Gig workers train AI by having conversations with strangers

Thousands of gig workers are earning money by having recorded conversations with strangers. The recordings are then sold as training data to artificial intelligence companies. Issie Lapowsky reports for Bloomberg Businessweek that the work is emotionally demanding, technically strict and raises serious questions about privacy and the future of human labor.

The platform at the center of the story is Babel Audio, owned by the two-year-old startup David AI Labs. Workers submit an audio clip to apply, and those who pass the screening can sign up for projects starting at $17 per recorded hour. Projects vary widely. Some involve open conversations on assigned topics. Others require reading scripts, annotating audio for emotional cues or role-playing characters such as therapists or pastors.

One worker, identified only as Gina, found herself sharing childhood trauma and memories of her Vietnam veteran father with a stranger playing a pastor. “He actually gave me some really good advice,” she says. “It took me aback.”

The emotional stakes are matched by strict technical requirements. A real-time meter tracks whether speakers are dominating or contributing too little to a conversation. Workers are evaluated on expressiveness, language proficiency, pause length and conversational depth. Swearing is banned. Sharing personal details is banned. And workers must never mention Babel itself, even when conversations touch on the topic of work.

Background noise is a common problem. Ryan, an Air Force veteran who joined the platform after his delivery income dried up, recorded calls between 11 p.m. and 4 a.m. to avoid his barking dogs. He frequently ended up discussing his divorce. Ashley, a comedian and YouTuber, was repeatedly penalized for popping her “p” sounds despite using a professional microphone. The platform eventually capped her at three calls per day. At 15 minutes each, the earnings barely added up.

The financial rewards can be significant for those who succeed. Gina earns around $600 per week on average and once made more than $10,000 in a single week after recording for 75 hours. “My voice was shot,” she says.

But the money comes with unease. A contributor agreement reviewed by Bloomberg Businessweek grants the company broad rights to use recordings for voice assistants, synthetic speech and other audio products. Workers are, in effect, helping to build the technology that may replace other workers.

“I’m not in the position to decide which income is morally best for me,” says Taylor, another Babel worker who earns roughly $200 a week on the platform.

Gina has since expanded into video projects that capture her facial expressions as well. Her reasoning is pragmatic: “At this point, I’m like, ‘They’ve already got my voice. I may as well make a little bit of money from it while I can.'”

The trend reflects a broader shift in AI development. According to a recent analysis by the freelance platform Upwork, demand for AI annotation and data labeling has increased by 154% since early last year.