- In supervised learning, models learn from labels that people often create, and that's where the demand for annotation work comes from.NIST CSRC glossary, "supervised learning"
- Not every model needs hand-made labels: self-supervised methods infer labels from the data itself, so raw recordings are useful too.NIST CSRC glossary, "self-supervised learning"
- The International Labour Organization (ILO) has studied working conditions on online micro-task platforms, where much annotation work is organized.ILO, Digital labour platforms and the future of work, 2018
- Research on Latin America and the Caribbean describes crowdworkers as "pivotal in labeling data" for global AI systems.arXiv 2501.06981
Annotation and rating work
Annotation is the AI data work most people have heard of, and it's what most job listings mean. You get data that already exists, such as photos, video clips, audio or chatbot replies, and you add what a model should learn from it: what's in the picture, where an object starts and stops, what was said, or which of two answers is better. If you're weighing this against recording, our guide for collectors shows what the recording side looks like in practice.
Annotation exists because many models learn from labels. NIST defines supervised learning as learning to predict "explicit (often human-generated) labels or output values for data." Someone has to make those labels, and very often that's a person at a computer.
Rating and writing work is close to it. You compare model answers, grade them against guidelines or write example answers yourself. Careful reading and clear writing count most, and some projects also want expertise in a subject. Our explainer on what AI training data is shows where labels fit in the bigger picture.
Data collection (recording) work
Data collection creates the raw material itself. For video, you record a real task as you do it: cooking, cleaning, stocking shelves or repairing something. Usually you wear a small camera on your head, so the footage shows your hands and tools from your own point of view. For gaming, you record the game screen together with your keyboard, mouse or controller inputs.
This footage doesn't always need a label to be useful. In self-supervised learning, NIST notes, the true labels are "automatically inferred from the training data". That's one reason large collections of everyday video are valuable.
The skills differ too. You need to do the task well, keep it in frame and follow the project's instructions. For the details, see our guides on how recording work works, how to record first-person video and recording gameplay for AI.
Annotation, rating and recording side by side
| Annotation (labeling) | Rating and writing | Data collection (recording) | |
|---|---|---|---|
| What you do | Tag images, mark objects in video, transcribe audio | Compare, grade or write AI answers | Record a real task from your own point of view |
| Where | At a computer | At a computer | Where the task happens: a home, a workplace or a gaming setup |
| Tools | A computer and the platform's labeling software | A computer and written guidelines | A head-worn camera for physical tasks, or screen and input recording for games |
| Skills that matter | Attention to detail and consistency | Reading, writing, sometimes subject expertise | Doing the task well and following the project's instructions |
| How work is checked | Agreement with guidelines and other labelers | Graded against guidelines | Footage reviewed against the project's requirements |
The table shows the usual shape of each kind of work. The details vary from one company and project to the next.
What research says about data work
Researchers have looked closely at data work and the people who do it. The ILO's 2018 report Digital labour platforms and the future of work "offers one of the first comparative studies of working conditions at five of the major, global, online micro-task platforms." Its 2021 flagship report on the role of digital labour platforms looks at how these platforms affect "enterprises, workers and society as a whole."
Academic work points the same way. The authors of Making Data: The Work Behind Artificial Intelligence write that the development of smart technologies "depends, at different stages, on a multitude of precarious, underpaid and invisible workers." A study of data enrichment work in Latin America and the Caribbean adds that "the global AI surge demands crowdworkers from diverse languages and cultures."
For you, that means reading the terms before you start. How you're paid, when, and for what should all be in writing.
How to choose the right kind of work
Before you apply anywhere, think through these points:
- Think about where you want to work. Annotation and rating happen at a computer; recording happens where the task is, at home, at your workplace or at your gaming setup.
- Start from what you already know well. A careful reader or writer may prefer rating, while a line cook, cleaner or mechanic already knows tasks worth recording.
- Check that you can get permission. Recording other people or a workplace needs releases and the site's permission; see what gets footage accepted.
- Look for written terms. Pay, timing and requirements should be visible before you commit.
- Notice whether anyone asks you for money. Real data work never charges you to start.
Common mistakes when picking AI data work
- Every "AI training job" looks alike from the title. The label covers labeling, rating and recording, and each suits different people.
- Don't pay to start. Fees for training, starter kits or "unlocking" tasks are warning signs; learn how to spot AI training job scams.
- You don't need a tech background for recording work. It depends on knowing an everyday task well, and you don't have to know how AI is built.
- Don't ignore privacy. Recording people without their release can get footage rejected and puts others at risk.
How TaskCapture fits in
Our collectors record video; they don't rate chats or write answers, and we don't hire people to drive robots. We collect three types of data: residential, commercial and gaming. Gaming data is the screen plus every keyboard, mouse or controller input in sync, with no camera, voice, chat or gamer handles. For physical work, you wear a small camera on a head strap or cap, so your hands stay free.
There is no fee to apply, and no coding or AI background is needed. You never pay for equipment: the project provides it, or the task needs no special equipment. Pay ranges from USD 5–15 per hour of accepted footage, depending on the project; it varies by project and is not guaranteed. You'd work as an independent contractor under the Data Collector Agreement, and the Privacy Policy sets out the consent rules.
Questions
What is data annotation work?
It's adding labels to existing data so an AI model can learn from it. That can mean tagging images, marking objects in video, transcribing audio or rating answers, usually at a computer.
Is data collection the same as data annotation?
No. Annotation labels data that already exists. Collection creates new data, for example by recording a real task from your own point of view or recording gameplay with its inputs.
Which kind of AI data work needs the least technical knowledge?
Requirements vary by project. Some annotation and rating work needs no coding, while specialist projects can require it. Annotation and rating ask for care and consistency at a computer. Recording asks you to know an everyday task well and follow the project's instructions.
Can I do both?
Yes. They're separate kinds of work, so read each company's terms on their own, and never pay to start either one.