AI data centers,
air pollution and health

What Harvard Chan actually published

The April 2026 Harvard Chan article is a question-and-answer interview with Michael Cork, a postdoctoral researcher who completed his Harvard biostatistics doctorate in 2025. Harvard describes his work as estimating health and economic effects associated with air pollution from energy infrastructure, including power plants and AI data centers.

Cork explains an analytical sequence: estimate emissions, model where pollution travels, translate additional exposure into health risk, and value outcomes such as hospital admissions, respiratory illness, premature mortality and lost productivity. That sequence is more informative than treating a permit limit, electricity forecast or health statistic as if each answered the same question.

Sources for this sectionHarvard Chan interview ↗

The Loudoun County case described in the interview

Harvard Chan reports that Cork and the Piedmont Environmental Council analyzed a Vantage Data Center project in Loudoun County using on-site gas turbines. The interview gives an estimated range of $53 million to $99 million in annual health damages and 3.4 to 6.5 additional premature deaths per year across the affected region.

Those numbers are model outputs, not a count of deaths observed and attributed to the facility. Their meaning depends on the emissions scenario, expected operating pattern, dispersion model, population exposure, concentration-response functions and economic valuation assumptions used in the underlying analysis.

Sources for this sectionHarvard Chan interview and cited Loudoun case ↗

Why the method matters

A facility air permit can describe authorized equipment, pollutants, operating limits and testing obligations. It does not automatically show how often equipment runs. Cork identifies incomplete operating information as a central challenge, which means a careful analysis must distinguish actual operation, a projected scenario and maximum permitted operation.

The same discipline applies to health language. A modeled increase in population risk is not proof that a named individual became ill because of one facility. Population-health models estimate changes across many people and places; they are designed for policy comparison, not individual diagnosis or legal attribution.

• Permit limits are not the same as actual emissions.
• Modeled exposure is not the same as monitored exposure at a specific home.
• Estimated health damages are not observed hospital or mortality counts.

The Francesca Dominici connection—stated narrowly

The Harvard interview identifies Francesca Dominici as Cork’s doctoral adviser and says the two co-founded EmPower Analytics Group. Dominici’s broader research on causal inference, air pollution and population health supplies methodological context for this field.

That relationship does not justify saying that Harvard University, Harvard Chan, Dominici or Cork supports Public Evidence Project. It also does not justify describing every consultant analysis as a peer-reviewed Harvard study. Each underlying report should be evaluated on its own sources, assumptions and review status.

How this source should be used in Virginia

For a Virginia facility, the responsible path starts with the government record: identify the permitted source, equipment, pollutant limits, dates and operating conditions. A health-impact analysis can then be presented as a separate layer with its scenario and assumptions disclosed.

The Harvard Chan interview is therefore best used as evidence that health-impact accounting is an established analytical question and that one Loudoun project has been studied. It is not evidence that every Virginia permit creates the same risk or that a permit alone proves harm.