Updated 2 November 2026
AI training work is evaluation. Whatever the listing calls it, you judge something an AI produced against rules someone else wrote, and your judgment becomes training data. The names change: grading answers, ranking versions, rating responses, annotating video, transcribing audio. The core stays the same. I have done this work on Mercor since August 2026, grading whether AI-generated writing sounds human. I have also been under contract with micro1 since September 2026.
This page shows what the tasks look like from the inside, how the platforms judge your work, and who tends to enjoy it. The platform facts come from Mercor’s, micro1’s and AfterQuery’s own pages, as they read on 2 November 2026.
The one question behind every task
On my Mercor project, every task comes down to one question: does this sound like a person, or like a machine? I grade AI-generated writing phrase by phrase, against the client’s own rules. The project has about 45,000 people on it, and I sit in the top 1% of performers.
The formats change more than the question. I started on slide decks. Then came long documents, chatbot answers, spreadsheet labels, web page text and descriptions of software changes. Each format hides machine-sounding writing in a different place, so each one trains your eye a little more.
Other projects ask other questions. Mercor’s own guide to the work lists the usual ones: is this answer correct, does it actually answer the question, is anything misleading or unsafe, and would someone trust it. Every project picks its own question. Your job is to answer it the same way, task after task. That consistency is the whole job: the same question, asked the same way, thousands of times.
The main kinds of task
Grading means giving a verdict on a piece of AI output against a set of rules. My own work is grading: I decide, phrase by phrase, if the writing passes as human.
Comparing and ranking means looking at two or more AI answers to the same request and picking the better one, then explaining why. Mercor’s guide explains why this matters: when trainers pick the better of two answers, the model learns what people actually prefer, not only what is technically correct. DataAnnotation and Outlier both list this kind of work on their own pages.
Writing responses means producing the answer yourself, so the model can learn from a good example. micro1 describes this as writing responses that show reasoning and problem-solving. Some projects ask you to write the prompts too.
Scoring against a rubric means rating an answer criterion by criterion, then writing a short rationale for each score. This is the core of micro1’s Educator/Assessor Writer role, which I describe further down.
Annotating means labelling data so a model can learn what it contains. AfterQuery’s help pages list the classic forms: sorting text into categories, marking the emotion in a sentence, tagging names of people, places and organisations, and labelling what appears in images. Video annotation does the same with moving images.
Transcribing means turning speech into exact text. Recording is newer: micro1 runs projects where contributors film everyday physical tasks, so robots can learn from real movements. Coding work exists too, where engineers check code that an AI wrote.
The life of a single task
micro1’s own pages describe the path of a task, and it matches what I see elsewhere. You work on a dedicated platform, and each task comes with its own instructions. On hourly contracts, Hubstaff tracks your time while you work. When you submit, the task goes through quality review before it counts as approved.
Then the feedback comes back. Reviewers point out what to fix, the pipeline instructions get updates, and you check them before your next task. Questions go to your team’s Slack channel. Over weeks, your performance record builds up task by task.
What RLHF work is like
RLHF stands for reinforcement learning from human feedback. In plain words, people compare AI answers, choose the better one and explain the choice, and the model learns from thousands of those choices. Most of the comparing, ranking and grading on these platforms feeds that process.
In practice, it feels like close reading under rules. Mercor’s guide describes the rhythm well: you review the rubric, complete tasks, write down your reasoning, pass quality checks, and adjust when the guidelines shift. Tasks come through a queue. You make many small decisions in a row, and each one has to match the rules, not your taste.
Audio and video work, and why I dropped video
I tried one video-annotation job on Mercor, and I hated it. Grading writing fits me far better. That is not a judgment on video work. It is a reminder that not every kind of AI work suits everyone. The label on a listing does not tell you how the hours will feel.
Transcription has its own arithmetic. One hour of audio takes two to two and a half hours of transcription work, because you play, pause, type and check every line. Keep that ratio in mind when a listing pays per audio hour. AfterQuery’s help pages list audio transcription among the annotation work that trains speech models.
In my experience, content grading is the best paid of these tasks, and the calmest. It asks for a good ear for language, steady attention and the discipline to apply someone else’s rules.
The assessor-writer kind of task
micro1’s Educator/Assessor Writer role shows what expert evaluation looks like. You compare AI-written responses for clarity, tone, helpfulness and instruction adherence. You score them against rubrics. You write rationales that explain each score. You also flag answers that sound fluent but have nothing to support them.
The interview for that role tested technical writing, editing and proofreading, analytical judgment, and writing rationales under constraint. The listing showed $90 to $140 an hour in October 2026. I reached the next step in that process, and the role matches what I already do on Mercor, from a different angle: there I judge the voice, here the substance.
If you teach, edit, translate or review other people’s work for a living, this is the kind of task to look for. Your profession already trained the habits it needs.
The skills each kind of task asks for
Grading asks for an ear and for discipline. You need to hear what sounds wrong, then name it with the project’s rules rather than your taste. Comparing and ranking asks for judgment you can explain: the better answer is not enough, you have to say why in a few clear lines.
Rubric scoring asks for rationale writing. micro1 tested exactly that in its assessor-writer interview, along with editing, proofreading and analytical judgment. Annotation asks for precision and patience, the same label on the same kind of thing, every time. Transcription asks for a careful ear and accurate typing, over long stretches.
Expert evaluation asks for your profession. Mercor’s guide gives the examples. Doctors judge clinical reasoning in medical answers, and lawyers check contract analysis and the citations in legal research. Engineers judge the quality and security of code. The platform supplies the rules. You supply the knowledge that tells a right answer from a confident wrong one.
How the platforms judge your work
On my Mercor project, the project scores my verdicts against an expected answer I never see. A mismatch counts as a mistake, and mistakes can get you offboarded. The client also reviews submitted tasks and sends feedback. And the client’s rules changed during the project, so what counted as right in one month could count as wrong in the next.
micro1 describes its own quality system in layers. You create the task. Peer reviewers check it for accuracy and guideline adherence. The operations team may run further checks, and in some pipelines the client gives the final approval. Feedback reaches you through Slack messages, comments on specific tasks and guidance in team channels.
micro1 also tracks quality scores, time per task and guideline adherence. If your performance needs to improve, your pipeline lead contacts you with specific guidance. Mercor’s guide describes the same principle: a reviewer compares your work to the project standard and settles any differences.
The practical lesson is simple. Read the guidelines again whenever they change, and treat every update as the new rule from that moment on.
What a working week looks like
Before any paid task, there is training. Mine was a long course covering nine task types, mostly around slide decks, and it was my only source of knowledge for the job. micro1 works the same way: its pipeline instructions are your source of truth, and some pipelines add training videos, sample tasks and quizzes before you can start.
Then the work arrives through a queue. Mine is spiky, and the project itself warns that it is intermittent. In September I called it inconsistent and exhausting. Some weeks are full and some are empty. These platforms are not jobs. They are queues of paid tasks, and they cannot give anyone a fixed monthly income.
Tools shape the week too. On micro1 you get a Slack channel for your team, an expert email account and access to the work platform. Hourly contracts add Hubstaff for time tracking. On Mercor, the Insightful tracker runs while you work. On both, the questions go to your team channel or your project lead, not to the wider internet.
How the work changes as you go
The work grows with you. On my project, the formats widened from slide decks to documents, chatbot answers, spreadsheets, web pages and software notes. Each new format brought new rules to learn on top of the old ones.
Mercor’s guide describes the wider path. It goes from labelling to evaluation, where you compare answers and back up your choice. Then it goes to rubric design, where you help define what a good answer is. micro1’s assessor-writer role sits in the middle of that path. It asks you to judge and to explain, with rubrics someone else wrote.
Who enjoys this work, and who hates it
People who enjoy it tend to like close reading and small decisions. They apply a rule the same way on the hundredth task as on the first. They can explain a judgment in two clear sentences. micro1’s own description of its best experts matches that: they take ownership, follow guidelines precisely, catch small details, write clear questions and keep learning from feedback.
People who hate it usually hit one of three walls. Some need a fixed income, and the queue cannot give it. Others find repetition draining, or resent rules that change mid-project. And some land on the wrong kind of task, the way I did with video. If you are unsure, start with the kind of task closest to what you already do every day.
The work also takes time from the rest of your life. I dropped more than half of my teaching hours for Mercor, including a full day a week. The money is real. So is the cost.
Do you need to code, or to know AI?
No, for most of this work. micro1 says plainly that you need neither programming nor AI experience, only your expertise, attention to detail and a willingness to learn. Mercor’s guide says the same about most AI trainer roles: writing skills, logical reasoning, domain knowledge and consistency matter more than code. I do not code, and I grade AI writing every week.
Coding matters only for coding projects, where engineers review code that an AI wrote. AfterQuery hires software engineers for that kind of work, and DataAnnotation runs a separate coding track. Everywhere else, your profession is the skill: a doctor judges medical answers, a lawyer judges legal ones, and a language teacher hears what sounds off in a sentence.
Mercor
For experts with a degree or real experience in their field who can hold a 15 to 20 minute AI interview in English. Most Mercor listings in October 2026 paid $50 an hour and up, with weekly pay.
[Nada’s photo beside the referral lines]
micro1
For people with a degree or real experience in their field who can hold a 20 to 40 minute interview in English. Each role shows its own pay.
This is my referral link, so please open it before you sign up. micro1 pays me a referral bonus, and nothing comes out of your pay.
AfterQuery
For professionals in fields such as law, medicine, finance, software or research, who live in a country AfterQuery can pay. Job boards listed its roles at typically $50 to $125 an hour in October 2026.
This is my referral link. AfterQuery pays me a referral bonus, and nothing comes out of your pay.
Frequently asked questions
What is AI training work?
Paid evaluation of what an AI produces. You grade, compare, rank, score, annotate or transcribe, against rules the project sets, and your judgments become training data for the model.
What does an AI data trainer do day to day?
They read the project guidelines and take tasks from a queue. For each task they give a verdict or write a response, then explain their reasoning. When the guidelines change, they adjust. Reviewers then check the work against the project standard.
What is a content grading job?
A task where you judge AI-generated content against a set of rules. In my case, I grade whether AI-generated writing sounds human, phrase by phrase, across slide decks, documents, chatbot answers and other formats.
What is RLHF work like?
Close reading under rules. You compare AI answers, pick the better one and explain why, many times in a row. The model learns from those choices, so consistency matters more than speed.
Is AI training work hard?
The tasks are rarely hard one by one. The difficulty is consistency: the same rule, applied the same way, across hundreds of tasks, while the guidelines keep evolving and a hidden expected answer scores you.
How long does transcription take?
In my experience, one hour of audio takes two to two and a half hours of transcription work.
Do I need to be an expert?
Not for every task. Generalist and bilingual roles exist, such as micro1’s at $30 to $65 an hour in October 2026, and DataAnnotation’s general track. Expert roles pay more because they need your profession.
Can I do this work in another language?
Yes. My third micro1 interview was for a bilingual French expert role. The first part ran in English, and the second in French, with Zara correcting me. Bilingual and multilingual work also appears on DataAnnotation and Alignerr.
Can I choose the kind of task?
Yes, through the listings you apply to. Each listing names its task, so read it before you apply. I tried video annotation once, then stayed with grading writing.
What equipment do I need?
Your own laptop and a good internet connection. Mercor also asks for a headset with a separate microphone, and some of its roles require an M-series MacBook. The time trackers, Insightful on Mercor and Hubstaff on micro1, run on your own computer.
Is it a full-time job?
No. The platforms are queues of paid tasks. Some weeks are full and some are empty, so treat the income as variable.
Which kind of task pays best?
In my experience, content grading pays best of the tasks I have seen, and it is the calmest. Expert evaluation roles list higher still: micro1’s Educator/Assessor Writer role listed $90 to $140 an hour in October 2026.