AI’s Least Glamorous Job

A toilet flush can carry an early warning that a virus is spreading. But a laboratory result alone cannot tell health officials how illness is affecting a community.

Predicting the next outbreak would be a dramatic use of artificial intelligence. My infectious-disease work brings laboratory and investigation records together, which requires checking what their dates and locations mean. That work suggests a less glamorous role for AI: helping health departments find those details. The wastewater example below is a proposal, not results from my study.

People with COVID-19 can release traces of the virus when they use the bathroom. Where homes connect to public sewers, wastewater flows toward a treatment plant. Operators collect samples before treatment and send them to a laboratory. The tests can reveal infections among people who never visit a doctor.

The results reflect the wastewater reaching that site. Homes with septic systems typically treat wastewater in a buried tank and surrounding soil instead of sending it to the plant. Local and state health officials therefore need to know which communities the samples represent before drawing conclusions about a wider area.

Even when samples are collected, getting results to health officials takes coordination. The Centers for Disease Control and Prevention describes limited staffing and longer shipping times for samples in small and rural programs. After testing, laboratories send results to health departments, where investigators must compare them with other disease reports.

Imagine a state health department’s disease investigator, an epidemiologist, comparing two reports. One shows rising coronavirus activity in town sewage; the other shows little change in COVID-19 hospital visits. Is one wrong?

Less time searching documents could leave investigators more time to investigate warnings, contact clinics, and explain risks to residents.

Not necessarily. In this example, the sewage report covers last week’s town samples; the hospital report counts this week’s visits across the county. They describe different periods and populations. Virus levels in sewage are not counts of sick patients.

To identify those differences, the investigator may need to read each report’s notes. I am interested in testing a large language model, an AI tool that answers questions about documents, to find the dates and areas covered. Asked when a sample was collected, it might distinguish “collected September 1” from “report issued September 5.” Using September 5 would make the sample look four days newer. Each proposed answer would appear beside its source passage.

Related research shows why those answers need checking. A 2026 study used a language model to match pregnancy and newborn health descriptions with standard medical terms. It found useful matches but also linked a condition present at birth with one acquired later. Accepting that match would change what the record says. The study tested medical terminology, not wastewater reports.

The epidemiologist would check each answer against the original report. They could ask the laboratory to clarify the sampling date or contact the local wastewater coordinator about neighborhoods served. I would instruct the model to flag missing information rather than guess.

Human review can fail, too. The World Health Organization warns that trust in automated answers can make people overlook errors. I would compare errors and time spent finding dates and locations with and without AI, including checks and corrections. A simpler alternative is for laboratories to label collection and reporting dates consistently on their forms. Sensitive records would remain in approved, protected systems.

AI cannot collect missing samples or permit agencies to share reports. But less time searching documents could leave investigators more time to investigate warnings, contact clinics, and explain risks to residents. The signal may start in a sewer. Its value comes from helping people decide what to do next.