Glossary

AI training data work

The umbrella category for paid tasks that improve AI models: labeling images and text, transcribing audio, evaluating and ranking model responses, writing prompts, and reviewing code. Work ranges from simple microtasks to expert-domain projects.

Annotation

Marking up data for machine learning: drawing boxes around objects in images, tagging entities in text, segmenting audio, or labeling video events. Annotation is the entry level of AI-training work — no degree required, but precision matters, since rejected batches are typically unpaid.

Audio hour

The standard billing unit for transcription work: one hour of recorded audio, regardless of how long the file takes to transcribe. Transcription platforms advertise pay per audio hour (or per audio minute), so a $60/audio-hour rate is not an hourly wage — beginners typically need four to six hours of work time per audio hour, making effective pay a fraction of the headline figure..

Used by: transcription platforms

Audio minute

One minute of recorded audio — the finer-grained cousin of the audio hour, used by platforms that price work per file rather than per hour of footage. A rate of $0.50 per audio minute equals $30 per audio hour.

Used by: transcription platforms

Moderated interview

A live, scheduled session in which a researcher leads the tester through tasks or questions in real time, usually over video call. Moderated sessions pay substantially more than unmoderated tests — often $30–$90 for an hour — because they require scheduling, reliability, and strong communication.

Prompt evaluation

AI-training work in which contributors rate and compare AI model responses to prompts — scoring helpfulness, accuracy, and safety, and explaining their judgments. It is the most widely available expert-tier AI task and typically pays per task or per hour.

Used by: AI training platforms

RLHF (reinforcement learning from human feedback)

The technique behind most AI-training gig work: humans rank or score model outputs, and those preferences train the model to behave better. RLHF is why platforms hire thousands of contributors to evaluate responses — the ranking data is the product.

Think-aloud protocol

The core method of usability testing: the tester narrates their thoughts — what they see, expect, like, and find confusing — while completing tasks on a website or app. Platforms train testers in the protocol during onboarding, and clear think-aloud commentary is the single biggest factor in passing the sample test and receiving more test invitations..

Used by: app and website testing platforms

Usability testing

Paid tasks in which real users explore a website, app, or prototype while verbalizing their experience, so product teams can find friction. Tests typically run 5–60 minutes, are completed from home with a microphone (often a webcam), and pay per completed test.

Verbatim transcription

Transcription that captures every spoken word exactly as said, including filler words ("um", "uh"), false starts, and sometimes non-verbal sounds. Verbatim work pays more than clean-read transcription because it is slower and more demanding.