The Berkeley Study

In The You On AI Encyclopedia

The study's methodology was ethnographic rather than statistical: the researchers embedded themselves in the company, attended meetings, watched screens, talked with workers, and documented patterns of behavior that quantitative surveys would have missed. This approach proved essential because the key findings concerned not what workers reported but what they did — and what they did diverged sharply from the conventional narrative that AI would reduce workload by automating routine tasks.

The three central findings map precisely onto Wiener's framework. First, AI did not reduce work; it intensified it. Workers adopted AI tools to accelerate existing tasks, then used the time saved to take on additional tasks, then expanded into domains previously belonging to other teams. The boundaries between roles blurred. Designers started writing code; engineers started drafting marketing copy. The scope of each person's job widened continuously, and the widening felt natural rather than imposed. Second, work seeped into pauses. Workers prompted on lunch breaks, snuck in requests during meetings, filled gaps of one or two minutes with AI interactions. The pauses that had served as cognitive rest — informally, invisibly — disappeared. Third, multitasking became the norm and fractured sustained attention. AI could handle low-effort tasks in the background while the human worked on something else, producing a sense of always-juggling that felt productive in the moment and exhausting in aggregate.

The study's cybernetic significance is that it documents, with empirical precision, the transition of a human-machine system from one operating mode to another. Before AI deployment, the company's workflows were governed by implementation friction: each task took a certain minimum time, each role had defined scope, each worker had periods of necessary rest built into the structure of the work itself. After AI deployment, the friction collapsed, the governors were removed, and the system entered positive feedback runaway — not because anyone chose it, but because the dynamics of the achievement society, previously constrained by mechanical slowness, now had nothing constraining them. The workers were not being exploited by external managers. They were exploiting themselves, in exactly the pattern Han had predicted and Wiener had warned against seventy-five years earlier.

The researchers' proposal — which they called 'AI Practice' — was a governor design: structured pauses built architecturally into the workday, sequenced rather than parallel workflows, protected mentoring time, behavioral training alongside technical training. Segal adopts this framework in You On AI as the operational specification of what he calls dams — the negative feedback structures that convert raw AI power into sustainable capability. The researchers stopped short of prescribing specific policy, but the implication is clear: the governors must be built into the architecture of work rather than left to the willpower of workers whose willpower is precisely what the positive feedback loop has overwhelmed.

Origin

Ye's doctoral work at UC Berkeley's Haas School of Business focused on technology adoption in organizations. Ranganathan, her advisor, had published extensively on labor and organizational dynamics. The AI ethnography began in summer 2025 as the next-generation AI coding tools were crossing the adoption threshold Segal documents in You On AI.

The Harvard Business Review publication in February 2026 gave the findings unusual reach: HBR's readership includes the executive audience whose deployment decisions the study's findings most directly concern.

Key Ideas

AI intensifies work, not reduces it. The aggregate effect of AI adoption was more work per worker, not less.

Task seepage. AI-assisted tasks colonized previously protected pauses, eroding the invisible structure that had enabled sustained attention.

Multitasking fractures attention. The ability to run AI tasks in parallel produced a juggling mode that felt productive but consumed cognitive resources invisibly.

Self-chosen, not imposed. Workers were not forced to work more; the internalized imperative converted possibility into compulsion.

Erik Brynjolfsson

"The biggest effect of general-purpose technologies on productivity comes not from the technologies themselves but from the complementary innovations that unlock their potential."

The Productivity J-Curve: How Intangibles Complement General Purpose Technologies · 2021

→ Erik Brynjolfsson — On AI: Chapter 2: The Shape of the J-Curve

Governors are architectural. The prescribed response — AI Practice — requires structural changes to work design, not individual willpower.