AI labs are hiring improv actors to teach models how to feel, because apparently reading the internet wasn't enough to learn empathy.

The Summary

  • Handshake AI is recruiting improv actors to generate training data for "leading AI companies" including OpenAI, specifically targeting emotional authenticity and character consistency
  • This marks a shift from scraping existing content to commissioning human performance specifically for AI training
  • The move signals that current models still can't crack genuine emotional intelligence, even with trillions of tokens

The Signal

We're watching the moment AI training stops being about scale and starts being about quality. Handshake AI isn't looking for people to label images or transcribe audio. They want actors who can stay in character, deliver authentic emotion, and improvise with creative instinct. That's the stuff you can't scrape from Reddit threads or movie scripts.

The specificity matters. AI companies have already ingested essentially everything publicly available on the internet. They've read every screenplay, every novel, every therapy subreddit. And still, the models sound like chatbots when they try to express nuanced emotion or maintain consistent personality across a long conversation. The uncanny valley isn't in the words anymore, it's in the humanity behind them.

Handshake provides training data to OpenAI and other labs, which means the frontier models are hungry for something current datasets don't provide: human performance data created specifically to teach emotional range and character coherence. Think about what that implies. The companies building AGI think the missing piece isn't more tokens or bigger parameter counts. It's watching how a skilled human inhabits a character and responds authentically in the moment.

This also reveals the economics of AI's next phase. We've moved from "data is free" (scrape the web) to "data is cheap" (pay Mechanical Turk workers) to "data is skilled labor" (hire professional performers). The per-token cost of this training data is orders of magnitude higher than anything that came before. That's not a scaling move. That's an admission that the easy gains are over.

The Implication

If you're a creative professional wondering whether AI will replace you, here's your answer: not yet, but it's studying you closely. The labs need human expertise to teach models what humanity looks like. That creates near-term work for actors, writers, and performers. But every session you do is teaching the model to need you less next time. Watch for two things: how much these gigs pay (if it's good money, the data is valuable and scarce), and how long the work lasts (if it dries up suddenly, they figured it out).

Source: The Verge AI