If you have seen an advert for an "AI trainer", "data annotator", "AI tutor" or "language specialist" and could not tell from the description what you would actually be doing all day, that is not your fault. The job titles in this industry are close to meaningless, and they change faster than the work does. This lesson gives you the thing the adverts leave out: what the work is, why it exists, and who pays for it.
Why a model needs people at all
A large language model is built in two broad stages, and only the second one employs you.
The first stage is pretraining. The model is shown an enormous volume of text and learns to predict what comes next. This is where it picks up grammar, facts, style, and a great deal of the internet's worst habits. Pretraining is expensive, automated, and involves no annotators. Nothing you do touches it.
What comes out of pretraining is not a useful assistant. It is a very good text predictor that will happily continue a sentence in any direction, including confidently wrong, unsafe, or simply unhelpful ones. It has no notion that a question is meant to be answered rather than extended, no sense of which of two plausible answers is better, and no reason to refuse anything.
The second stage, usually called post-training, is where that gets fixed, and it runs on human judgement. Someone has to show the model what a good answer looks like. Someone has to decide, between two answers that both look fine, which one is actually better and why. Someone has to try to make the model behave badly so the failure can be found before a customer finds it. There is no automated substitute for this, because the whole point is to inject a preference that the model does not have and cannot derive on its own.
That is your job. You are not "teaching a robot" in any mystical sense. You are supplying examples of judgement, at scale, in a format a training pipeline can consume.
The four families of task
Almost every piece of AI training work you will ever be offered is a variation on one of four things. Learning to recognise which one you are looking at is the single most useful skill in this lesson, because the definition of "good work" is completely different in each.
1. Demonstration. You are given a prompt and you write the ideal response yourself. The model later learns to imitate it. This is the most writing-heavy family and often the best paid, because it needs genuine subject knowledge. A request to "write the response a skilled tax accountant would give to this question" is a demonstration task.
2. Preference comparison. You are shown the same prompt with two or more model responses and asked which is better, usually with a written justification. This is the backbone of what is often called RLHF, reinforcement learning from human feedback. You are not writing content; you are ranking it, and your reasoning matters as much as your choice.
3. Annotation and classification. You apply labels to content against a defined scheme. Is this message spam. Does this response contain medical advice. Which of these nine categories does this query belong to. Individually simple, and deceptively hard to do consistently across hundreds of items.
4. Evaluation and red-teaming. You score a single response against a rubric, or you deliberately try to provoke a failure: get the model to contradict itself, produce something unsafe, or break a rule it was told to follow. Red-teaming is the one family where "making it go wrong" is success.
A great deal of confusion comes from platforms using their own names for these. One company's "SFT project" and another's "response authoring" are both demonstration work. "Side-by-side", "pairwise" and "A/B ranking" are all preference comparison. Learn the four shapes and the branding stops mattering.
Who is actually paying you
It is worth understanding the money's path, because it explains almost everything that frustrates people about this industry.
At the top are a small number of AI labs building models. They need human data in large volumes, in bursts, with specialist skills they do not employ in-house. Hiring thousands of people directly, in dozens of countries, for work that might last six weeks, is not something a research lab wants to do.
So they buy it from data vendors, companies whose entire business is delivering labelled data to a specification. The vendors in turn recruit a large, flexible pool of contributors, which is you, usually as an independent contractor rather than an employee.
Three consequences follow directly from that structure, and none of them are personal:
- Work is project-shaped, not job-shaped. A project exists because a client bought it. When the client's data need is met, the project ends, often abruptly and often with no warning. A queue going empty is usually a finished project, not a judgement on you.
- The rules come from someone you will never speak to. The rubric you are working to was written to a client's specification. This is why guidelines sometimes seem arbitrary, and why "but my way is better" is not a winning argument.
- Quality is measured obsessively. The vendor's product is your output. If it fails the client's quality bar, the vendor loses the contract. That is why the scoring can feel harsh: your work is the thing being sold.
What the work actually feels like
Two honest expectations, because a lot of recruitment material is vague about both.
Most of this work is reading. Even on writing-heavy projects, the majority of your time goes on carefully reading a prompt, a rubric, and one or more responses before you produce anything. People who are fast readers with high attention to detail do well. People who skim do badly, and the quality systems in Lesson 3 are specifically built to catch skimming.
And it is genuinely cognitively demanding in short bursts. It is not mindless clicking, whatever the phrase "data labelling" suggests. A preference comparison on a technical topic can require you to verify a factual claim, notice a subtle instruction the response ignored, and articulate the difference in three sentences. That is real work, which is precisely why it is paid rather than automated.
What to take from this lesson
- Your work exists because post-training needs human judgement that cannot be automated.
- Nearly every task is demonstration, preference comparison, annotation, or evaluation and red-teaming. Identify which before you start.
- You are usually a contractor to a vendor, not an employee of a lab. Project-shaped work, externally written rules, and heavy quality measurement all follow from that.
- The work rewards careful reading far more than fast clicking.