AI Training Jobs Weekly

1. 9 Nov 2026 · 4:52

   How reviewers grade your work, and what gets a task rejected

   On most AI training platforms you are paid for the work a reviewer approves, not the hours you log. What the platforms' own rules say about approval, idle time, originality and good faith, and how a rejection shapes the work you are offered next.

   Transcript

   On most AI training platforms, the hour you work is not the hour you are paid for. You are paid for the work a reviewer approves. This week's lesson is how that review works, and what gets a task rejected.

   Lesson one: approval is the unit of pay. AfterQuery's help center says pay is based on approved work. micro1 says hourly work is paid on the hours the project team approves, and that on flat-fee work no payment is made for incomplete, rejected or unapproved submissions. DataAnnotation's terms say rejected work is not paid. Sovrano AI pays monthly, for approved work only. So a task is not done when you submit it. It is done when someone signs it off. Count your earnings the same way: the hours that were approved, not the hours you logged, and expect the two to differ in your first weeks.

   Lesson two: the clock is watching. On tracked contracts, Sovrano AI counts time as idle when its tracker records no activity for 90 seconds, and idle time is not paid. SME Careers, part of SuperAnnotate, pays for active time only. Its tracker can take screenshots and screen recordings while the timer runs, it switches to idle after 2 minutes without keyboard or mouse activity, and time spent in AI tools your project instructions do not allow is not paid. Learn your platform's tracking rule before you start the timer, not after your first payout.

   Lesson three: originality is checked. Handshake's help center says a prompt taken from a textbook does not count as original work, and asks you to keep your drafts for 90 days. Mindrift says you may not directly copy content from other sources, including automated tools. And one Outlier worker wrote on Trustpilot that an automated check flagged their typing as AI and froze 320 dollars of earned pay. That is one person's account, not a finding, but the lesson holds: write your own work, and keep your drafts, because they are your evidence if a review goes against you.

   Lesson four: complete, and in good faith. Mercor pays its work trials whether or not you are selected, but only for a complete, good-faith submission, and it says it can withhold payment for low-effort or incomplete work. This is how the labs grade their own models too. OpenAI had 262 doctors write 48,562 criteria for grading ChatGPT's medical answers, and each answer is checked against them one by one. Your task will be read the same way. Treat the instructions as the rubric: before you submit, check every requirement against your answer, line by line. In practice that means the small things: the format the task asks for, the length, the sources it requires, the tone, and every part of a question with several parts. A reviewer with a rubric gives no credit for what you meant.

   Lesson five: feedback is rare, so your record speaks for you. Mercor says it cannot share specific feedback on a rejection because of the volume of applications. DataAnnotation says the work available to you depends on the demand for your skills and on your past performance on the platform. So a rejected task can cost you twice: once in pay, and again in the work you are offered next. Find your platform's dispute window before you need it. Sovrano AI, for example, gives you 14 days to dispute idle time by replying with evidence.

   Lesson six: the cap on the clock. Handshake AI pays hourly work up to each task's time cap, so a task that takes you twice its cap is paid as one cap. Turing's rate card pays some experts per approved task, between 150 and 300 dollars. Before you take a task, read its cap or its unit of pay, and decide whether you can finish it inside that.

   To sum up. You are paid for approved work, not logged time. Know your tracker's idle rule and your task's time cap. Write your own work and keep your drafts. Treat the instructions as the rubric and submit complete work. And protect your record, because it shapes the work you are offered next. The platform rules in this episode come from their own help pages and terms, read between the 1st and the 8th of November 2026. A posted rate is an advertised rate, not a guaranteed wage. Next week: how to tell a real AI training project from a scam. This was AI Training Jobs Weekly.

2. 2 Nov 2026 · 5:27

   Why some accepted experts never get a project, and what changes that

   On 19 Oct 2026 one AI training platform picked 75 of the 401 people who applied through our job board that day; of the 998 who applied to the same roles over the next 12 days, it picked none. Why acceptance puts you in a pool rather than on a project, why projects take people in windows, why the first two days after a hire matter, and why experts keep more than one platform.

   Transcript

   On the 19th of October, one large AI training platform picked 75 of the 401 people who applied through our job board that day. The day before, it picked 1 of 134. The day after, 1 of 244. And of the 998 people who applied to those same roles over the next 12 days, it picked none. The postings had been up for weeks, some for months, and they stayed up the whole time. This is AI Training Jobs Weekly. Today, as promised last week: why some accepted experts never get a project, and what changes that.

   Lesson one: being accepted puts you in a pool, not on a project. The platforms say so in their own words. micro1 says being certified means you are part of its certified talent pool and eligible for opportunities. Alignerr says you might be on a generalized waiting list until a project that fits your profile starts. Handshake AI says projects are not self-serve and are not available immediately after applying. And Mercor adds one that most people miss: contributors are not automatically moved to a new project after one ends. Our own records show how many steps sit between a yes and a paycheck. Of the 113 people who applied through our board and were placed on a project at one of these platforms by the 1st of November, 6 had earned anything so far. Most were placed in the last two weeks, and first payments often wait for a weekly approval, so part of that gap is simply time. But it shows where the steps are: accepted, then matched to a project, then actually working, then paid. Each one is a separate step, and each one loses people.

   Lesson two: projects take people in windows, not in a queue. That is what the 19th of October shows. The roles people applied to were the same on every one of those days, and so was the mix of how old those postings were. What differed was the day. A platform can keep a posting up for months, while the project behind it takes its whole intake in a day or two, from the people who are eligible and active right then. That platform had not closed a single role since September. So the age of a posting tells you almost nothing about whether anyone is being picked from it. What changes your odds is being ready before a window opens. Finish the assessments, the identity and tax checks and your availability now, not when an invitation arrives. Mercor puts it most directly: a regularly updated, assessment-complete profile with clear availability is the single most significant driver of the offers it sends. And on Mercor, once you pass an assessment, you are considered for new roles that use it, even if you never apply to them. And watch for new programs. In the week to the 1st of November our hourly detection caught 21, from AfterQuery's first agentic and research coding roles to new transcription and audio work at micro1.

   Lesson three: the first two days after you are hired decide a lot. At another platform, of 12 people hired through our board, 2 have worked so far, and both started within 2 days of being hired. The other 10 show no work logged yet. Those are small numbers, but the shape is clear: work tends to start at once or to stall. The platforms reward the same thing. Mercor asks people to accept or decline offers before they expire and to reply to follow-ups promptly. Handshake AI warns that fellows who stay inactive while tasks are available may have their project access reassigned. So when you are hired, do the onboarding, the paperwork and the first task that same day.

   Lesson four: an hourly rate is not a promise of hours. The Wall Street Journal profiled a Princeton chemical engineering PhD who had work through Mercor at $70 an hour. His hours ran from 4 to 20 a week, and then the work dried up. In the open roles we read on the 1st of November, 53.9% of those that state their weekly hours ask for under 20. Plan for contract work that comes in bursts, not for a salary.

   Lesson five: keep more than one door open. Of the 12,840 people looking for this work on our board, 22% already work on at least one AI training platform, and 11% on two or more. And of the open roles that state a type, 65.5% are contract work and 3.1% are full-time. The people already doing this work treat it as a portfolio, not a job. If you hold a credential in law, medicine, finance or a science, two or three platforms in your field will keep you in more windows than one.

   To sum up. Accepted is a pool, not a project. Projects take people in windows, so be fully set up before one opens. Start the day you are hired. Expect hours in bursts. And keep more than one platform in your field. The figures in this episode come from our job board's own records and the public job postings we read, as of the 1st and 2nd of November 2026, and a posted rate is an advertised rate, not a guaranteed wage. Next week: how reviewers grade your work, and what gets a task rejected. This was AI Training Jobs Weekly.

3. 26 Oct 2026 · 6:52

   What the AI labs hire for inside their own teams

   Of 3,515 open jobs at 25 AI labs on 26 Oct 2026, 90 were on the teams that buy, run and evaluate human data. What those teams do, why the instructions you work from are written there, why a lab name on a posting is no pay premium, and how contract work leads to the jobs inside.

   Transcript

   On the 26th of October, the 25 AI labs whose public job boards we read every day had 3,515 jobs open. Just 90 of them were on the teams that buy, run and evaluate human data, the teams behind the work experts are paid to do. That is 2.6%. The night before, the AI training platforms we follow listed 4,336 open roles. The labs keep only a small part of this work in house, and they buy most of it. This is AI Training Jobs Weekly. Today, as promised last week: what the AI labs hire for inside their own teams, and what that means for the work you do for them.

   First, what those 90 jobs are. The largest group is evaluations: 22 roles at 10 labs, people who build the tests that measure what a model can and cannot do. Next is human data operations: 20 roles at 8 labs, the people who run data programs. Then AI tutors and raters, 17 roles at only 4 labs: Cohere, Mistral AI, Skild AI and xAI. 14 roles collect robot data, at 4 robotics companies. The rest buy or license data, build training environments, or research what makes data good. So the hands-on expert work itself is a small part of what the labs hire for. Most of it is bought: OpenAI's procurement team lists human data services among the professional services it buys.

   Lesson one: the instructions you work from start at the lab. OpenAI's own posting for a data program manager on its human data team says it plainly. The job is to be a key interface between OpenAI's external vendors and the AI trainers: to gather requirements, write instructions, define success criteria, calibrate the trainers, and assess the labeled data and give them feedback. It pays $207,000 to $230,000 a year in San Francisco. So on a project for a lab like OpenAI, the platform is in the middle, but the guidelines, the examples and the bar come from people like this. Treat the guidelines as the client's specification. Expect them to change: xAI's own tutor posting lists working from evolving instructions as part of the job. And calibration rounds are where you are judged, so aim to match the rubric every time, not just to be brilliant once.

   Lesson two: evaluations are what the labs hire for most. 22 of the 90 roles build or run them. Anthropic's posting for a data operations manager on its human data team includes building evaluation frameworks and quality standards that ensure data meets the bar for training. The grading you do is the raw material of that work. A rubric is a measuring instrument, and a measurement is only useful if two experts grading the same answer would reach the same verdict. So write every score as if a second expert will check it against the same rubric: name the criterion, point to the part of the answer, and say why. A judgment another expert can retrace is one an evaluation can use.

   Lesson three: working directly for a lab is not automatically the better deal. A few labs do hire tutors themselves. xAI's posting for an AI tutor in legal and compliance asks for a law degree and at least 2 years of practice, and pays US based candidates $50 to $100 an hour. For contractors, it says hours vary widely, that most projects may take at least 10 hours a week on average, and that it cannot hire in Wyoming or Illinois. Across the labs, the 22 contract roles that state an hourly rate pay $18 to $100 an hour. Now compare the platforms. On the 22nd of October, the first 13 roles we read on Surge AI's own job board had a typical rate of $300 an hour, and one of them, for constitutional lawyers, advertised $200 to $400. On the 19th, Mercor opened 13 roles in medicine and nursing at a typical $200 an hour. A lab's name on a posting is not a premium. Compare the rate, the hours and the terms.

   Lesson four: if you want a career on the lab side, the door is operations, not more tasks. 37 of the staff roles state a yearly salary. The middle of the typical range is $284,000, and the ranges run from $100,000 to $850,000, with research engineers at the top. The jobs closest to what an expert contractor knows are the program and operations roles. Anthropic's data operations manager pays $270,000 to $365,000 a year. It asks for 3 or more years in operations, consulting, product or program management, and it lists as a plus experience with data collection, labeling or annotation operations, and with managing vendors or external contractors. So your contract work counts, if you describe it the way these postings do: the projects you calibrated on, the quality bar you met, the guidelines you helped improve. Most of these jobs are based in cities like San Francisco, New York, London and Paris.

   Lesson five, briefly: the labs also buy expertise in bulk. 6 roles at Mistral AI, OpenAI, Reflection AI and Thinking Machines Lab buy or license data. OpenAI's posting for a science data program manager, from the 13th of October, says its current work focuses on biology and chemistry, and ranges from acquiring existing datasets to shaping research collaborations. So if you work at a hospital, a research lab or a firm with rare records, your field may reach these models through a deal your employer signs, not only through tasks you do.

   To sum up. The labs buy most of their human data work, so a platform is usually your way in. The instructions come from the lab and calibration is where you are judged, so be consistent. Grade in a way another expert could repeat. A lab's name is not a premium. And the inside jobs are in operations, where your contract work counts. The lab figures in this episode come from the public job boards of 25 AI labs, read on the 26th of October 2026, and a posted rate is an advertised rate, not a guaranteed wage. Next week: why some accepted experts never get a project, and what changes that. This was AI Training Jobs Weekly.

4. 18 Oct 2026 · 5:38

   What experts should know before applying for AI training work

   Why AI companies now pay professionals to train their models, what the work actually is, and five things an expert should know first: the platform sets the pay, a posted range is decided after you qualify, the first interviewer is software, timing beats polish, and scarce credentials have leverage. Plus the one scam to recognise.

   Transcript

   A medical expert can find the same kind of AI training work advertised at $135 an hour on one platform and $40 on another. Same field. Same kind of task. If you are a professional thinking about this work, that gap is the first thing worth understanding, and it is not the last. This is AI Training Jobs Weekly. Today: what an expert should know about this market before applying anywhere.

   First, why experts at all. A few years ago most of this work was general: label an image, pick the better of two answers. Models now handle much of that themselves. What AI companies pay for now is judgment they cannot get cheaply: a doctor who can see that an answer is confidently wrong, a lawyer who can tell a weak argument from a strong one, an engineer who can break a solution. When OpenAI worked on how ChatGPT handles conversations about mental health, it said it worked with more than 170 mental health experts. In the public job postings we track, more than 4 in 10 open AI training roles now name a profession in their title, and in law, medicine and finance work it is nearly 3 in 4.

   So what does the work look like? Mostly three things. You write hard questions in your field that the model gets wrong. You write the answer an expert would give, as a reference. And you grade the model's answers against a rubric: is it accurate, is it safe, would a professional sign it. One current posting for licensed psychologists puts it plainly: review how AI models respond to clinical psychology scenarios, and judge whether they got the clinical reasoning, ethics and risk handling right. That one is a 3 week project. Short projects are normal in this work.

   Lesson one: the platform decides your pay as much as your field does. In mid October, medicine and nursing roles had a median advertised rate of $135 an hour at Mercor and $40 at Alignerr. Law roles, $133 at micro1 and $60 at SuperAnnotate. Software roles, $98 at Handshake AI and $31 at OpenTrain AI. Looking only at roles open to US residents, the best paying platform in a typical kind of work advertises about twice what the worst one does. So before you apply, compare what each platform pays for your field, not what the industry pays on average.

   Lesson two: the number on a posting is a range, and where you land in it is decided after you qualify. A psychologist posting at micro1 reads $100 to $200 an hour. Which end you get depends on your credentials and how you score in their assessment. And an hourly rate is not a promise of hours. Projects open, fill and close. Some weeks there is plenty of work and some weeks there is none, and some platforms pay per task instead of per hour.

   Lesson three: your first interviewer is probably software. At micro1, the first round is a conversation with its AI recruiter, called Zara: open questions about your field, and scenario questions about how you would handle a case. You get a result right away on whether you meet their certification bar. Mercor also screens experts with an AI interview. Treat it like a real interview. Speak in specifics, reason out loud, and use the vocabulary of your field. And know that passing gets you certified, not booked. The work starts when a project needs someone with your background, and that can take days or weeks.

   Lesson four: timing beats polish. Applicants pile into new listings: a role in its first week draws more than 6 times the interest of one that has been up for more than three months. The typical listing in our data is up for about 18 days, and projects fill. Apply in the first days a role appears, and keep your profile and credentials ready so you can.

   Lesson five: the crowd is not where the money is. Writing and editing roles draw two and a half times the applicant interest of the average role, and they pay the least of the expert fields, a median of $48 an hour. Law pays the most, $109 an hour, and draws barely more than half the average interest per opening. Chemistry roles draw about a tenth of it, and they stay open longer. If you hold a scarce credential, you have leverage: you can be choosy about platforms and projects.

   And one warning. Real platforms pay you. They never ask you to pay them. A fast growing scam, which the US Federal Trade Commission calls a task scam, pays you small amounts at first to earn your trust, then asks for a deposit to unlock your earnings. The FTC counted about 20,000 reports in the first half of 2024, against about 5,000 in all of 2023. Its rule is simple: never pay anyone to get paid.

   To sum up. Compare platforms for your own field. Read a posted range as a range. Prepare for the AI interview like a real one. Apply in the first days. And use your scarcity. The figures in this episode come from public job postings as of the 17th of October 2026, and a posted rate is an advertised rate, not a guaranteed wage. Next week: what the AI labs themselves are hiring for, inside their own teams. This was AI Training Jobs Weekly.

The same week in writing is in the weekly news, and every open role with its pay is on the job list.