More than six years after residents of Flint, Michigan, suffered widespread lead poisoning from their drinking water, hundreds of millions of dollars have been spent to improve water quality and bolster the city’s economy. But residents still report a type of community PTSD, waiting in long grocery store lines to stock up on bottled water and filters. Media reports Wednesday said former governor Rick Snyder has been charged with neglect of duty for his role in the crisis.
Snyder maintains his innocence, but he told Congress in 2016, “Local, state and federal officials—we all failed the families of Flint.”
One tool that emerged from the crisis is a form of artificial intelligence that could prevent similar problems in other cities where lead poisoning is a serious concern. BlueConduit, an analytics startup that says it uses predictive modeling to find lead pipes, offered promising results in Flint, but the city’s complex politics ended its use prematurely.
Now, four years later and 100 miles away, officials in Toledo, Ohio, facing concerns about lead pipes, want to use the technology. They hope to avoid the problems that surfaced in Flint by expanding community outreach and involvement. The Ohio Department of Health estimates that as many 19,000 children in the state have elevated levels of lead; children in Toledo tested positive for lead poisoning at nearly double the statewide rate, according to a 2016 report from the Toledo Lead Poisoning Prevention Coalition.
Lead is a crippling neurotoxin that can cause lifelong developmental problems in children and is toxic to adults even at low exposure levels. Last year, Toledo committed to a 30-year project to find and replace the estimated 30,000 lead pipes in the city. In October, a coalition including the city, local activists, and a nonprofit group received a $200,000 grant from the Environmental Protection Agency to use BlueConduit’s technology to find lead pipes.
Started in 2019 by Jacob Abernethy and Eric Schwartz, BlueConduit grew out of a University of Michigan project to identify lead pipes in Flint. Abernethy says the startup has contracts with organizations governing 50 cities to help replace lead pipes.
BlueConduit uses statistical techniques to predict which neighborhoods and households are most likely to have lead pipes, based on dozens of factors: the age of the home, the neighborhood, proximity of other homes where lead has been found, utility records, and more. Given a list of addresses, BlueConduit offers a ranking based on the likelihood of a lead service line. Cities can use the ranking to prioritize homes for excavations to examine the pipes.
“You can think of this not so much as ‘These homes have lead, these homes don’t,’” Schwartz says. “What we’re saying is, here’s the rank ordering of probabilities. And if your goal is reducing the amount of time people in the community are living with lead pipes, this is the way you should start going down the list.”
Alexis Smith, community program and technical associate at Freshwater Future, a nonprofit working with Toledo, says one appeal of Toledo’s approach is the input from residents, as well as the algorithms.
“It’s going to take the knowledge of homeowners and information not just from the city, but from the residents,” she says. “It really put our mind at ease that this isn’t just something that’s going to happen to us. We’re going to be working as a part of this program.”
Balancing tech and community perspectives is essential so residents don’t feel as though their concerns are secondary to algorithms. During the Flint project, BlueConduit’s model offered promising results, but it was met with a divided community and deep mistrust in leadership.
In 2017, Schwartz and Abernethy, professors of marketing and computer science, respectively, worked with Flint officials, who were initially impressed by the team’s predictive model. That year roughly 70 percent of the homes identified by the model turned out to have lead pipes. The city later signed a deal with AECOM, a Los Angeles-based engineering firm, that declined to use the pair’s predictive modeling. The following year, without the model, accuracy dropped to roughly 15 percent.