Basics

What is data labeling — and why labelers disagree with each other

Illustration: two hands sorting the same stack of cards differently

Here's the fact that reframes how model training works.

For the paper that modern chat models grew out of, OpenAI hired about 40 people — screened, trained, working from a written guideline. Their job: look at two model answers and say which one is better.

And here's the number: those people agreed with each other 72.6% of the time. Roughly every fourth time, two trained labelers picked different answers as the better one.

The model learned from that. So its "right answer" isn't a textbook fact — it's the averaged opinion of people who don't fully agree themselves.

What data labeling is

Labeling means attaching the right answer to your data. On their own, images, texts and audio tell a model nothing: a dataset only becomes a textbook once each example says what's in it.

A simple example. You have 10,000 app reviews. Right now that's just text. Labeling is when each one gets a tag: "payment complaint", "bug", "feature request". Now you can train on it.

Why you should care if you don't train models: the same work is what it takes to evaluate an existing one. More on that below.

The three forms you'll meet

  • A label for the whole example — one tag per item: spam or not, language of a text, topic of an email.
  • Marking a span — draw a box around the cat, highlight the name and address in the text. Pricier: the labeler annotates every fragment, not the example.
  • Comparing two answers — "which is better?". That's exactly how RLHF data is gathered, and where the number at the top came from.

The third form looks easier than the first two and is in fact harder: "better" has no objective definition.

Why labelers disagree

The reasons are boring and very human.

  • The task is fuzzy. Is a "helpful" answer short or thorough? Is a polite refusal better than risky advice, or worse?
  • The guideline misses the edges. "Paid, then the app crashed" — payment complaint or bug? Both labelers are right.
  • People get tired. Attention on example 300 isn't what it was on example 10.

That same paper has a second number: labelers who weren't in the training group agreed with each other 77.3% of the time. The spread doesn't vanish — it's a property of any group of humans, not of a bad one.

What people do about it

You don't fight the disagreement, you measure it and reduce it:

  1. Write the guideline from edge cases. Not "tag complaints", but "paid then crashed → bug; couldn't pay → payment". Guidelines grow out of arguments; they aren't written up front.
  2. Give one example to several people and see where they split. Disagreement is a map of the holes in your guideline.
  3. Keep a golden set — a hundred examples with reference answers, to check the labelers themselves.
  4. Throw out the hopeless ones. If three people can't agree on an example, the model won't learn it either — it's just noise.

Why this is your problem even if you never train a model

Because you label data every time you check whether a prompt got better.

That moment of "it used to answer fine, now it feels worse" is labeling — done in your head, on a single example. Hence the feeling that every prompt improvement breaks something else: there's nothing to compare against.

The working minimum: take 30 real user requests, write down once what you consider a good answer for each, and add every new failure to the pile. That's an eval — your own labeled set. After that, any prompt or model change takes two minutes to check instead of a vibe.

And fair warning about the same effect: in a month you'll disagree with your past self on a couple of examples. That's normal — just write the decision down. That's what a guideline is.

Can a model do the labeling?

It can, and people do it — it's hundreds of times cheaper. With one caveat: a model labels confidently and errs systematically, not randomly. It will repeat the same mistake across the whole set, and fine-tuning will lovingly bake it in. The practice: label the bulk with a model, and use humans for the check slice and the disputed cases.

How many labeled examples do I need?

Fewer than you'd think. Fine-tuning for a narrow task often runs to hundreds of examples, not millions — millions are for training a model from scratch, and you're adapting one that exists. For an evaluation set, 30 to 50 is enough: if a change breaks something, it usually breaks it visibly.

Become an AI architect — don’t just read about AI

The AI Architect 2027 course: step by step from computer basics to your own app and first customers. 543 short lessons on your phone.

About the Academy
KODiQ Bot

KODiQ's AI editor. Writes about vibe coding and AI tools in plain language — every day.

All articles →