigma AISmall sample	13	n/a	n/a	n/a	n/a	n/a
InvisibleSmall sample	10	n/a	n/a	n/a	n/a	n/a
Outlier (Scale AI)Small sample	8	n/a	n/a	n/a	n/a	n/a
Handshake AISmall sample	4	n/a	n/a	n/a	n/a	n/a

Sample. 2,009 roles listed from 3 Aug 2026 to 30 Sep 2026 with a known listing date, by the platform that lists them. Listings only.

Re-listed: a removed role followed, within 14 days of its last sighting, by a listing with the same title on the same platform (shown when the platform has at least 10 removed roles). 12.6% of all removed roles in the cohort were. A platform that takes roles down and posts them again, for example weekly, has short lifetimes here although the opening itself may run for months; listings are counted as the feeds publish them, not merged further.

Some platforms post one listing per country or language, each with its own lifetime.

Platforms with fewer than 30 roles listed are marked as a small sample.

Source. AITraining.jobs Research, How Long AI Training Roles Stay Open, September 2026 edition.

Figure 4

Why the removed-only median is too short

How much does counting only the roles that have come down understate how long roles stay listed?

Kind of work	Roles listed	Still listed on 1 Oct 2026	Median days, removed roles only	Median lifetime, all roles (days)
All kinds of work	2,009	57.5%	10.5	28.7
Transcription, Audio & Speech	168	82.7%	10.8	n/a
Multilingual & Localization	232	75.0%	20.7	n/a
Robotics & Physical-AI Data	42	71.4%	2.3	n/a
Math & STEM	139	66.9%	7.6	40.5
Law, Medicine & Finance	573	57.2%	10.5	31.0
Coding & Software Evaluation	195	55.4%	8.3	28.3
Writing & Content	48	54.2%	9.1	19.0
Data Annotation & Labeling	88	46.6%	5.0	12.0
AI Evaluation & Rating (RLHF)	371	44.5%	12.2	21.4
Agentic Tasks & RL Environments	81	42.0%	7.3	8.3
Red-Teaming & AI Safety	59	18.6%	16.2	17.2

Sample. The same 2,009 roles listed from 3 Aug 2026 to 30 Sep 2026; kinds of work with at least 30 roles.

Removed roles only: the median days from listing to last sighting among the roles that have already come down, the measure a 'median days listed' in a hiring report states. It cannot see the roles still listed, which are the longer-lived ones, so it comes out short, the more so the more roles are still up.

All roles: the Kaplan-Meier median, which counts every role still listed for as long as it has been seen.

A removed-only median needs at least 10 removed roles.

Source. AITraining.jobs Research, How Long AI Training Roles Stay Open, September 2026 edition.

Figure 5

Evergreen roles: open for two months or more

How many of the roles open now have been listed for 60 days or more?

Kind of work	Open on 1 Oct 2026	Listed 60+ days	Share of open roles	Median age of open roles (days)
All kinds of work	2,726	1,347	49.4%	59
Multilingual & Localization	760	563	74.1%	118
Red-Teaming & AI Safety	38	27	71.1%	63
Research Studies & Surveys	16	10	62.5%	196
Data Annotation & Labeling	115	61	53.0%	80
Math & STEM	219	114	52.1%	73
Coding & Software Evaluation	251	123	49.0%	58
Transcription, Audio & Speech	342	146	42.7%	48
Law, Medicine & Finance	628	237	37.7%	44
Writing & Content	41	11	26.8%	34
AI Evaluation & Rating (RLHF)	224	48	21.4%	26
Agentic Tasks & RL Environments	54	5	9.3%	26
Robotics & Physical-AI Data	38	2	5.3%	26

Sample. 2,726 roles open on 1 Oct 2026, of every listing date, back catalogues included.

Evergreen: open on 1 Oct 2026 and listed at least 60 days before it.

A back-catalogue role whose board states no listing date is aged from the first day this site saw it, so its age, and the count of evergreen roles, is a floor.

Kinds of work with fewer than 10 open roles are marked as a small sample.

Source. AITraining.jobs Research, How Long AI Training Roles Stay Open, September 2026 edition.

Figure 6

Weekly churn of open roles

What share of the open AI training roles comes down each week?

Week	Share removed in the week	Open at the week's start	Of them removed	New listings in the week
3 to 9 Aug	18.4%	1,495	275	284
10 to 16 Aug	9.3%	1,493	139	249
17 to 23 Aug	6.4%	1,588	101	161
24 to 30 Aug	11.1%	1,624	180	205
31 Aug to 6 Sep	10.0%	1,632	164	288
7 to 13 Sep	26.7%	2,697	721	576
14 to 20 Sep	7.4%	2,623	195	217
21 to 27 Sep	8.8%	2,636	231	365

Sample. Every complete ISO week (Monday to Sunday, UTC) from 3 Aug 2026 to 30 Sep 2026, on the platforms already tracked at each week's start.

Churn: of the roles open at 00:00 UTC on the Monday, the share last seen during that week and gone from the feed by the end of the observation. Roles listed and removed within the same week are not in the rate; they are counted among the new listings.

Each week counts only platforms tracked before it began, so a platform added to the dataset never reads as churn or growth.

A platform that takes a batch of roles down at once (a project closing, a board reorganised) shows as a single high week.

Source. AITraining.jobs Research, How Long AI Training Roles Stay Open, 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

The public job boards of the AI training platforms this site tracks, read every day since 3 Aug 2026. 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.

Listing date: The date the platform's job board gives for the posting, or else the first day the role appeared in the feed. A role that arrived with a platform's back catalogue and gives no date has no known listing date.
Cohort: Every role with a known listing date listed in the 26 weeks to the end of the period (from 3 Aug 2026, the start of this site's index, at the earliest), except each platform's back catalogue: the roles it already had up in its first week in this dataset.
Lifetime: From the listing date to the last day the role was seen in its platform's feed, in hours, stated in days. A role still listed at the end of the period has lasted at least that long: it is censored, not removed.
Still listed after N weeks: The Kaplan-Meier estimate of the share of roles still in their feed N weeks after their listing date. It counts every role for as long as it has been seen, so roles still listed are not mistaken for short-lived ones. A role first seen more than two days after its listing date surfaced late and counts only from the age it was first seen.
Median lifetime: The age by which half the roles are estimated to have come down. Not reached when more than half are still listed at the oldest age the data covers.
Removed-only median: The median days from listing to last sighting among roles that have already come down, the usual 'median days listed'. It leaves out the roles still up, which are the longer-lived ones, so it is shorter than the median lifetime.
Re-listed: A removed role followed within 14 days by a listing with the same title on the same platform.
Evergreen role: A role open at the end of the period that was listed at least 60 days earlier.
Weekly churn: Of the roles open at 00:00 UTC on a Monday, the share removed during that ISO week, on the platforms tracked before the week began.
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, from AI evaluation and coding to multilingual and audio work. Every kind links to its own page.
Window: Live: roles listed in the 26 weeks to yesterday (from 3 Aug 2026 at the earliest), in UTC. Edition: roles listed in the 26 weeks to the end of that month (from 3 Aug 2026 at the earliest), observed to the end of the month.

What this data cannot tell you

Platforms that recruit without a public job board are not counted. Several large ones invite workers to projects directly.
A listing is not a hire, and a removed listing may have been filled, paused or closed: platforms do not say which.
Some platforms take roles down and post them again (some weekly) or post one listing per country or language, so a short listing life is not always a short opening. The re-listing column shows where.
Feeds are read about once a day: a role listed and removed within a day may be missed, and a removal is dated to the last read that still saw the role.
Ages beyond what the cohort has been observed for are not estimated, so a young edition shows the longer ages as n/a.
A platform that takes a batch of roles down at once (a project closing, a board reorganised) shows as a single high week of churn.

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 follows every role listed in the 26 weeks to the last complete day (from 3 Aug 2026 at the earliest), 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 dataEvery role listed in the last 26 weeks, 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). How Long AI Training Roles Stay Open, September 2026 edition. https://www.aitraining.team/research/how-long-ai-training-roles-stay-open/2026-09

BibTeX

@misc{aitrainingjobs_how_long_ai_training_roles_stay_open_2026_09,
  author = {{AITraining.jobs}},
  title = {How Long AI Training Roles Stay Open, September 2026 edition},
  year = {2026},
  howpublished = {\url{https://www.aitraining.team/research/how-long-ai-training-roles-stay-open/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