Northville teacher workshop
The Mechanical
Turk
Where is the work
happening?
A hidden human player
made the moves.
Jason Kridner
Amazon Mechanical Turk
Human work becomes
training data
- 01A requester defines a task.
- 02Workers label, sort
or judge examples. - 03The data can support
model training.
Who sets the task?
Who gets paid?
Paper activity
One word at a time
PROMPT
Tell a short story about a surprising discovery at school
RESPONSE
At recess, we found a…
- Read the prompt and
the whole response. - Add one word,
then pass the paper. - When it feels complete,
write STOP.
After the paper activity
What shaped the story?
Same prompt.
Different stories.
Which words felt likely?
Where did someone surprise you?
You brought experience and judgment.
Language models generate tokens,
which can be parts of words.
This activity explores context and choice. It does not simulate model training.
A 2022 training example
Human feedback
in InstructGPT
- 01
Demonstrate
People write example answers.
The model learns from them. - 02
Compare
People rank model responses.
A reward model learns
their preferences. - 03
Fine-tune
Further training uses
the reward model’s scores.
This is a separate example of human labor in AI, not a claim that OpenAI used MTurk.
Questions for our classrooms
Whose judgment
counts?
Whose examples define “good”?
Who has access to the tools?
Who does the work,
and who receives the value?
What would you want to know before using an AI tool with students?