The fallout from Flint, Michigan’s lead-contaminated drinking water has been far-flung and long-lasting.
Some outcomes from the crisis have been a mix of positive and negative, making them all the more worthy of closer attention. That’s the case with a Google-driven, machine learning (ML) model that was developed to predict which homes in Flint were most likely to be served by lead infrastructure and therefore at the highest risk of drinking water contaminated with lead.
At first, it seemed like a positive solution stemming from a negative situation. Using the ML model, workers chose over 8,000 homes to inspect and 70 percent of them did end up needing replacement pipes. The effort was so successful that, at the end of the year, Flint gave a national engineering firm, AECOM, a $5 million contract to accelerate it.
“The artificial intelligence was supposed to help the City dig only where pipes were likely to need replacement,” The Atlantic reported. “Through 2017, the plan was working.”
But then the positive momentum was lost. Throughout 2018, fewer inspections were uncovering lead drinking water infrastructure. AECOM abandoned the artificial intelligence model that had been so successful, instead choosing homes to inspect based on a range of factors, like which homes had active water accounts, maps, and historic water cards. Mayor Karen Weaver ordered the firm to investigate every house on selected blocks, without any particular methodology.
“A statistical analysis submitted to the court shows this year’s method finds lead at a worse rate than randomly picking a home for the work,” according to a Michigan Live report at the time. “The current method’s 21 percent success rate for finding lead contrasts with last year’s success rate of 80 percent.”
Furthermore, the failed efforts cost Flint $19.4 million just to rebury copper pipes that they had excavated for no reason.
Following a reimbursement dispute between Michigan, the City of Flint, and AECOM, the engineering firm’s contract has been renewed with an additional $1.1 million allocated for the work and an apparent return to the algorithmic model that was so successful in 2017. Even with this return, however, local residents are understandably upset over the process that’s brought them to this point.
“They’re saying now they’ll use the predictive model to plan. They should’ve been using the model,” resident Arthur Woodson said during a public comment period on the extension at a city council meeting, according to Michigan Live. “Not one time have we been able to speak to AECOM. This is wrong.”
The story of Flint’s algorithm-driven lead service line replacement effort seems to contain some good and bad. On the positive side, computer scientists were able to drive an effective solution to a problem faced by cities all over the country. Still, dysfunction and confusion meant that the model was not utilized to its fullest potential and money and time were wasted while the city’s residents suffered.
Like the saga of Flint’s lead contamination crisis in general, this solution was anything but straightforward.