Wiener’s Feedback Loops: Ethical AI for Utilities

Norbert Wiener’s The Human Use of Human Beings is 75 years old. That makes it older than the modern electric grid as we operate it today—restructured markets, SCADA everywhere, renewables on the margin, batteries behaving like power plants, and algorithms deciding what happens next.

And yet Wiener’s central idea—feedback—is now the defining technical and operational reality for energy and utilities. We’ve built an industry where sensors, controllers, forecasting models, and autonomous optimization tools are tightly coupled. When something changes, the system responds. The system’s response changes conditions again. The loop never stops.

Paul Jones’s recent poem reflecting on Wiener frames it in human terms: “feedback loops of love and grace,” mixed with unease, contingency, and the reminder that “every web conceals its spider.” That line hits especially hard in utilities, where AI-based automation can feel like a black box sitting between operators and the grid.

This post is part of our AI in Robotics & Automation series. If you work in operations, asset management, grid modernization, or utility innovation, here’s the stance I’ll take: most utilities don’t have an AI problem—they have a feedback problem. Fix that, and AI becomes safer, more useful, and easier to scale.

Why Wiener still matters for AI in energy automation

Wiener’s message wasn’t “machines are coming.” It was more specific: when you connect humans and machines through feedback, you’re designing behavior—not just software.

In utilities, that’s not abstract philosophy. It’s daily work:

• Volt/VAR optimization pushes voltage down → customer devices respond → load shape changes → voltage changes again.
• Price signals shift demand → demand response fires → wholesale price changes → signals change.
• Predictive maintenance flags a transformer → crews reprioritize work → loading shifts → failure risk changes.

These are cybernetic systems in the purest sense: measurement, decision, action, and consequences feeding back into the next measurement.

Wiener also insisted that the purpose of automation should be human-centered—not because it’s “nice,” but because human incentives, attention, and judgment are part of the control loop. If you ignore them, the system becomes brittle.

Myth-bust: “More automation means fewer humans”

Most companies get this wrong. In grid operations, automation doesn’t remove people—it changes the failure modes.

You get fewer routine decisions, but more “uh-oh, what’s happening?” decisions. When things go wrong, they go wrong faster (because the system is optimized for speed). That means:

• Operators need clearer model intent, not just alerts.
• Engineers need traceability, not just accuracy.
• Leadership needs governance that fits real operational tempo, not quarterly AI ethics workshops.

That’s Wiener’s point in modern language: control without understanding isn’t control.

“Every web conceals its spider”: the real risk with AI control loops

The poem’s “spider” is the hidden agency inside a system: who set the objective function, what data got excluded, which constraint got relaxed, what failure got quietly accepted.

In energy AI, the “spider” usually shows up as one of these:

1. Unexamined objectives (optimize cost… but whose cost?)
2. Opaque model boundaries (the model doesn’t “see” critical constraints)
3. Human workaround loops (operators learn to ignore or game the tool)
4. Automation bias (people defer to the model even when they shouldn’t)

Here’s a utilities-specific example I’ve seen variations of: a dispatch optimizer for grid-connected battery energy storage systems (BESS) maximizes market revenue. It performs well—until a heat wave. Then it keeps cycling aggressively because the price signal says “go,” while thermal conditions and auxiliary load penalties quietly pile up. Revenue looks great on paper, but degradation accelerates and availability drops right when reliability risk is highest.

That isn’t “AI being bad.” It’s a feedback loop designed around the wrong state variables.

What responsible AI looks like in utilities

Responsible AI in energy and utilities is not a poster on a wall. It’s a set of design choices:

• Objectives: reliability and safety constraints are first-class, not afterthoughts.
• Observability: operators can see why the tool recommends an action.
• Controllability: humans can intervene quickly, and the system degrades gracefully.
• Accountability: you can audit decisions after the fact.

Wiener would call this “the human use” of automated systems: use machines to extend judgment, not replace it.

From cybernetics to smart grids: where AI actually helps (and where it bites)

AI in utilities is often described as “smart grid AI” or “grid optimization,” but it’s more useful to map it to automation layers—the same way we think about robotics and industrial automation.

Layer 1: Perception (sensing and interpretation)

This is where AI is most mature and least scary:

• Outage prediction from vegetation, weather, and historical faults
• Fault detection and classification on feeders
• Asset health estimation (transformers, breakers, cables)
• Customer load disaggregation (where permitted)

The risk here is mostly data integrity: drift, bias, and false confidence.

Layer 2: Planning (recommendations)

This is the “copilot” layer:

• Crew routing and work prioritization
• Switching sequence recommendations
• DER hosting capacity analysis
• BESS dispatch recommendations with degradation-aware constraints

This layer succeeds when it’s designed like good automation in robotics: clear handoffs, clear boundaries, and operator trust earned over time.

Layer 3: Control (closed-loop autonomy)

This is where feedback loops become existential:

• Autonomous voltage regulation with inverter-based resources
• Automated feeder reconfiguration
• Real-time frequency response coordination across storage fleets
• Microgrid islanding and resynchronization logic

Closed-loop control is powerful—and unforgiving. If you don’t model the human role in that loop, you end up with the operational version of a robot that moves fast but can’t explain why it’s in the hallway.

A practical framework: the “Operator-Centered Feedback Loop” checklist

If you’re deploying AI in grid operations, substation automation, or BESS management, use this checklist to keep Wiener’s principles practical.

1) Define the loop explicitly

Write it down:

• Measured signals: voltage, frequency, temperature, SOC, breaker status, market price, etc.
• Decision output: setpoint, schedule, alarm, switching plan
• Actuators: inverters, breakers, tap changers, dispatch instructions
• Update frequency: milliseconds, seconds, 5-minute, day-ahead

If you can’t describe the loop, you can’t govern it.

2) Put reliability constraints above optimization goals

A rule that holds up in real utilities: optimize inside constraints, not past them.

Examples of constraints that should be non-negotiable:

• thermal limits (asset and ambient)
• protection coordination boundaries
• minimum reserve margins for black-start or contingency response
• cycling limits tied to warranty and degradation models

3) Make “why” available at operator speed

Explainability isn’t a research project; it’s an HMI requirement.

Useful “why” artifacts in control rooms:

• top 3 drivers behind a recommendation
• constraint binding indicators (what limit is active)
• confidence plus reason for low confidence (missing telemetry, drift, out-of-family conditions)

4) Design graceful degradation and manual recovery

Assume the model will fail—because it will.

Plan for:

• sensor dropouts
• communications latency
• bad weather creating out-of-distribution conditions
• adversarial or corrupted inputs

A strong pattern is fallback modes that are simpler, slower, and safer.

5) Close the loop with learning that doesn’t surprise people

If your model retrains, operators should know:

• when it changed
• what changed
• what performance shifted
• how to roll back

Surprise learning is how trust dies.

“We are old enough to be friends”: human-machine teamwork that works

Jones’s poem ends with a surprisingly operational truth: “We are old enough to be friends.” In utilities, friendship looks like calibrated reliance—knowing when to lean on automation and when to slow down.

In the AI in Robotics & Automation world, the best systems share a few traits:

• They reduce cognitive load without hiding critical context.
• They treat humans as part of the system, not an external auditor.
• They improve through measured iteration, not big-bang rollouts.

In energy and utilities, that means piloting AI tools in shadow mode, validating against disturbance scenarios, and training operators on failure behavior—not just normal behavior.

The goal isn’t “full autonomy.” The goal is a grid that stays stable when people are tired, weather is ugly, and assets are stressed.

What to do next if you’re implementing AI in utilities

If you’re trying to generate leads, the temptation is to pitch AI as the answer. I’d rather you sell the truth: AI is an amplifier. It amplifies whatever control philosophy, data discipline, and operational clarity you already have.

Concrete next steps that actually move projects forward:

1. Inventory your control loops (grid, DER, BESS, outage) and label which ones AI touches today.
2. Choose one high-value loop where better perception or planning reduces risk (not just cost).
3. Run a joint design workshop with operations, protection engineers, IT/OT security, and vendors.
4. Set measurable success metrics that include reliability and human factors (override rate, alarm fatigue, time-to-recovery).

Wiener warned that freedom is “always a contingency.” In grid terms, reliability is contingent on decisions made under uncertainty. AI can help—if you treat feedback loops as the product, not an implementation detail.

Where are you comfortable letting automation close the loop in your organization—and where do you still need a human hand on the dial?