education

the hidden human labor of training AI models.

10.28.2024

6 minute read

Behind the gleaming facade of artificial intelligence lies an illusion similar to the shadows in Plato's Cave.

Like the prisoners who mistake flickering projections for reality, we see AI's outputs without glimpsing the human hands that cast them.

Today, behind generative AI's impressive demonstrations of language, vision, and reasoning isn’t autonomous intelligence, but a vast hidden workforce of over 500,000 underpaid and undervalued people.

A key part of AI's development is data labeling, where people work to manually tag vast amounts of information to teach AI systems how to interpret inputs and generate appropriate outputs.

For example, someone might spend hours labeling traffic images so that autonomous vehicles can recognize nuances.

Others might categorize social media posts as 'threatening' or 'safe' to train content moderation algorithms.

The people that do this work are essential to making AI systems seem intelligent, revealing a reality reminiscent of Gilbert Ryle's "ghost in the machine," though not in the traditional Cartesian sense.

While AI is often portrayed as autonomous, it is shaped by the labor of individuals who invest their bodies, minds, and emotions into the process — yet their contributions remain largely invisible to us.

A recent investigation by TIME revealed that OpenAI employed workers in Kenya earning less than $2 per hour to label some of the darkest content imaginable — vile descriptions of violence, abuse, and hatred, so that ChatGPT could be brought to market “safely.”

This labor was crucial because early versions of the model, like GPT-3, often produced harmful and offensive outputs, drawing from toxic internet content used in its training.

These workers, who help make AI models "safe," report recurring traumas from their exposure to toxic and explicit content, with little to no mental health support.

This supply chain is deliberately opaque.

Most workers remain unaware of who their labor ultimately serves, as tech companies use layers of subcontractors to employ them.

Hence, they often don't know they're labeling data for industry giants leading the AI race.

This has formulated into a multi-billion dollar industry dependent on undervalued workforces in countries like Kenya, India, Colombia and the Philippines, often logging up to 18-hour days for minimal pay.

As companies seek to minimize the costs of this labor, the very systems these people help build may also soon replace them.

Advances in AI, especially in large language models and sophisticated embedding techniques, are beginning to automate the data labeling process itself to need less and less human review.

What we call 'artificial intelligence' is less a triumph of machine autonomy than a monument to hidden human labor - its sophistication built upon countless hours of underpaid work and unacknowledged trauma.

Peeling back the curtain reveals that our most advanced AI systems are not self-sufficient thinking machines but the result of a vast, global workforce that shape AI systems for our use.

AI is human-centered, and we must reshape the industry to see, reward, and empower those who really contribute to its advancement.

Calvary Rogers, Founder