Nearly 1 in 5. That's the share of AI data annotation workers who have experienced homelessness, according to a new investigation into the workforce that trains large language models like the one behind ChatGPT.
Data annotation - the process of labeling text, rating AI responses, and flagging harmful content so models learn what's useful and what's dangerous - is foundational to how AI works. Without it, ChatGPT doesn't know which answer is better, which output crosses a line, or how to interpret an ambiguous request. But the people doing this work are predominantly gig workers: no benefits, no guaranteed hours, paid by the task, often competing globally.
What the work actually pays
A typical annotation task might pay a few cents per labeled image or a dollar or two for rating a pair of AI responses. Workers pick up tasks in batches with no income guarantees. Quality requirements are strict - rejection rates run high - but pay doesn't reflect the cognitive skill involved. For US-based workers competing against annotators in lower-cost countries, the economics are brutal.
That's partly why the homelessness figure hits hard. These aren't workers in regions where per-task rates stretch further. These are American workers doing skilled, repetitive cognitive labor and still unable to afford stable housing.
AI companies typically hire annotators through third-party platforms like Scale AI or Remotasks, which creates distance between model developers and the people whose daily work makes those models functional. OpenAI and others can point to contractors rather than employees when labor conditions come up.
The gap that's getting harder to ignore
When a new ChatGPT model ships, press coverage focuses on benchmark scores and capability improvements. The training data - and the humans who labeled it - rarely appear in the announcement.
This investigation adds to a documented record. TIME's 2023 reporting on Kenyan workers doing content moderation for OpenAI found people earning less than $2 per hour and experiencing psychological harm from exposure to disturbing material. The pattern is consistent: the more capable AI becomes, the more invisible labor goes into building that capability.
In a demo video accompanying the investigation, workers described inconsistent pay, sudden task droughts, and the gap between what they earn and what the products they help build are worth commercially.
The tools get smarter. The people training them often can't pay rent.