Also called: data annotation · Last checked: 7 November 2026
In short: Data labelling is the work of tagging raw data so an AI model can learn from it. People mark what each item shows, such as “spam”, “cat” or “angry customer”, and the model learns to copy those judgements.
Table of Contents
How it works
- Write the guide. A team defines the labels and gives clear examples of each, so every labeller makes the same call.
- Label the items. People, often helped by software, tag thousands of texts, images or recordings one by one.
- Check and train. Reviewers check a sample for mistakes, and the labelled set is then used to train or test the model.
Where you’ll meet it
Tools such as Label Studio and services such as Amazon Mechanical Turk organise labelling work. When you rate a ChatGPT answer with a thumbs up, you add a label too.
What it means for your business
- Your experts hold the labels. Staff who know your products and customers can label data better than any outside crowd.
- Budget for it. Labelling can take more time and money than building the model. Include it in any AI project plan.
- Check your suppliers. If you buy labelled data or labelling services, ask how workers are paid and supported.
Watch out for
- Inconsistent labels. If labellers disagree, the model learns the confusion. Test agreement on a sample first.
- Harm to workers. Labelling violent or abusive content can damage people’s mental health.
Example
Sama. In January 2023 TIME reported that Kenyan workers at Sama, an outsourcing firm, labelled descriptions of sexual abuse, hate speech and violence to help OpenAI build a filter for ChatGPT. TIME said they took home about $1.32 to $2 an hour. Sama said workers earned $1.46 to $3.74 an hour after tax, and it had ended the work in February 2022.
Data labelling vs training data
| Data labelling | Training data | |
|---|---|---|
| What it is | The task of tagging data | The examples a model learns from |
| Who does it | People, helped by software | Collected from many sources |
| How they relate | Turns raw data into usable data | Often includes labelled data |
Frequently asked questions
Can AI label data instead of people?
Partly. AI models can now produce first-pass labels quickly and cheaply. People still need to check samples, handle hard cases and write the rules the labels follow.
How much labelled data does a project need?
It depends on the task. Adjusting an existing model may need a few hundred good examples, while training a new one from scratch can need many thousands.
Is it safe to send company data to a labelling service?
Only with a contract that covers confidentiality, data protection and deletion. Remove personal details first where you can, and check where the workers and servers are based.
Related terms
Training Data
Supervised Learning
Reinforcement Learning from Human Feedback