Mark Zuckerberg told employees this week that Meta's AI agents aren't developing fast enough — but the company had zero trouble rolling out mandatory keystroke and mouse-tracking surveillance on its own workforce. The disconnect tells ordinary Americans everything they need to know about where Big Tech's competencies really lie.
At an internal town hall Thursday, Zuckerberg admitted that four months after a major restructuring designed to accelerate AI development, progress has stalled. "The kind of trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected," he said, according to Reuters. The bets Meta made earlier this year "haven't come to fruition yet."
That's quite the admission given what Meta has spent. The company is projected to pour up to $145 billion into AI infrastructure this year. In May, Meta slashed roughly 8,000 jobs — 10% of its global workforce — while simultaneously reassigning some 7,000 employees into AI roles. Zuckerberg previously framed the cuts as a matter of capital spending priorities, telling employees the company runs on two cost centers: compute and people, and headcount would keep bending toward the former.
So the people got bent. The compute? Not delivering.
TNW reported that Zuckerberg and other executives were "super optimistic" about coding tools like Anthropic's Claude Code early this year, expecting them to translate into faster progress. They didn't. Zuckerberg offered employees a vague promise of "more significant benefits" within three to six months but declined to specify which products or teams would deliver them. A Meta spokesperson declined to comment when pressed by Reuters.
But while Meta's AI agents flounder, its surveillance apparatus works just fine. CTO Andrew Bosworth told the same town hall that an internal keystroke and mouse-tracking tool — rolled out in April with no opt-out option — will now resume only on an opt-in basis. Business Insider reported that the mandatory version, part of what Meta calls its Model Capability Initiative, sparked an internal backlash after a leak exposed employee conversations and keystrokes to their own colleagues. Bosworth acknowledged the rollout damaged morale and trust — but also noted the tool generated more useful data than expected.
They surveilled their own people efficiently. They just couldn't build the product they claimed justified it.
The culture inside Meta's AI operation is deteriorating. Business Insider previously reported that Bosworth warned staff morale was "probably one of the worst it's ever been" in the company's 20-year history. Meta recently gave engineers the option to leave its Applied AI task force — a climbdown employees dubbed "the undraft" — after forcibly reassigning thousands to the unit.
Zuckerberg still insists Meta is on a "journey to superintelligence." Maybe. But the company that can't build a working AI agent managed to build a keystroke logger and deploy it against its own workforce without asking permission. Americans who trust this company to police speech while pursuing "superintelligence" should ask themselves which capability Meta will actually perfect first.
