The Algorithm That Found Flint's Lead Pipes

Machine learning is now one of the most effective tools for finding lead water pipes before they poison anyone, a use case proven in Flint, Michigan, where University of Michigan researchers built a model that predicted which homes had lead service lines with roughly 97 percent accuracy according to contemporary reports. This matters because millions of lead service lines still carry drinking water to homes across the United States, and the water crisis in Flint showed what happens when lead corrodes into what flows from the tap. The pipes are underground, records are chaotic, and digging up every line to check is prohibitively expensive. Prediction turned guesswork into a ranked work plan.

The Flint crisis began in 2014 when the city switched its water source without proper corrosion control, allowing lead to leach from aging pipes into the drinking water of a city of roughly 100,000 people. By the time the danger was acknowledged, residents had been drinking contaminated water for more than a year, and children's blood lead levels had risen.

A City That Did Not Know Where Its Lead Was

When Flint committed to replacing its lead and galvanized steel service lines, it confronted an absurd problem: nobody knew which of the city's roughly 55,000 homes had them. Property records were incomplete and contradictory. Hand drawn water department maps spanning decades listed pipe materials inconsistently, and in many cases not at all. The naive approach, digging up every line to inspect, would have cost the city years and hundreds of millions of dollars it did not have, with most of the digging wasted on homes that turned out to have copper pipes.

Jacob Abernethy, Eric Schwartz, and students at the University of Michigan proposed a different approach: treat it as a machine learning problem. Using every excavated pipe as a labeled example, their model estimated the probability that each home in the city had a lead or galvanized service line, based on dozens of factors including the age and size of the home, property records, neighborhood, and the hand drawn maps of the water department.

Hydrovacing Down a Ranked List

The predictions were put to work. As The Atlantic reported in 2019, Flint's Fast Start program used hydrovac trucks, which expose pipes with jets of water rather than backhoes, to verify pipe material at homes the model prioritized. Each dig updated the model, which continuously retrained on new results and refined the citywide map. Reporting on the effort cited a success rate around 97 percent for the algorithm's predictions, meaning crews digging where the model sent them found what it expected nearly every time.

The university team estimated at the time that roughly three out of four Flint homes had lead or galvanized lines, a figure that defined the scale of replacement work. The work became a template: the researchers spun their methods into a company, BlueConduit, which has since brought predictive lead pipe mapping to dozens of other cities facing the same invisible inventory problem.

A National Inventory Deadline

The problem extends far beyond Flint. The Environmental Protection Agency estimates that millions of lead service lines remain in use across the United States, concentrated in older industrial cities and low income neighborhoods. New federal rules require water systems to inventory their service lines and have pushed replacement timelines forward, and federal infrastructure funding has made money available for the digging. The binding constraint, as in Flint, is knowing where to dig.

That is why predictive tools have spread. BlueConduit and similar analytics approaches are being adopted by cities from the Midwest to the South, each deployment starting the same way Flint's did, with messy records, a first set of digs, and a model that gets sharper with every shovel.

The Lesson of Flint

Flint's crisis was a failure of governance and chemistry, and no algorithm can undo the harm done to the children who drank that water. What the machine learning work demonstrated is narrower but important: in the recovery, when a city must act with incomplete information under enormous time and budget pressure, AI can convert buried uncertainty into a prioritized plan. The pipes could not hide from a model that learned the city's own records better than the city knew them.

Lead in drinking water is a slow emergency that predates AI and will outlast any single technology. But the combination of mandated inventories, funded replacement, and statistical tools that tell crews where to dig first represents genuine progress for the millions of households still connected to the twentieth century's most consequential plumbing mistake.