.svg)
What is Data Annotation in AI?
Data annotation (data labeling) is the process of marking raw inputs—text, images, audio, or video—with categories, spans, or attributes so supervised machine learning models can learn from examples.
High-quality labels define the ceiling for model performance. Errors, inconsistent guidelines, and poisoned labels become training-time failures that are hard to debug later. Teams often combine in-house labeling, vendors, crowdsourcing, and human-in-the-loop review.
Mitigations
- Written guidelines, inter-annotator agreement checks, and gold sets.
- Audit for systematic bias and adversarial/poisoned samples.
- Version labels with the dataset so regressions are reproducible.
Related reading on Giskard
Treat labels as a security surface — Poisoned or inconsistent annotations create silent model failures—test for data-poisoning and quality regressions before you trust new labeled sets. Learn more: Data poisoning on enterprise LLM apps.
Authoritative reference: Google ML Crash Course: Data
