The dark human cost behind ChatGPT's safety filters

Everyone is talking about how smooth and helpful ChatGPT feels. But behind that polite conversational style lies an uncomfortable reality.

According to an investigation by Time, OpenAI relied on low-wage workers in Kenya to filter toxic content out of the system, paying many of them less than $2 an hour.

Here is how one of the biggest breakthroughs in tech was built on deeply troubling labor practices.

The human detox filters behind the chatbot

Back in November 2021, OpenAI was wrestling with a well-known flaw in GPT-3. Because the model ingested hundreds of billions of words from all over the open internet, it regularly generated offensive, toxic, and hateful responses.

To clean up the dataset, OpenAI partnered with Sama, a San Francisco-based outsourcing firm that ran operations in Kenya.

Workers in Kenya were assigned to review and label horrifying text snippets, covering topics like child sexual abuse, torture, suicide, murder, bestiality, incest, and self-harm.

All four Sama workers interviewed by Time reported lasting mental scars. One employee described reading the horrific text as feeling like torture, leaving them emotionally disturbed by the weekend. While Sama claimed counseling was available at any time, workers said the therapy sessions were ineffective and rarely used due to heavy workloads.

An image project breaks the partnership

Things escalated in February 2022, when OpenAI launched a separate project asking the Kenyan workers to collect sensitive images containing sexual and violent content.

OpenAI defended the request, telling Time the work was essential to make future AI tools safer.

Within weeks of starting the task, Sama canceled all of its contracts with OpenAI. Reports suggest the decision was partly driven by PR fallout from a similar moderation scandal Sama was facing over its work with Facebook.

OpenAI later blamed the situation on a miscommunication, claiming it never requested the most extreme imagery and never viewed what was sent. But when the partnership collapsed, Kenyan workers paid the price: some lost their jobs entirely, while others were demoted to lower-paying gigs.

Why it matters

Silicon Valley loves to pitch generative AI as an ethereal, purely technological breakthrough. In reality, modern AI still relies heavily on poorly paid human labor to clean up the garbage before models are shipped to the public.

For everyday users, having a safe, non-toxic chatbot feels effortless. But that safety barrier was effectively bought with the psychological distress of workers in East Africa who were paid less than $2 an hour.

For the tech industry, this report marks a turning point. As giants pour billions into AI, supply chains for human data labeling and safety filtering are going to face fierce scrutiny from regulators and the public alike.

We think this reliance on bottom-dollar overseas labor is a massive ethical liability. Trying to cut corners on the most emotionally devastating work in tech might save a balance sheet in the short term, but it exposes AI companies to devastating reputational damage.