AI companies want to harvest improv actors' skills to train AI on human emotion
Briefly

AI companies want to harvest improv actors' skills to train AI on human emotion
"If you've got strong creative instincts, the ability to authentically portray emotion, and are capable of staying true to a character's voice throughout a scene, there's a job listing calling for your experience. The catch: You won't be performing in a theater, a film studio, or an underground performance space. You'd be using your talents to train an AI model for 'one of the leading AI companies,' according to the open role posted by Handshake."
"AI models are often described as 'jagged,' meaning they're typically great at some surprisingly complex tasks but fail deeply at some simple ones. AI companies are trying to fix the gaps in their models' knowledge with specialized data labeling, and companies like Handshake, Mercor, and Scale AI have adjusted accordingly, hiring professionals in a wide range of industries."
"Handshake's demand for training data tripled last summer, and the company surpassed a $150 million run rate in November, scrambling to keep up with demand. Handshake and its competitors have touted their networks of tens of thousands (or more) professionals in white-collar industries, from chemists and doctors to lawyers and screenwriters."
Handshake and similar companies provide specialized training data to major AI labs like OpenAI by hiring professionals across various industries, including creative fields. AI models exhibit uneven performance—excelling at complex tasks while failing at simpler ones—creating demand for niche data to address these gaps. Handshake's training data demand tripled in one year, reaching a $150 million run rate. The company now recruits actors, comedians, screenwriters, and other creative professionals alongside white-collar workers like doctors and lawyers. These professionals increasingly worry that by training AI systems, they're accelerating the obsolescence of their own careers and industries.
Read at The Verge
Unable to calculate read time
[
|
]