Key findings
- 1
Rating or scoring responses against guidelines is the task mentioned most often, in 43.1% of the 4,000 roles read, ahead of annotating, labelling or tagging data (39.1%).
- 3
The median role read mentions 3 different tasks; 57.5% mention three or more, and 6.6% name none.
- 4
The highest share in any kind of work is red-teaming or adversarial testing in Red-Teaming & AI Safety: 81.6% of its roles read mention it, against 2.9% of all roles read.
- 5
Among tasks with at least 3 priced roles on 3 or more platforms, roles mentioning reviewing others' work advertise the highest median ($80/hr) and roles mentioning performing physical tasks for robot data the lowest ($15/hr), against $40/hr for all roles read.
- 6
The platforms differ most on rating or scoring responses against guidelines: 100.0% of Cohere's roles read mention it, against 0.0% at TransPerfect DataForce.
- 7
Beyond judging and authoring, 19.4% of roles read mention recording or transcribing audio and video, 10.4% writing or reviewing code, and 2.9% red-teaming or adversarial testing.
Figure 1
Tasks AI training roles describe
What do AI training roles ask people to do?
| Task | Roles mentioning it | Share of roles read |
|---|---|---|
| Rate or score responses against guidelines | 1,725 | 43.1% |
| Annotate, label or tag data | 1,564 | 39.1% |
| Create evaluation tasks, benchmarks or test cases | 1,146 | 28.7% |
| Fact-check or verify accuracy | 910 | 22.8% |
| Write prompts or questions | 840 | 21.0% |
| Translate or localise | 693 | 17.3% |
| Give feedback on reasoning or explanations | 651 | 16.3% |
| Record audio, video or photos | 578 | 14.5% |
| Write ideal responses or demonstrations | 536 | 13.4% |
| Solve problems step by step | 502 | 12.6% |
| Hold conversations or sessions | 466 | 11.7% |
| Compare or rank model responses | 459 | 11.5% |
| Edit or rewrite text | 406 | 10.2% |
| Transcribe audio or video | 302 | 7.6% |
| Review or debug code | 300 | 7.5% |
| Write rationales for judgments | 276 | 6.9% |
| Write code or tests | 271 | 6.8% |
| Review agent trajectories or tool use | 257 | 6.4% |
| Review for safety or policy | 238 | 5.9% |
| Write rubrics, criteria or guidelines | 223 | 5.6% |
| Review others' work (QA) | 184 | 4.6% |
| Take part in surveys or studies | 169 | 4.2% |
| Red-team, jailbreak or adversarial testing | 115 | 2.9% |
| Do hands-on professional work (models, analyses, simulations) | 94 | 2.4% |
| Perform physical tasks for robot data | 47 | 1.2% |
Sample. 4,000 roles open on 31 Oct 2026 whose job board publishes a full description (AfterQuery, Aligned Labs, Anthropic, Anyone AI, Appen, Babel Audio, Braintrust, Claru, Cohere, DataAnnotation, DATAmundi, Edison Scientific, Elsevier Health, EvolutionIQ, Fleet AI, Handshake AI, Headway, iMerit, Innodata, Invisible, Inworld AI, JazzX AI, Kastle, Lifted (Upwork), Limbic, Mercor, micro1, Mindrift, Mistral AI, OneForma, OpenTrain AI, Outlier (Scale AI), Pasiflora AI, Perle Labs, Pila8, PointClickCare, Prolific, RealTrainAI, RemoExperts, Remotasks, Rise Data Labs, RWS TrainAI Community, Sama, Scale AI Human Frontier Collective, Shaip (Collaborate by Shaip), Sigma AI, Snorkel AI Expert Community, Sovrano AI, SuperAnnotate (SME), Surge AI, TELUS Digital AI, Terac, TransPerfect DataForce, Trellis, Turing, Vals AI, Vetto, Volga Partners, Weekday, Welo Data and xAI (SpaceXAI)), 91.3% of the 4,381 open roles; the boards that publish descriptions are expert networks, so these roles lean towards expert work.
A role mentions a task when its description contains one of the task's documented patterns, anywhere in the text: a mention is not necessarily the main task, and a role often mentions several.
Hiring steps are not tasks: a description saying applications are read against a rubric, or that applying starts with an AI interview, does not count as writing rubrics or holding conversations.
Source. AITraining.jobs Research, What AI Trainers Do, October 2026 edition.
Figure 2
Tasks by kind of work
Which tasks does each kind of AI training work involve?
| Task | Domain Experts (1,445) | Multilingual (840) | Transcription & Audio (526) | Coding (259) | Math & STEM (232) | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Rate or score responses against guidelines | 58% | 44% | 12% | 50% | 33% | 32% | 56% | 23% | 0% | 53% | 72% | 0% |
| Annotate, label or tag data | 29% | 57% | 52% | 29% | 18% | 54% | 55% | 9% | 13% | 42% | 40% | 19% |
| Create evaluation tasks, benchmarks or test cases | 47% | 9% | 7% | 32% | 49% | 14% | 19% | 50% | 2% | 34% | 24% | 25% |
| Fact-check or verify accuracy | 24% | 36% | 1% | 33% | 35% | 8% | 21% | 8% | 1% | 29% | 36% | 0% |
| Write prompts or questions | 25% | 27% | 1% | 32% | 35% | 11% | 10% | 8% | 0% | 37% | 60% | 0% |
| Translate or localise | 1% | 63% | 15% | 2% | 3% | 17% | 6% | 2% | 4% | 3% | 12% | 0% |
| Give feedback on reasoning or explanations | 26% | 9% | 0% | 25% | 33% | 3% | 16% | 3% | 0% | 13% | 28% | 0% |
| Record audio, video or photos | 1% | 18% | 51% | 4% | 0% | 10% | 21% | 23% | 45% | 0% | 0% | 6% |
| Write ideal responses or demonstrations | 19% | 9% | 0% | 21% | 16% | 3% | 5% | 41% | 0% | 42% | 32% | 6% |
| Solve problems step by step | 10% | 19% | 1% | 22% | 43% | 3% | 5% | 4% | 0% | 3% | 8% | 0% |
| Hold conversations or sessions | 5% | 20% | 12% | 14% | 30% | 2% | 25% | 1% | 1% | 8% | 0% | 6% |
| Compare or rank model responses | 10% | 19% | 0% | 18% | 3% | 5% | 35% | 3% | 0% | 26% | 32% | 0% |
| Edit or rewrite text | 6% | 28% | 4% | 3% | 3% | 13% | 8% | 1% | 0% | 0% | 48% | 0% |
| Transcribe audio or video | 0% | 12% | 37% | 0% | 0% | 2% | 0% | 0% | 5% | 0% | 0% | 0% |
| Review or debug code | 3% | 2% | 0% | 75% | 4% | 1% | 10% | 17% | 0% | 5% | 0% | 6% |
| Write rationales for judgments | 10% | 4% | 5% | 6% | 3% | 4% | 13% | 1% | 0% | 18% | 16% | 13% |
| Write code or tests | 4% | 0% | 2% | 51% | 5% | 1% | 6% | 26% | 1% | 18% | 8% | 13% |
| Review agent trajectories or tool use | 6% | 3% | 4% | 6% | 4% | 4% | 9% | 54% | 8% | 13% | 8% | 13% |
| Review for safety or policy | 8% | 4% | 1% | 6% | 3% | 7% | 8% | 3% | 0% | 79% | 16% | 6% |
| Write rubrics, criteria or guidelines | 10% | 1% | 6% | 2% | 4% | 1% | 3% | 8% | 0% | 3% | 12% | 0% |
| Review others' work (QA) | 6% | 0% | 9% | 9% | 4% | 1% | 3% | 3% | 0% | 11% | 0% | 0% |
| Take part in surveys or studies | 5% | 7% | 0% | 3% | 1% | 1% | 6% | 1% | 1% | 3% | 4% | 44% |
| Red-team, jailbreak or adversarial testing | 4% | 0% | 0% | 3% | 1% | 1% | 5% | 4% | 0% | 82% | 0% | 6% |
| Do hands-on professional work (models, analyses, simulations) | 5% | 0% | 0% | 2% | 3% | 1% | 1% | 10% | 0% | 0% | 0% | 0% |
| Perform physical tasks for robot data | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 0% | 49% | 0% | 0% | 0% |
Sample. 4,000 roles open on 31 Oct 2026 whose job board publishes a full description (AfterQuery, Aligned Labs, Anthropic, Anyone AI, Appen, Babel Audio, Braintrust, Claru, Cohere, DataAnnotation, DATAmundi, Edison Scientific, Elsevier Health, EvolutionIQ, Fleet AI, Handshake AI, Headway, iMerit, Innodata, Invisible, Inworld AI, JazzX AI, Kastle, Lifted (Upwork), Limbic, Mercor, micro1, Mindrift, Mistral AI, OneForma, OpenTrain AI, Outlier (Scale AI), Pasiflora AI, Perle Labs, Pila8, PointClickCare, Prolific, RealTrainAI, RemoExperts, Remotasks, Rise Data Labs, RWS TrainAI Community, Sama, Scale AI Human Frontier Collective, Shaip (Collaborate by Shaip), Sigma AI, Snorkel AI Expert Community, Sovrano AI, SuperAnnotate (SME), Surge AI, TELUS Digital AI, Terac, TransPerfect DataForce, Trellis, Turing, Vals AI, Vetto, Volga Partners, Weekday, Welo Data and xAI (SpaceXAI)), 91.3% of the 4,381 open roles; the boards that publish descriptions are expert networks, so these roles lean towards expert work. Kinds of work whose roles read outweigh the average of the kinds (at least 5.6 roles read, the weight of the average fitted on the table's own counts); each cell is the share of that kind's roles read that mention the task (roles read in the column header).
A role mentions a task when its description contains one of the task's documented patterns, anywhere in the text: a mention is not necessarily the main task, and a role often mentions several.
Every kind of work with a role read is shown.
Source. AITraining.jobs Research, What AI Trainers Do, October 2026 edition.
Figure 3
Advertised pay of roles mentioning each task
Which tasks come with the highest advertised pay?
| Task | Median per hour | Typical pay | Roles stating pay |
|---|---|---|---|
| Review others' work (QA) | $80 | $60 to $125 | 106 |
| Write ideal responses or demonstrations | $70 | $40 to $115 | 462 |
| Do hands-on professional work (models, analyses, simulations) | $70 | $40 to $115 | 79 |
| Red-team, jailbreak or adversarial testing | $70 | $40 to $100 | 81 |
| Write rubrics, criteria or guidelines | $70 | $35 to $125 | 203 |
| Create evaluation tasks, benchmarks or test cases | $65 | $40 to $120 | 937 |
| Write code or tests | $65 | $50 to $110 | 199 |
| Review or debug code | $60 | $35 to $100 | 245 |
| Give feedback on reasoning or explanations | $55 | $35 to $110 | 611 |
| Write rationales for judgments | $50 | $35 to $150 | 216 |
| Rate or score responses against guidelines | $50 | $35 to $120 | 1,503 |
| Review for safety or policy | $50 | $35 to $100 | 191 |
| Review agent trajectories or tool use | $40 | $20 to $105 | 155 |
| Compare or rank model responses | $40 | $25 to $100 | 389 |
| Annotate, label or tag data | $35 | $25 to $95 | 1,150 |
| Hold conversations or sessions | $35 | $35 | 372 |
| Edit or rewrite text | $35 | $25 to $70 | 375 |
| Fact-check or verify accuracy | $35 | $35 to $80 | 852 |
| Solve problems step by step | $35 | $35 to $85 | 443 |
| Write prompts or questions | $35 | $35 to $80 | 778 |
| Record audio, video or photos | $30 | $15 to $55 | 328 |
| Take part in surveys or studies | $30 | $25 to $150 | 128 |
| Translate or localise | $30 | $25 to $40 | 509 |
| Transcribe audio or video | $25 | $15 to $50 | 114 |
| Perform physical tasks for robot data | $15 | $11 to $50 | 13 |
Sample. 4,000 roles open on 31 Oct 2026 whose job board publishes a full description (AfterQuery, Aligned Labs, Anthropic, Anyone AI, Appen, Babel Audio, Braintrust, Claru, Cohere, DataAnnotation, DATAmundi, Edison Scientific, Elsevier Health, EvolutionIQ, Fleet AI, Handshake AI, Headway, iMerit, Innodata, Invisible, Inworld AI, JazzX AI, Kastle, Lifted (Upwork), Limbic, Mercor, micro1, Mindrift, Mistral AI, OneForma, OpenTrain AI, Outlier (Scale AI), Pasiflora AI, Perle Labs, Pila8, PointClickCare, Prolific, RealTrainAI, RemoExperts, Remotasks, Rise Data Labs, RWS TrainAI Community, Sama, Scale AI Human Frontier Collective, Shaip (Collaborate by Shaip), Sigma AI, Snorkel AI Expert Community, Sovrano AI, SuperAnnotate (SME), Surge AI, TELUS Digital AI, Terac, TransPerfect DataForce, Trellis, Turing, Vals AI, Vetto, Volga Partners, Weekday, Welo Data and xAI (SpaceXAI)), 91.3% of the 4,381 open roles; the boards that publish descriptions are expert networks, so these roles lean towards expert work. 3,185 of the 4,000 roles read state an hourly rate (median $40/hr).
A role mentions a task when its description contains one of the task's documented patterns, anywhere in the text: a mention is not necessarily the main task, and a role often mentions several.
Pay is the hourly figure a posting advertises (the middle of a stated range), counted once per distinct rate as the job board counts it. The median and typical pay are the board's: typical pay runs from the 25th percentile of what roles pay (the middle of each advertised range) to the 90th-percentile advertised ceiling, stated from 3 distinct rates. A task's median also needs rates from at least 3 platforms; others are marked and listed after the ranked tasks. 1.00x is the median of all roles read.
A role mentioning several tasks counts in each, and tasks travel with kinds of work and seniority: a pay difference is an association, not the price of the task.
Source. AITraining.jobs Research, What AI Trainers Do, October 2026 edition.
Figure 4
Tasks by platform
How do the platforms that publish descriptions differ in the work they describe?
| Task | AfterQuery (229) | Aligned Labs (15) | Anthropic (1) | Anyone AI (133) | Appen (42) | Babel Audio (47) | Braintrust (77) | Claru (42) | Cohere (4) | DataAnnotation (112) | DATAmundi (8) | Edison Scientific (1) | Elsevier Health (6) | EvolutionIQ (1) | Fleet AI (2) | Handshake AI (184) | Headway (1) | iMerit (9) | Innodata (31) | Invisible (533) | Inworld AI (1) | JazzX AI (1) | Kastle (1) | Lifted (Upwork) (45) | Limbic (1) | Mercor (237) | micro1 (302) | Mindrift (90) | Mistral AI (3) | OneForma (17) | OpenTrain AI (40) | Outlier (Scale AI) (46) | Pasiflora AI (7) | Perle Labs (67) | Pila8 (4) | PointClickCare (1) | Prolific (124) | RealTrainAI (4) | RemoExperts (32) | Remotasks (3) | Rise Data Labs (108) | RWS TrainAI Community (143) | Sama (3) | Scale AI Human Frontier Collective (10) | Shaip (Collaborate by Shaip) (9) | Sigma AI (81) | Snorkel AI Expert Community (1) | Sovrano AI (36) | SuperAnnotate (SME) (379) | Surge AI (13) | TELUS Digital AI (67) | Terac (37) | TransPerfect DataForce (44) | Trellis (1) | Turing (175) | Vals AI (1) | Vetto (20) | Volga Partners (59) | Weekday (108) | Welo Data (184) | xAI (SpaceXAI) (17) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Rate or score responses against guidelines | 54% | 40% | 0% | 20% | 0% | 0% | 48% | 14% | 100% | 62% | 0% | 100% | 100% | 0% | 100% | 94% | 100% | 89% | 16% | 32% | 0% | 0% | 0% | 36% | 100% | 36% | 38% | 0% | 0% | 6% | 43% | 98% | 57% | 0% | 0% | 0% | 25% | 0% | 72% | 67% | 93% | 16% | 0% | 60% | 0% | 0% | 0% | 53% | 91% | 54% | 27% | 24% | 0% | 0% | 46% | 0% | 25% | 76% | 44% | 24% | 18% |
| Annotate, label or tag data | 8% | 13% | 0% | 0% | 19% | 45% | 38% | 36% | 100% | 1% | 25% | 100% | 0% | 100% | 0% | 8% | 100% | 67% | 97% | 30% | 100% | 0% | 100% | 62% | 0% | 23% | 20% | 21% | 100% | 41% | 43% | 91% | 71% | 91% | 0% | 100% | 29% | 0% | 31% | 0% | 91% | 59% | 100% | 0% | 0% | 99% | 0% | 8% | 90% | 8% | 34% | 14% | 18% | 100% | 52% | 0% | 10% | 90% | 22% | 38% | 100% |
| Create evaluation tasks, benchmarks or test cases | 65% | 33% | 100% | 90% | 0% | 2% | 34% | 2% | 100% | 30% | 0% | 100% | 0% | 100% | 100% | 50% | 0% | 0% | 10% | 7% | 0% | 0% | 0% | 13% | 0% | 46% | 27% | 43% | 33% | 12% | 43% | 0% | 0% | 0% | 0% | 0% | 23% | 0% | 38% | 0% | 19% | 8% | 0% | 80% | 44% | 0% | 100% | 92% | 12% | 69% | 13% | 43% | 2% | 0% | 61% | 100% | 15% | 44% | 43% | 17% | 18% |
| Fact-check or verify accuracy | 2% | 0% | 0% | 2% | 0% | 0% | 33% | 5% | 50% | 0% | 0% | 0% | 17% | 0% | 0% | 1% | 0% | 67% | 7% | 63% | 0% | 0% | 0% | 9% | 0% | 5% | 10% | 0% | 0% | 0% | 3% | 0% | 0% | 0% | 0% | 100% | 20% | 0% | 3% | 67% | 13% | 0% | 0% | 0% | 0% | 0% | 0% | 3% | 83% | 0% | 13% | 11% | 2% | 0% | 26% | 0% | 5% | 71% | 12% | 5% | 0% |
| Write prompts or questions | 16% | 27% | 0% | 2% | 0% | 0% | 8% | 2% | 25% | 27% | 0% | 0% | 0% | 0% | 0% | 27% | 0% | 0% | 65% | 60% | 0% | 0% | 0% | 9% | 0% | 11% | 6% | 0% | 0% | 6% | 30% | 0% | 0% | 0% | 0% | 0% | 1% | 0% | 13% | 33% | 3% | 8% | 0% | 0% | 0% | 0% | 0% | 3% | 56% | 15% | 5% | 3% | 5% | 0% | 21% | 0% | 0% | 39% | 5% | 2% | 0% |
| Translate or localise | 1% | 0% | 0% | 1% | 0% | 0% | 31% | 0% | 25% | 37% | 0% | 0% | 0% | 0% | 0% | 1% | 0% | 44% | 0% | 28% | 100% | 0% | 0% | 4% | 0% | 0% | 3% | 0% | 0% | 47% | 8% | 0% | 0% | 70% | 0% | 0% | 21% | 0% | 0% | 33% | 3% | 5% | 0% | 0% | 0% | 90% | 0% | 0% | 31% | 0% | 25% | 0% | 7% | 0% | 17% | 0% | 5% | 14% | 2% | 54% | 71% |
| Give feedback on reasoning or explanations | 6% | 27% | 0% | 35% | 0% | 0% | 31% | 10% | 25% | 17% | 0% | 0% | 0% | 0% | 50% | 4% | 0% | 0% | 0% | 4% | 0% | 0% | 0% | 4% | 0% | 18% | 8% | 21% | 0% | 6% | 10% | 0% | 0% | 2% | 0% | 0% | 16% | 0% | 19% | 0% | 6% | 0% | 0% | 0% | 0% | 0% | 0% | 19% | 70% | 54% | 16% | 8% | 0% | 0% | 12% | 0% | 10% | 75% | 17% | 1% | 18% |
| Record audio, video or photos | 20% | 0% | 0% | 0% | 76% | 13% | 33% | 62% | 0% | 0% | 50% | 0% | 0% | 0% | 0% | 2% | 0% | 0% | 10% | 6% | 0% | 0% | 0% | 29% | 0% | 18% | 6% | 0% | 0% | 12% | 8% | 94% | 29% | 12% | 100% | 0% | 21% | 0% | 3% | 0% | 0% | 32% | 0% | 0% | 78% | 88% | 0% | 0% | 9% | 0% | 10% | 3% | 30% | 0% | 6% | 0% | 0% | 7% | 2% | 18% | 71% |
| Write ideal responses or demonstrations | 45% | 0% | 0% | 9% | 0% | 0% | 25% | 5% | 25% | 40% | 0% | 0% | 0% | 0% | 50% | 1% | 0% | 0% | 3% | 0% | 0% | 0% | 0% | 7% | 0% | 20% | 7% | 46% | 33% | 0% | 28% | 0% | 57% | 0% | 0% | 0% | 0% | 0% | 44% | 33% | 4% | 0% | 0% | 0% | 0% | 0% | 100% | 8% | 36% | 8% | 5% | 3% | 0% | 0% | 19% | 0% | 15% | 2% | 20% | 1% | 0% |
| Solve problems step by step | 3% | 7% | 0% | 1% | 0% | 0% | 0% | 5% | 75% | 12% | 0% | 0% | 0% | 0% | 0% | 1% | 0% | 0% | 0% | 60% | 0% | 0% | 0% | 11% | 0% | 5% | 5% | 0% | 0% | 0% | 5% | 0% | 0% | 2% | 0% | 0% | 7% | 0% | 3% | 0% | 1% | 0% | 0% | 20% | 0% | 0% | 0% | 14% | 4% | 0% | 12% | 11% | 0% | 0% | 24% | 0% | 10% | 29% | 9% | 1% | 18% |
| Hold conversations or sessions | 0% | 0% | 0% | 0% | 2% | 11% | 0% | 2% | 0% | 0% | 25% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 7% | 61% | 0% | 0% | 0% | 4% | 0% | 2% | 0% | 0% | 0% | 0% | 3% | 98% | 0% | 8% | 75% | 0% | 10% | 0% | 6% | 0% | 0% | 6% | 0% | 0% | 11% | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 39% | 0% | 3% | 0% | 10% | 3% | 0% | 8% | 0% |
| Compare or rank model responses | 3% | 0% | 0% | 2% | 0% | 0% | 20% | 7% | 50% | 26% | 0% | 0% | 0% | 0% | 0% | 6% | 0% | 78% | 3% | 0% | 0% | 0% | 0% | 2% | 0% | 3% | 2% | 0% | 33% | 0% | 5% | 91% | 57% | 0% | 0% | 0% | 27% | 0% | 19% | 67% | 0% | 10% | 0% | 0% | 0% | 0% | 0% | 8% | 56% | 0% | 3% | 3% | 0% | 0% | 18% | 0% | 5% | 2% | 6% | 2% | 0% |
| Edit or rewrite text | 1% | 0% | 0% | 0% | 7% | 2% | 29% | 2% | 50% | 9% | 0% | 0% | 17% | 0% | 0% | 1% | 0% | 11% | 65% | 14% | 0% | 0% | 0% | 4% | 0% | 0% | 10% | 0% | 0% | 6% | 10% | 0% | 0% | 0% | 0% | 0% | 22% | 0% | 0% | 0% | 1% | 9% | 0% | 0% | 0% | 0% | 0% | 0% | 38% | 39% | 6% | 0% | 5% | 0% | 5% | 0% | 10% | 15% | 5% | 5% | 0% |
| Transcribe audio or video | 0% | 0% | 0% | 0% | 14% | 45% | 3% | 7% | 0% | 0% | 25% | 0% | 0% | 0% | 0% | 1% | 0% | 11% | 0% | 0% | 0% | 0% | 0% | 40% | 0% | 3% | 9% | 0% | 0% | 12% | 23% | 0% | 14% | 91% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 0% | 33% | 98% | 0% | 0% | 0% | 0% | 12% | 0% | 2% | 0% | 1% | 0% | 0% | 10% | 0% | 13% | 71% |
| Review or debug code | 8% | 0% | 0% | 22% | 0% | 0% | 4% | 5% | 25% | 8% | 0% | 0% | 0% | 0% | 0% | 1% | 0% | 0% | 0% | 7% | 0% | 0% | 0% | 7% | 0% | 6% | 7% | 0% | 0% | 0% | 3% | 0% | 0% | 0% | 0% | 0% | 5% | 0% | 19% | 0% | 0% | 0% | 0% | 30% | 0% | 0% | 0% | 11% | 16% | 0% | 6% | 19% | 0% | 0% | 29% | 0% | 5% | 0% | 15% | 0% | 0% |
| Write rationales for judgments | 2% | 0% | 0% | 0% | 0% | 0% | 1% | 12% | 25% | 7% | 0% | 0% | 0% | 0% | 0% | 5% | 0% | 67% | 10% | 0% | 0% | 0% | 0% | 16% | 0% | 22% | 16% | 21% | 0% | 6% | 3% | 0% | 43% | 0% | 0% | 0% | 8% | 0% | 13% | 0% | 1% | 0% | 0% | 0% | 0% | 0% | 0% | 6% | 5% | 0% | 3% | 3% | 0% | 0% | 16% | 0% | 30% | 20% | 7% | 7% | 0% |
| Write code or tests | 13% | 0% | 0% | 21% | 0% | 0% | 1% | 7% | 25% | 4% | 0% | 0% | 17% | 0% | 0% | 2% | 0% | 0% | 0% | 1% | 0% | 100% | 0% | 22% | 0% | 5% | 9% | 0% | 0% | 0% | 23% | 0% | 0% | 8% | 0% | 0% | 9% | 0% | 13% | 0% | 3% | 0% | 0% | 100% | 22% | 1% | 0% | 17% | 7% | 8% | 8% | 11% | 0% | 0% | 21% | 0% | 10% | 0% | 16% | 1% | 0% |
| Review agent trajectories or tool use | 5% | 0% | 0% | 8% | 0% | 0% | 1% | 19% | 75% | 0% | 0% | 0% | 0% | 0% | 100% | 0% | 0% | 0% | 10% | 1% | 0% | 0% | 0% | 16% | 0% | 9% | 1% | 46% | 0% | 0% | 10% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 3% | 0% | 30% | 0% | 0% | 100% | 3% | 5% | 62% | 2% | 11% | 0% | 0% | 50% | 0% | 0% | 0% | 7% | 0% | 0% |
| Review for safety or policy | 1% | 0% | 100% | 0% | 0% | 0% | 1% | 10% | 25% | 1% | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 11% | 13% | 1% | 0% | 0% | 0% | 0% | 100% | 11% | 5% | 0% | 0% | 0% | 10% | 0% | 57% | 0% | 0% | 0% | 1% | 0% | 0% | 0% | 0% | 0% | 0% | 30% | 0% | 0% | 0% | 8% | 28% | 8% | 18% | 0% | 5% | 0% | 12% | 100% | 10% | 10% | 4% | 2% | 0% |
| Write rubrics, criteria or guidelines | 0% | 0% | 0% | 4% | 0% | 0% | 0% | 0% | 25% | 3% | 0% | 0% | 0% | 0% | 0% | 0% | 100% | 0% | 0% | 8% | 0% | 0% | 0% | 0% | 0% | 22% | 8% | 0% | 67% | 0% | 5% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 19% | 0% | 11% | 0% | 0% | 0% | 0% | 0% | 0% | 50% | 1% | 8% | 3% | 8% | 0% | 0% | 7% | 100% | 0% | 2% | 17% | 0% | 77% |
| Review others' work (QA) | 5% | 0% | 0% | 13% | 0% | 0% | 4% | 2% | 0% | 5% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 3% | 0% | 0% | 0% | 0% | 0% | 0% | 9% | 3% | 0% | 0% | 0% | 0% | 0% | 0% | 69% | 0% | 0% | 2% | 0% | 19% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 42% | 4% | 0% | 2% | 3% | 0% | 0% | 10% | 0% | 0% | 0% | 7% | 2% | 0% |
| Take part in surveys or studies | 0% | 0% | 0% | 0% | 0% | 0% | 1% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 3% | 0% | 0% | 0% | 6% | 0% | 0% | 0% | 0% | 0% | 0% | 100% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 3% | 1% | 0% | 0% | 73% | 0% | 0% | 1% | 0% | 0% | 3% | 0% | 0% | 0% |
| Red-team, jailbreak or adversarial testing | 0% | 27% | 0% | 0% | 0% | 0% | 1% | 2% | 25% | 9% | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 0% | 3% | 1% | 0% | 0% | 0% | 4% | 0% | 11% | 2% | 0% | 0% | 0% | 13% | 0% | 14% | 0% | 0% | 0% | 0% | 0% | 3% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 8% | 7% | 8% | 8% | 0% | 0% | 0% | 5% | 0% | 5% | 0% | 2% | 0% | 0% |
| Do hands-on professional work (models, analyses, simulations) | 4% | 0% | 0% | 3% | 0% | 0% | 1% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 4% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 2% | 5% | 22% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 1% | 0% | 6% | 0% | 1% | 0% | 0% | 0% | 0% | 0% | 0% | 3% | 1% | 0% | 3% | 3% | 0% | 0% | 6% | 0% | 5% | 2% | 8% | 1% | 0% |
| Perform physical tasks for robot data | 0% | 0% | 0% | 0% | 2% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 1% | 0% | 0% | 3% | 0% | 0% | 0% | 0% | 4% | 0% | 0% | 1% | 33% | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 0% | 5% | 0% | 0% | 0% | 0% | 0% | 0% | 2% | 0% | 2% | 0% |
Sample. 4,000 roles open on 31 Oct 2026 whose job board publishes a full description (AfterQuery, Aligned Labs, Anthropic, Anyone AI, Appen, Babel Audio, Braintrust, Claru, Cohere, DataAnnotation, DATAmundi, Edison Scientific, Elsevier Health, EvolutionIQ, Fleet AI, Handshake AI, Headway, iMerit, Innodata, Invisible, Inworld AI, JazzX AI, Kastle, Lifted (Upwork), Limbic, Mercor, micro1, Mindrift, Mistral AI, OneForma, OpenTrain AI, Outlier (Scale AI), Pasiflora AI, Perle Labs, Pila8, PointClickCare, Prolific, RealTrainAI, RemoExperts, Remotasks, Rise Data Labs, RWS TrainAI Community, Sama, Scale AI Human Frontier Collective, Shaip (Collaborate by Shaip), Sigma AI, Snorkel AI Expert Community, Sovrano AI, SuperAnnotate (SME), Surge AI, TELUS Digital AI, Terac, TransPerfect DataForce, Trellis, Turing, Vals AI, Vetto, Volga Partners, Weekday, Welo Data and xAI (SpaceXAI)), 91.3% of the 4,381 open roles; the boards that publish descriptions are expert networks, so these roles lean towards expert work. Each cell is the share of the platform's open roles read that mention the task (roles read in the column header).
A role mentions a task when its description contains one of the task's documented patterns, anywhere in the text: a mention is not necessarily the main task, and a role often mentions several.
Listings only: a share of a platform's open roles, not of its hires or projects. Platforms write descriptions to their own template, so a task one platform's template always names counts in every one of its roles.
A finding compares platforms on one task only among those that read on their own for its share: their roles read outweigh the average of the platforms, the weight fitted on the platforms' own counts of that task.
Source. AITraining.jobs Research, What AI Trainers Do, October 2026 edition.
Figure 5
How many different tasks a role describes
Do AI training roles describe one task or a mix?
| Distinct tasks mentioned | Roles | Share of roles read |
|---|---|---|
| No task named | 262 | 6.6% |
| 1 task | 646 | 16.2% |
| 2 tasks | 791 | 19.8% |
| 3 tasks | 598 | 15.0% |
| 4 tasks | 749 | 18.7% |
| 5 tasks | 490 | 12.3% |
| 6 or more tasks | 464 | 11.6% |
Sample. 4,000 roles open on 31 Oct 2026 whose job board publishes a full description (AfterQuery, Aligned Labs, Anthropic, Anyone AI, Appen, Babel Audio, Braintrust, Claru, Cohere, DataAnnotation, DATAmundi, Edison Scientific, Elsevier Health, EvolutionIQ, Fleet AI, Handshake AI, Headway, iMerit, Innodata, Invisible, Inworld AI, JazzX AI, Kastle, Lifted (Upwork), Limbic, Mercor, micro1, Mindrift, Mistral AI, OneForma, OpenTrain AI, Outlier (Scale AI), Pasiflora AI, Perle Labs, Pila8, PointClickCare, Prolific, RealTrainAI, RemoExperts, Remotasks, Rise Data Labs, RWS TrainAI Community, Sama, Scale AI Human Frontier Collective, Shaip (Collaborate by Shaip), Sigma AI, Snorkel AI Expert Community, Sovrano AI, SuperAnnotate (SME), Surge AI, TELUS Digital AI, Terac, TransPerfect DataForce, Trellis, Turing, Vals AI, Vetto, Volga Partners, Weekday, Welo Data and xAI (SpaceXAI)), 91.3% of the 4,381 open roles; the boards that publish descriptions are expert networks, so these roles lean towards expert work.
A role mentions a task when its description contains one of the task's documented patterns, anywhere in the text: a mention is not necessarily the main task, and a role often mentions several.
Out of the 25 tasks in the taxonomy. A role naming none describes its work in words no pattern reads (a project overview, a list of requirements only).
Source. AITraining.jobs Research, What AI Trainers Do, October 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:
11,482 roles tracked, each with its history
Every role on 62 job boards, read since 13 Jul 2026 (every hour since 18 Oct 2026): its pay and place, and the first and last day it was open.
58,632 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.
Building AI and need the experts behind this data? (you contract them yourself) or (we deliver the expert work).
Investing in AI-data companies? Hiring radar is these records role by role, every day, as files.
Method and definitions
The public job boards of the AI training platforms this site tracks, read every hour. Full descriptions are read for the roles whose feed publishes them. 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.
- Mention
- A description mentions a task when it contains one of the task's documented patterns (whole words, case-insensitive, never across two lines), anywhere in the text: the duties, the requirements or a note on the project. A mention is not necessarily the main task, and a role often mentions several.
- Task
- One of 25 activities a description can name, from writing prompts, ideal answers, rubrics and evaluation tasks, through rating, ranking, fact-checking and red-teaming model output, to coding, annotating, transcribing, translating, recording and physical tasks for robot data. Each has a published pattern, tested on real sentences.
- Not a task
- Hiring steps and company boilerplate are kept out: an application read against a rubric, an AI interview, 'no screen recording' or 'professionals, not annotators' do not count as writing rubrics, holding conversations, recording or annotating.
- Pay by task
- The median hourly rate advertised by the roles mentioning a task. A median needs 3 roles stating a rate, and is ranked only when they come from at least 3 platforms; a median resting on fewer platforms is shown, marked, and not ranked.
- Coverage
- Only roles whose job board publishes a full description are read (Mercor, micro1, AfterQuery and Braintrust, about three in ten open roles); each figure states how many that is.
- Open role
- A role listed by a given day and still in its platform's feed that day.
- 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: the last 28 complete days, in UTC. Edition: one calendar month, in UTC.
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. Some platforms keep evergreen roles open for months, and some post one listing per language or country.
- Descriptions are published by about three in ten open roles, all at expert networks: the figures describe those roles, lean towards expert and senior work, and say little about microtask crowds and the platforms that publish titles only.
- A mention is not the main task, and a requirement counts like a duty ('experience in translation' mentions translating). Platforms write to their own templates, so a task one template always names counts in every one of its roles.
- A pattern can miss a task phrased in an unusual way; nothing is inferred beyond what the text says. A pay difference between tasks is an association: tasks travel with kinds of work and seniority.
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
- October 2026 edition (this page)Freezes on first reading
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 AI Trainers Do, October 2026 edition. https://aitraining.jobs/research/what-ai-trainers-do/2026-10
BibTeX
@misc{aitrainingjobs_what_ai_trainers_do_2026_10,
author = {{AITraining.jobs}},
title = {What AI Trainers Do, October 2026 edition},
year = {2026},
howpublished = {\url{https://aitraining.jobs/research/what-ai-trainers-do/2026-10}},
}More research
- Languages in Demand for AI TrainingWhich languages are AI labs hiring for, what do they pay and how much competition is there?Read the report
- Expert Fields in Demand for AI TrainingWhich professions and disciplines are AI labs hiring to train models, and what do they pay?Read the report
- What AI Training Roles RequireWhat do AI training roles ask for: degrees, credentials, experience, hours, tests?Read the report
- AI Training Job TitlesWhat is AI training work called, and does the title say anything about the pay?Read the report
- Programming Languages in AI TrainingWhich programming languages are AI labs hiring coders to train models in, and what does each pay?Read the report
- The Expert Scarcity Index for AI TrainingWhich expertise and languages are hardest to recruit for AI training work right now?Read the report
- Real-World Data Collection WorkWho is hiring people to collect real-world data for robots and physical AI, where, and at what pay?Read the report
- The State of AI Training PayWhat does AI training work pay, by kind of work and by platform?Read the report