Leaked audio posted on social media by More Perfect Union captured an internal exchange in which Meta Platforms Inc CEO Mark Zuckerberg addressed employee concerns about the company collecting device activity to train artificial intelligence systems.
During the meeting, an employee raised questions about the practice after recent company announcements. Zuckerberg explained that Meta's AI training strategy relies on three pillars: research and architecture, infrastructure, and data quality. He said that models learn by watching people perform tasks and that employees represent a higher quality data source than the contract workers commonly used elsewhere.
"If we're trying to teach the models coding, for example, then having people internally build tools or solve tasks that help teach the model how to code, we think is going to dramatically increase our model's coding ability," Zuckerberg said.
Zuckerberg sought to draw a clear line between using employee-generated signals to train models and monitoring individuals. He said that no humans are reviewing individual employee activity and that content is stripped out where possible. The CEO added that the company is not using the data for surveillance or to measure employee performance, saying its sole purpose is to improve AI systems.
He also acknowledged that Meta could have explained the initiative more effectively. Zuckerberg framed the decision to limit disclosure as a balance between transparency and protecting strategically sensitive details from competitors, describing AI development as one of the most competitive fields in history. He said that the company may continue similar training efforts with employee data if the approach proves effective.
The audio and Zuckerberg's comments highlight how Meta positions internal work activity as a component of its AI training data strategy while asserting safeguards against human review and operational surveillance. The CEO's remarks underscore both a data-quality justification for using employee signals and a recognition that communication around the program did not meet employee expectations.
Summary
Leaked internal audio shows Zuckerberg telling staff that Meta trains AI models by observing employee activity on computers, framing employees as a higher-quality data source, denying human review of individual activity, and acknowledging communication shortcomings.
Key points
- Meta trains AI using observations of employees performing tasks on their computers, which the company views as higher-quality data than contract worker activity - impacted sectors: technology, AI research, and enterprise software.
- Zuckerberg described AI development as resting on research and architecture, infrastructure, and data quality, and gave coding as an example of tasks used to teach models.
- The company says no humans review individual employee activity and that content is stripped where possible; Meta also indicated the approach may continue if effective.
Risks and uncertainties
- Employee trust and internal morale could be affected if communication about data use remains inadequate - relevant to human resources and corporate governance within technology firms.
- Perceptions of workplace surveillance, even if the company says data is not used for performance tracking, could raise regulatory or reputational scrutiny in the technology and AI sectors.