The bias hiding in your meter data

In 2020, researchers at the University of Michigan and Stanford analyzed millions of smart meter readings and found something utilities hadn't systematically checked for: solar adoption models trained on historical customer data consistently underpredicted rooftop solar potential in lower income and majority Black neighborhoods, the same neighborhoods where interconnection queues (the process by which distributed energy resources get approved to connect to the grid) already moved slower. The data wasn't wrong. It was faithfully reflecting a past shaped by redlining, uneven grid investment, and unequal access to financing. The model just automated it forward.

This is the core problem of this lesson: AI systems in utilities don't need biased intent to produce biased outcomes. They just need biased history, and most meter, billing, and grid data has plenty of it.

Where bias enters: three high-stakes use cases

Disconnections and collections

Utilities increasingly use predictive models to prioritize which delinquent accounts to flag for disconnection or refer to collections, and which to route toward payment plans or assistance programs. Training data usually includes payment history, credit proxies, and neighborhood-level features.

The risk: ZIP code and payment history correlate strongly with race and income in the US, given decades of housing discrimination. A model optimizing for "likelihood of eventual payment" can end up disconnecting service more aggressively in historically underserved areas, even with race explicitly excluded as a variable, a phenomenon called proxy discrimination.

Dynamic pricingDynamic pricingAutomatically adjusting prices in real time based on demand, competition or user behaviour to optimise revenue, margin or conversion.View full definition → and demand response

Time of use rates and dynamic pricing (electricity prices that shift by hour based on grid conditions) are expanding across US states and EU markets under smart meter rollouts. AI models set price signals and enroll customers into demand response programs.

The risk: households without flexible schedules, often shift workers, renters without smart thermostats, or multigenerational homes with someone home all day, can't shift usage to cheap hours. They pay more, structurally, while wealthier households with home batteries and smart HVAC capture the savings. The model isn't biased against them by design; it just assumes a flexibility that not everyone has.

DER interconnection approvals

DER (distributed energy resources: rooftop solar, batteries, EVs) interconnection uses AI-assisted grid capacity models to approve or queue applications. If historical grid investment was uneven, as documented in numerous US utility service territory studies, the model will show "less hosting capacity" in underinvested areas, not because of physics but because of decades of deferred maintenance and thinner infrastructure.

The risk: this creates a feedback loop. Low investment leads to low modeled capacity, which leads to more denials, which leads to continued low investment.

Why this is a governance problem, not just a modeling problem

Regulators are starting to treat this as core utility governance, not a side technical issue.

• In the US, the Federal Energy Regulatory Commission (FERC) and state public utility commissions (PUCs) increasingly require utilities to disclose disconnection data by demographic proxy (income, geography) under state-level equity mandates, for example in California under the California Public Utilities Commission (CPUC) disconnection reporting rules.
• The EU AI Act (entered into force 2024, phased obligations through 2026 to 2027) classifies certain utility AI uses, including creditworthiness-adjacent scoring like disconnection risk models, as "high-risk AI systems," triggering mandatory risk assessments, human oversight, and documentation requirements before deployment. See the European Commission's AI Act overview for the official framework.
• The US doesn't yet have a federal AI-specific law for utilities, but the Federal Trade Commission (FTC) has signaled it will apply existing anti-discrimination and unfair practices authority to algorithmic decisions, and several state energy offices now require equity impact assessments before AI-driven rate design changes.

The practical implication for a utility: you cannot treat model deployment as purely an engineering sign-off. Legal, regulatory affairs, and customer equity teams need a seat before launch, not after a complaint.

The equity checks utilities should run before deployment

A workable pre-deployment checklist, adapted from emerging practice at investor-owned utilities and guidance like NIST's AI Risk Management Framework:

1. Disparate impact testing. Run model outputs (disconnection flags, price tiers, DER approvals) segmented by ZIP code, income band, and where legally permitted, demographic proxies. Compare outcome rates across groups, not just average accuracy.

2. Proxy variable audit. Check whether "neutral" features (credit score, ZIP code, meter age, prior outage count) are highly correlated with protected characteristics. A simple correlation matrix often surfaces this in an afternoon.

3. Counterfactual fairness spot checks. Take a real customer record, change only the ZIP code or name, rerun the model, see if the output flips.

4. Human review threshold for high-stakes actions. No fully automated disconnection. Require human sign-off above a defined risk score, especially for accounts flagged during extreme weather.

5. Community and regulator disclosure. Publish a plain-language model card describing what data trains the model, what it optimizes for, and known limitations, before regulatory filing.

A simple bias check in practice

Here's the kind of lightweight test a utility data team would run before greenlighting a disconnection model:

import pandas as pd

# outcomes: predicted risk score + actual demographic proxy (income tercile)
df = pd.read_csv("disconnection_model_outputs.csv")

disparity = df.groupby("income_tercile")["flagged_for_disconnection"].mean()
print(disparity)

# flag if lowest income tercile is flagged >1.5x more than highest
ratio = disparity.iloc[0] / disparity.iloc[-1]
if ratio > 1.5:
    print(f"WARNING: disparity ratio {ratio:.2f}, review before deployment")

This isn't sophisticated statistics. That's the point: most utilities could run a check like this today and haven't, because nobody owned the requirement.

What "good" looks like: emerging practice

A few real, verifiable signals of the field maturing:

• National Grid and several US investor-owned utilities have published equity frameworks tied to DER interconnection queue reform, partly in response to state regulatory pressure (New York's Reforming the Energy Vision proceeding is a well-documented example).
• The Department of Energy's Office of Economic Impact and Diversity has funded research explicitly on algorithmic equity in utility disconnection and low-income program targeting.
• Several EU member state regulators, under the AI Act's high-risk category rules, now require a documented "fundamental rights impact assessment" before deploying automated decision systems affecting essential services like electricity, gas, and water.

None of this eliminates bias. It creates a paper trail and a forcing function to check before harm compounds at scale, which is the realistic bar for governance in 2026.

🎬 [VIDEO: "How Algorithms Can Discriminate (Without Meaning To)" - youtube.com/results?search_query=algorithmic+bias+explained - a short explainer on how proxy variables create unintended discrimination in automated decision systems, useful background for the disparate impact concept above]

Key Takeaways

• AI models in utilities (disconnections, dynamic pricing, DER approval) can encode historical inequities like uneven grid investment or discriminatory lending, producing biased outcomes even without biased intent.
• Proxy discrimination is the central mechanism: variables like ZIP code or payment history correlate with protected characteristics even when those characteristics are excluded from the model.
• Regulatory exposure is real and growing: the EU AI Act classifies many of these systems as high-risk with mandatory impact assessments, while US state PUCs and the FTC are applying existing authority to algorithmic decisions.
• Practical pre-deployment checks (disparate impact testing, proxy audits, counterfactual spot checks, human review thresholds) are low-cost and can be run before launch, not retrofitted after a regulatory complaint.
• Governance ownership matters as much as the technical fix: equity review needs a formal seat in the model deployment process, not an afterthought triggered by a lawsuit or press story.