TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Ethics & Provenance · AGE 17—20 · FACTUAL · EVOLVING
The people inside the model
People rating, moderating and writing rules sit between a raw model and a usable assistant.
The idea
A raw model is not a usable assistant. Between the two sits work: people rating answers so the model learns which replies to prefer, people writing the rules it is tuned on, people labelling the worst material on the internet so it learns to refuse. For the assistants built around 2022–23, much of that work was done in Kenya, the Philippines, India and elsewhere, for low pay, some of it exposing workers to disturbing content for hours a day. Newer pipelines mix human judgment with AI feedback, written constitutions and classifiers. The human share has changed shape, not vanished.
Why it matters
Architecture already asks where the marble was quarried and who laid the brick. AI has a supply chain too. Pretending it does not is a choice, and it is yours.
See it in the studio
The moderation desk is in Nairobi; the labelling desk is also in Ranchi, where women employed by annotation firms mark hand joints, brain scans and road scenes — Karishma Mehrotra's 2022 reportage Human Touch follows them, and the 2024 film Humans in the Loop shows one such desk in Jharkhand. That labelling side has its own card on this map, The labour inside the dataset; this card is the work that comes after training. Now picture a studio in Bengaluru that specifies certified timber and fair-wage contractors, and runs a hundred prompts a day. Its specification sheet has a blind spot, and the people in that blind spot live two states away — or a continent away.
Watch for this
"The model learned it by itself." The model learned its manners from feedback — human ratings, and increasingly AI feedback trained on them — and learned to refuse from humans who read what it must refuse, and from classifiers built on their labels. "By itself" is never the whole story.
Try it
Read the 2023 Time investigation of the Kenyan workers who labelled material for OpenAI. Then find one current model's own account of how it was aligned — the system card or technical report — and list which steps it says were done by people and which by models or classifiers. Write a materials-style declaration for that tool — origin, labour, conditions — with honest blanks where you could not find out.
Prove it
Explain what human raters and moderators add between a raw model and a usable assistant, say how that work is now shared with AI feedback and automated checks, and draw the parallel between the worker behind a prompt and the labourer behind a building material.
How it works
The rating step is called reinforcement learning from human feedback, RLHF. Later methods — reinforcement learning from AI feedback, training against a written constitution, synthetic preference data, automated graders and safety classifiers — reduce how often a person has to read each example; but the classifier was trained on examples a person once labelled, and people still write the rules, rate the hard cases and red-team the result. The moderation step is content labelling, usually outsourced through contractors. The Time report documented one such contract in Nairobi, published in 2023, with take-home pay reported between roughly $1.32 and $2 an hour — one contract, one generation of models, not a description of every pipeline since. India is a large part of this workforce. Conditions, pay, methods and geography change as companies and contractors change, which is why this card is EVOLVING. The labour itself is a lasting fact; the terms of it, and its share of the work, are not.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.