What is data labeling?

AI and people intelligence

Data labeling is the process of adding meaningful tags, categories, or annotations to raw data, such as text, images, or records, so a machine learning model can learn from it or be tested against it. People, software, or both apply the labels, and each labeled example shows the model what the correct answer looks like.

How data labeling works

A label is the answer attached to an example. It might be a category for an image, a topic for a sentence, or a judgment about a record. The EU-U.S. Terminology and Taxonomy for Artificial Intelligence defines supervised learning as "machine learning that makes use of labelled data during training." The model learns the pattern that links each input to its label, and a separate set of labeled test data shows how well it learned.

Labels can be applied by people working from shared guidelines, by software, or by a mix of the two, where software proposes labels and people check them. Specialists are needed when a label depends on domain judgment that a general annotator would not have.

The EU AI Act treats labeling as part of data governance. Article 10 says the data governance practices for high-risk AI systems must cover "relevant data-preparation processing operations, such as annotation, labelling, cleaning, updating, enrichment and aggregation," along with an examination of possible biases in the data.

Why data labeling matters for talent teams

A model can only be as accurate as the labels it learns from and is tested against. A study presented at NeurIPS 2021 (Northcutt, Athalye, and Mueller) found label errors in the test sets of 10 commonly used computer vision, language, and audio datasets, an average of at least 3.3% across the 10. At least 6% of the ImageNet validation set was mislabeled. The authors also studied how those errors can change benchmark results.

Hiring raises the stakes. The EU AI Act lists AI systems used to recruit or select people, including systems that analyze and filter applications and evaluate candidates, as high-risk (Annex III, point 4), so the Article 10 data governance requirements apply to them. This page is a general overview, not legal advice.

Labels in talent data also carry judgment. Deciding whether someone's experience counts as leading a team through rapid growth, or whether two job titles mean the same role, is a call an experienced recruiter makes differently from a general annotator. When evaluating AI for hiring, talent teams can ask:

Who defines the labels: Whether people with domain expertise set and review them.
How labels are checked: Whether there is a review step for errors and disagreements.
How labels stay current: Whether they are refreshed as careers and companies change.
What a user can see: Whether each recommendation shows the evidence behind its labels.

How Findem approaches data labeling

Findem's Labeling Engine turns human expertise into Success Signals and Relationship Signals. Career and company records are first organized into structured 3D data, and duplicates, noise, and outdated information are removed to create a reliable foundation for expert labeling. Recruiters, researchers, and domain specialists then define the patterns that signal potential, performance, and fit, and Findem continuously refreshes and enriches those signals.

According to the Findem platform page, more than 200 experts contribute to this work, Findem has labeled more than 2 million Success Signals, and each profile carries 75 to 100 Success Signals. Findem also publishes regular releases of new expert-labeled attributes.

Labeled signals inform recommendations. Findem's AI does not reject candidates or make hiring decisions, and people review results before they influence hiring decisions. Read more in Responsible AI at Findem.

Related terms

Sources

EU-U.S. Terminology and Taxonomy for Artificial Intelligence, first editionEU-U.S. Trade and Technology Council, published by NIST
EU AI Act, Article 10: Data and data governanceEuropean Commission AI Act Service Desk
EU AI Act, Annex III: High-risk AI systemsEuropean Commission AI Act Service Desk
Pervasive label errors in test sets destabilize machine learning benchmarksNorthcutt, Athalye, and Mueller, NeurIPS 2021 Datasets and Benchmarks Track
Findem Labeling Engine
Success Signals
The People Intelligence Platform

Frequently asked questions.

It is the work of attaching the correct answer, such as a category or tag, to examples in a dataset. Supervised learning uses those labeled examples during training, and labeled test data is used to measure how accurate the trained model is.

Usage varies. Many teams use the two words interchangeably. Others use annotation for marking up content, such as highlighting a phrase in a document, and labeling for assigning a category. The EU AI Act lists annotation and labelling side by side as data-preparation steps.

A model learns from its labels and is judged against them, so errors carry through. A NeurIPS 2021 study estimated an average label error rate of at least 3.3% across the test sets of 10 widely used datasets and found that such errors can change which model appears to perform best.

People, software, or both. Software can apply labels at scale, and people check them or handle the cases that need judgment. When a label depends on domain knowledge, such as what counts as relevant leadership experience, specialists in that field define and review it.

Findem's Labeling Engine structures career and company data, then recruiters, researchers, and domain specialists label the patterns that signal potential, performance, and fit. The results are Success Signals and Relationship Signals, which inform recommendations that people review before any hiring decision.