Key findings

1
1.5% of the 1,591 AI training roles open on 31 Aug 2026 (24) must be done on-site or hybrid; 75.5% are remote and 23.0% do not say.
Figure 1
2
366 open roles state no arrangement in their title, location or workplace field, 94.3% of them at Invisible, so the remote share is a floor.
Figure 2
3
Data Annotation & Labeling has the largest on-site share of any kind of work with at least 10 open roles (3.3%, 1 roles); 7 kinds of work have any on-site roles at all.
Figure 1
4
3 platforms list on-site roles; Mercor lists the most (14, 58.3% of all on-site roles).
Figure 2
5
On-site roles are open in 5 countries; the United States has the most (15, 62.5%), and the place named most is the San Francisco Bay Area (15 roles).
Figure 3
6
Read from their titles, the largest groups of on-site roles are domain expert evaluation (14 roles, 58.3%) and in-person data collection (9, 37.5%).
Figure 5

Figure 1

On-site, remote or not stated, by kind of work

Which kinds of AI training work have to be done at a site rather than from home?

Kind of work	On-site or hybrid	Remote	Not stated	On-site roles	Open roles
All kinds of work	1.5%	75.5%	23.0%	24	1,591
Research Studies & SurveysSmall sample	16.7%	83.3%	0.0%	1	6
Data Annotation & Labeling	3.3%	80.0%	16.7%	1	30
Law, Medicine & Finance	2.3%	78.2%	19.5%	10	431
Multilingual & Localization	1.6%	59.7%	38.7%	7	432
Math & STEM	1.3%	59.5%	39.2%	2	158
Transcription, Audio & Speech	0.8%	87.7%	11.5%	1	122
Coding & Software Evaluation	0.8%	76.2%	23.0%	1	126
AI Evaluation & Rating (RLHF)	0.5%	97.3%	2.3%	1	219
Agentic Tasks & RL Environments	0.0%	100.0%	0.0%	0	17
Red-Teaming & AI Safety	0.0%	100.0%	0.0%	0	16
Robotics & Physical-AI Data	0.0%	100.0%	0.0%	0	17
Writing & Content	0.0%	100.0%	0.0%	0	17

Sample. 1,591 roles open at the end of 31 Aug 2026; 24 on-site or hybrid.

On-site or hybrid: the title or location names a site ('Onsite', 'On-site London', 'In-person', 'Hybrid'), or a job board that publishes a workplace field (Lever, Ashby, Mercor's own board) marks the role not remote. Remote: the title or location says remote, or the board flags the role remote.

Not stated: neither the title, the location nor a workplace field says how the work is done. Most such roles are country-level freelance listings whose full description says remote further down, so both the on-site and the remote shares are floors, not estimates of the rest.

Hybrid roles are counted with on-site ones: the listing fields do not separate them.

Source. AITraining.jobs Research, On-Site AI Training Work, August 2026 edition.

Figure 2

On-site, remote or not stated, by platform

Which platforms list AI training work that has to be done in person?

Platform	On-site or hybrid	Remote	Not stated	On-site roles	Open roles
Mercor	4.3%	95.7%	0.0%	14	325
TransPerfect DataForce	29.0%	12.9%	58.1%	9	31
Sigma AI	1.3%	98.7%	0.0%	1	79
Invisible	0.0%	34.0%	66.0%	0	523
micro1	0.0%	100.0%	0.0%	0	292
AfterQuery	0.0%	100.0%	0.0%	0	161
Braintrust	0.0%	98.6%	1.4%	0	69
xAI (SpaceXAI)	0.0%	92.6%	7.4%	0	27
Appen	0.0%	100.0%	0.0%	0	24
Terac	0.0%	100.0%	0.0%	0	18
Alignerr	0.0%	100.0%	0.0%	0	17
Mindrift	0.0%	100.0%	0.0%	0	13
Outlier (Scale AI)Small sample	0.0%	100.0%	0.0%	0	9
RemotasksSmall sample	0.0%	100.0%	0.0%	0	3

Sample. 1,591 roles open at the end of 31 Aug 2026, by the platform that lists them (listings, not hires).

Source. AITraining.jobs Research, On-Site AI Training Work, August 2026 edition.

Figure 3

Where on-site AI training roles are, by country

In which countries is in-person AI training work open?

Country	On-site roles	Share of on-site roles	Cities named	Platforms
United States	15	62.5%	San Francisco Bay Area	2
India	4	16.7%	Hyderabad, Pune	1
United Kingdom	2	8.3%	London	1
Norway	1	4.2%	Bergen	1
Singapore	1	4.2%	Singapore	1
Not stated	1	4.2%	n/a	1

Sample. 24 on-site or hybrid roles open at the end of 31 Aug 2026, by the country their location names.

Read from the location text the posting publishes. A role open at sites in two countries counts once in each, so shares can add up to more than 100%.

A city's country is taken from the location itself (the country or US state beside it, or the one country its list names). Two capital letters after a city are read as a US state. A city that cannot be matched is listed with no country.

Source. AITraining.jobs Research, On-Site AI Training Work, August 2026 edition.

Figure 4

Cities with on-site AI training roles

Which cities have the most in-person AI training work?

City or metro area	Country	On-site roles	Naming only this city	Platforms
San Francisco Bay Area	United States	15	15	2
Hyderabad	India	2	2	1
London	United Kingdom	2	2	1
Pune	India	2	2	1
Bergen	Norway	1	1	1
Singapore	Singapore	1	1	1

Sample. 23 of the 24 on-site roles open at the end of 31 Aug 2026 name a city; the 25 cities named most.

A role that may be worked at any of several sites counts once in each city, so 'Naming only this city' counts the roles tied to that city alone.

Cities of one metropolitan area are merged: Sunnyvale, Burlingame, Santa Clara and San Francisco count as the San Francisco Bay Area; Redmond and Bellevue as the Seattle area; Brooklyn as New York City.

Source. AITraining.jobs Research, On-Site AI Training Work, August 2026 edition.

Figure 5

What on-site AI training roles are

What do the AI training roles that must be done in person involve?

Work, read from the title	On-site roles	Share of on-site roles	Main kind of work	Example title	Platforms
Domain expert evaluation	14	58.3%	Law, Medicine & Finance	Accounting & Audit Domain Expert	1
In-person data collection (studies, recording sessions and their moderators)	9	37.5%	Multilingual & Localization	Camellia Onsite Collection - Sunnyvale, CA	1
Linguistic and language projects	1	4.2%	Multilingual & Localization	English Linguistic Projects - On-site (London)	1

Sample. 24 on-site or hybrid roles open at the end of 31 Aug 2026.

Each role counts once, in the first group whose words its title contains, in the order listed in the method (robotics first, then data collection, speech and voice, labelling, linguistic, policy, expert work).

Source. AITraining.jobs Research, On-Site AI Training Work, August 2026 edition.

Figure 6

Advertised pay of on-site and remote roles, within each kind of work

Does in-person AI training work advertise more than remote work of the same kind?

Kind of work	On-site median (per hour)	On-site roles stating a rate	Remote median (per hour)	Remote roles stating a rate	On-site against remote
Law, Medicine & Finance	$93	10	$88	328	+5.7%
Math & STEMSmall sample	n/a	2	$85	87	n/a
Coding & Software EvaluationSmall sample	n/a	1	$78	90	n/a
AI Evaluation & Rating (RLHF)Small sample	n/a	1	$75	211	n/a
Transcription, Audio & SpeechSmall sample	n/a	0	$36	81	n/a
Data Annotation & LabelingSmall sample	n/a	0	$22	20	n/a
Multilingual & LocalizationSmall sample	n/a	0	$40	175	n/a
Research Studies & SurveysSmall sample	n/a	0	n/a	3	n/a

Sample. Roles open at the end of 31 Aug 2026 that state an hourly rate, in the kinds of work with on-site roles.

Medians need at least 8 roles stating a rate on each side; the comparison is made within each kind of work so the mix of work is not read as a premium.

On-site roles advertising an annual salary, a fee per session or no figure are not in the medians.

Source. AITraining.jobs Research, On-Site AI Training Work, August 2026 edition.

Figure 7

On-site roles over the window, like for like

Is in-person AI training work growing or shrinking?

Work arrangement	Open on 1 Aug 2026	Open on 31 Aug 2026	Change	New listings in the window
On-site or hybridSmall sample	9	24	+166.7%	24
Remote	1,130	1,201	+6.3%	892
Not stated	381	366	-3.9%	22
All roles	1,520	1,591	+4.7%	938

Sample. Roles open at the start of 1 Aug 2026 and at the end of 31 Aug 2026 on the platforms tracked throughout; new listings over the whole window.

Open counts compare only platforms tracked before the window opened, so a platform added during it does not read as growth.

New listings: roles first seen in the window on a platform already tracked before that week (a platform's first week is its back catalogue).

Source. AITraining.jobs Research, On-Site AI Training Work, August 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: each open role's title, its location text and the workplace field its board publishes. 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.

On-site or hybrid: The role must be done at a named place at least part of the week: its title or location says so ('Onsite', 'On-site London', 'In-person', 'Hybrid'), or a job board that publishes a workplace field of its own (Lever, Ashby, Mercor's board) marks it not remote. Hybrid roles are counted here because the listing fields do not separate them from on-site ones.
Remote: The role's title or location says remote (or work from home, worldwide), or its board flags it remote.
Not stated: Neither the title, the location nor a workplace field says how the work is done. On most boards the stored remote flag is inferred from the word 'remote' near the top of the posting, so a missing word is not read as on-site.
City or metro area: Read from the location text: a city's country is the country or US state beside it, or the one country its list names, never guessed. Cities of one metropolitan area are merged (Sunnyvale and Burlingame count as the San Francisco Bay Area, Redmond as the Seattle area). A role open at several sites counts once in each.
Work, read from the title: Each on-site role falls in the first of these groups whose words its title contains: robotics; in-person data collection (collection, study, recording, participants, moderators); speech and voice; labelling and data quality; linguistic projects; policy and safety evaluation; domain expert work; other.
Advertised pay: The hourly figure a posting states (the middle of a stated range), compared within each kind of work. Medians need at least 8 roles stating a rate on each side.
New listing: A role that first appeared during the window on a platform already tracked before that week. A platform's first week in this dataset is its back catalogue and is not counted as new.
Open role: A role listed by a given day and still in its platform's feed that day.
Window: Live: the last 28 complete days, in UTC. Edition: one calendar month, in UTC.

What this data cannot tell you

Both shares are floors. When this report was built (10 Oct 2026), a check of the full descriptions of the largest group of 'not stated' roles found nearly all of them say remote further down the posting, and a few postings that require an on-site session say so only in their description.
Hybrid and fully on-site roles cannot be told apart from the listing fields: descriptions of some roles counted on-site ask for on-site days each week rather than every day.
Location text says where a role is worked or where its applicants must live, and the two are not always distinguishable; two capital letters after a city are read as a US state.
Platforms that recruit without a public job board are not counted. Several large ones invite workers to in-person studies directly.
A listing is not a hire, and some platforms post one listing per language or site, which weighs a city with many such listings more.

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 editionFrozen 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). On-Site AI Training Work, August 2026 edition. https://aitraining.run/research/on-site-ai-training-work/2026-08

BibTeX

@misc{aitrainingjobs_on_site_ai_training_work_2026_08,
  author = {{AITraining.jobs}},
  title = {On-Site AI Training Work, August 2026 edition},
  year = {2026},
  howpublished = {\url{https://aitraining.run/research/on-site-ai-training-work/2026-08}},
}

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