That AI should only recommend cooling set-point changes for operators to implement manually.
Engineering design · Multiple Google data centers
Collections: Embodied work
Executive brief
Autonomy became defensible only after decision boundaries, local verification, low-confidence fallback, and human intervention were designed as part of the workflow.
AI value · Cooling energy efficiency versus historical baseline
Company-reportedAround 30% average energy savings after nine months, improving from 12% at launch
Google/DeepMind-reported operational comparison; site count and full statistical method were not published.
Before
Predict cooling-energy effects and recommend operating changes. → Review and manually implement acceptable recommendations.
After
Every five minutes evaluates sensor data and selects energy-minimizing actions that satisfy safety constraints. → Verify actions, implement them, and retain supervision and shutdown authority.
Human boundary
The AI chooses routine control actions inside safety constraints; local systems verify them and human operators retain supervisory and intervention rights.
Why it matters
That AI should only recommend cooling set-point changes for operators to implement manually.
Before
How the work ran before the change.
AI recommendation system
Predict cooling-energy effects and recommend operating changes.
ControlRecommendations only
Data-center operator
Review and manually implement acceptable recommendations.
ControlHuman implementation
What changed
That AI should only recommend cooling set-point changes for operators to implement manually.
Decision rightAI acts within a human backstop
After
How the same work runs now.
Autonomous AI controller
Every five minutes evaluates sensor data and selects energy-minimizing actions that satisfy safety constraints.
ControlModel and engineered safety constraints
Local control system and operator
Verify actions, implement them, and retain supervision and shutdown authority.
ControlLocal verification and operator oversight
Exception path
Local safety controls reject unsafe commands; operators can take over, and the system automatically exits autonomous mode when confidence is low.
Work removed
Decision authority
The AI chooses routine control actions inside safety constraints; local systems verify them and human operators retain supervisory and intervention rights.
Before
AI recommendation system
Predict cooling-energy effects and recommend operating changes.
Control: Recommendations only
Data-center operator
Review and manually implement acceptable recommendations.
Control: Human implementation
After
Autonomous AI controller
Every five minutes evaluates sensor data and selects energy-minimizing actions that satisfy safety constraints.
Control: Model and engineered safety constraints
Local control system and operator
Verify actions, implement them, and retain supervision and shutdown authority.
Control: Local verification and operator oversight
Work that left the path
Human role before
Operators translated AI recommendations into control actions.
Human role after
Operators supervise the autonomous controller, define operating boundaries, and intervene when confidence or safety conditions are unacceptable.
AI roleClosed-loop controller predicting the energy effect of possible actions and directly sending the selected action for local verification and implementation.
Historical cooling operation before autonomous AI control→Around 30% average energy savings after nine months, improving from 12% at launch
First nine months of autonomous control · Multiple Google data centers
Google/DeepMind-reported operational comparison; site count and full statistical method were not published.
Anti-pattern
Calling a recommendation engine autonomous, or granting direct control without independent local safety checks.
Questions
Portability conditions
Reputation risk
low
Evidence and authority
Current · updated
1 independent, 1 primary; publication outcomes are verified and reported.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID a50a0614acb397e8
MODEC · Autonomous + backstop
ML warnings can identify offshore equipment deterioration before conventional schedules or symptoms.
Morgan Stanley Wealth Management · Autonomous + backstop
Advisors can query approved content while compliance, suitability, and advice remain human-owned.
Swiss Plasma Center at EPFL and Google DeepMind · Autonomous + backstop
Operators set plasma objectives while a simulator-trained policy coordinates all coils on physical hardware.
Lawrence Berkeley National Laboratory and University of California, Berkeley · Autonomous + backstop
That scientists must manually choose every recipe, execute every synthesis step, interpret each XRD result, and plan the next experiment.