← The sketchbook
On the job
Training an AI, one judged task at a time
AI training work is rating answers, writing examples, checking facts and following a rubric, usually after a test, paid by the hour or the task, then reviewed.
Drawn in 6 stages, the way an illustrator fills a field sketchbook. Press play to watch it drawn.
Training an AI, one judged task at a time FreelanceSites sketchbook
ADMIT ONE ROW SEAT rate answers check facts write examples the rubric per hour test first or per task then a quality review The rubric wins. Guessing gets sent back. Finished sketch: 6 stages, 448 pencil marks
Watch it drawn 1× Skip to finished Blocking in the robot, the laptop and four tasks Outlines and washes The four jobs, the test and the pay A small sweat drop and rubric ticks Shading the robot's right side and the desk Caption
What the sketch shows
Some platforms pay people to show a model what a good answer looks like. You are not chatting with a robot for fun. You are making a stack of small judgments, and someone checks them against a written standard.
Four kinds of task show up again and again. You rate two chatbot answers and say which one followed the instructions. You write an example yourself, so the model has a target worth imitating. You check a claim against a source, which means opening the source, not nodding because the sentence sounds smooth. And you follow a rubric: a list of what earns a high score and what earns a low one. When your taste and the rubric disagree, the rubric wins.
Before the paid queue opens, most projects ask for a qualification test or a sample. Fail it and that project stays closed. Pass it and tasks appear, until the project has enough data and the queue goes quiet. That stop and start is normal. It is not a salary, and a quiet week is not proof you did the last batch wrong.
Pay, and the review after it
Pay is quoted per hour or per task. It changes with the subject. A careful rubric in a technical field and a simple rating task are not priced the same, and a rate on one site is not the rate on the next. The listings beside this sketch carry whatever the platforms have published. Nothing drawn here is a number you should budget with. A quality review sits at the end. A person, or a second pass of the rubric, scores your tasks. Rushing, guessing a fact, or copying an answer from the model you are supposed to be judging gets the work sent back. It can also close the project for you. Slow, sourced, and matched to the rubric beats a huge pile of sloppy tasks.
Key points
The usual tasks are rating two answers, writing an example, checking a fact against a source, and scoring with a written rubric. A qualification test or a sample usually comes before paid work, and projects stop when they have enough data. Pay is per hour or per task and changes with the subject. Reviewers send weak work back. The drawing names no rates.
