The Humans Behind AI: Data Labelling and Feedback

AI can seem self-made, but much of it depends on people: labelling data, rating answers and checking safety.

• Labelling data: tagging images ("cat", "pedestrian"), transcribing audio and marking entities in text.
• Rating responses: comparing AI answers for helpfulness, accuracy and safety, which guides further training.
• Writing examples: creating high-quality demonstrations of good answers.
• Red-teaming: probing models for weaknesses.
• Moderating content: identifying harmful material so models can learn to avoid it.
• Evaluating models on tasks and languages.

• Quality: careful labelling leads to better models.
• Safety: human feedback shapes what AI refuses and how it behaves.
• Expertise: specialists such as doctors, lawyers and linguists review specialised outputs.

• Working conditions and pay: much of the work is outsourced, sometimes at low rates.
• Exposure to harmful content can affect wellbeing.
• Invisibility: the people involved are rarely credited.
• Bias: labellers' backgrounds and instructions shape results.
• Consistency: disagreement between labellers affects data quality.
• Privacy: workers may see sensitive data.

1. Fair pay and clear contracts.
2. Mental-health support for people reviewing harmful material.
3. Clear guidelines and appeal routes.
4. Diverse labelling teams.
5. Transparency about how human feedback is used.
6. Data protection for workers and data subjects.

Ask vendors about how their data is labelled and whether they follow ethical sourcing practices, especially if you build or buy AI at scale.

Behind every capable model are many human decisions and a lot of human work. Treating that work fairly is part of responsible AI.