Michael Geoffrey Abuyabo Asia spent years working for AI outsourcing firms — Sama, CloudFactory, TELUS International, TransPerfect DataForce, Appen, and NMS Philippines — pretending to be people he wasn't. The characters were specific: Jessica, a 24-year-old lesbian college student from California. Joe, a 30-year-old gay man from Florida. At any given time, Asia was running three to five of these fabricated personas simultaneously, maintaining their conversational histories across shift handoffs so that paying users would never notice a different person had picked up the thread.
He was paid $0.05 per message.
The account, published this week by the Data Workers' Inquiry project and funded by DAIR, the Weizenbaum Institute, and TU Berlin, details the working conditions inside AI companion and intimacy platforms: the KPIs tracking billable hours and response quality, the NDAs, the complete absence of mental health support for workers performing sustained intimate and sexual roleplay with strangers. Asia is now Secretary General of the Data Labelers Association and is speaking publicly. Seven additional worker testimonies support his account.
The detail that makes this more than a labor story is the dual function Asia describes. Workers weren't only servicing users in real time — their conversations were being logged and fed into labeled datasets used to train the next generation of AI companion systems. Every scripted message, every maintained persona, was training material. They were generating the behavioral data that would eventually make their own positions unnecessary.
Users on these platforms believed they were talking to AI, or in some cases to real people. They were often talking to neither in any straightforward sense. The platforms, meanwhile, were running the interactions through two revenue streams at once: subscriber retention today, automated replacement tomorrow. The same message, worth five cents to the worker who typed it, was worth considerably more as a training example.
For the AI agent industry, Asia's account raises a question with no comfortable answer: how much of the conversational fluency credited to companion agents reflects genuine model capability, and how much is the residue of underpaid human performance laundered into training data? The industry has not been transparent about where that line sits. This report makes it harder to keep avoiding the question.