Meta’s employee keystroke plan made HN focus on privacy, not training data

The Hacker News thread around Reuters’ Meta story passed 747 points because it hit a raw nerve: not “AI needs training data,” but what happens when employee mouse movement and keystroke traces become part of that data supply. The reported plan says Meta will start capturing those signals for AI training. HN readers pushed the discussion toward workplace privacy, operational risk, and whether internal trust boundaries can survive this kind of collection.

The sharpest concern was purpose limitation. The thread repeatedly returned to the reported assurance that the data would not be used for performance assessment, only for model training. Many commenters did not treat that as enough. Developer machines are not clean lab devices. They show terminals, tickets, customer identifiers, incident notes, dashboards, credentials, and the messy path by which work actually gets done. Even if the initial use is training, the capture surface itself is broad.

• Mouse movement and keystrokes can expose workflow, hesitation, review patterns, and sensitive context.
• Routine endpoint logging feels different from building an always-on training feed out of employee activity.
• AI training purpose does not by itself answer consent, retention, access control, or audit questions.

Community discussion noted that “training only” is not a privacy model. It is a stated use case that still needs technical limits. The reaction was also shaped by Meta’s reputation: a company long criticized for consumer data practices is now asking its own employees to accept a deeper form of instrumentation. That made the thread less about one HR policy and more about a broader pattern in AI systems. As model builders chase realistic work traces, the next scarce resource may be internal legitimacy.

The bigger lesson for AI teams is that training data collected inside a company still needs product-grade governance. Clear retention limits, redaction, access logs, opt-out paths, and independent review matter because employee activity is not just another dataset. It is the working surface of the people being asked to trust the system.

The source is the Reuters report, with the HN discussion at https://news.ycombinator.com/item?id=47851948.