Water Efficiency

Lead Pipes and Machine Learning

How far are we willing to trust technology?

  • Email This Post Email This Post

More than three years after it was determined that lead from distribution pipes in Flint, MI, was leaching into drinking water and affecting residents’ health, thousands of homes in the city still have lead pipes. The pipe replacement effort has been complicated by both politics and residents’ mistrust of a machine-learning model designed to help determine which homes had pipes made of lead.

The Metropolitan Water Reclamation District of Greater Chicago is seeking a visionary Executive Director. The District is an award-winning wastewater agency which has been a leader in protecting the Chicago area water environment for over a 120 years. For information and to apply, click here or contact [email protected]The District is an Equal Opportunity Employer.

A grave challenge emerged when city administrators launched an excavation and replacement program in 2016. The city’s record system consisted of thousands of old cards used to catalog public infrastructure, making it difficult to determine which homes have lead pipes. Officials soon realized that Flint not only had a lead pipe problem, it also had an information problem.

A group of volunteer computer scientists, led by Jacob Abernethy of Georgia Tech and Eric Schwartz of the University of Michigan, also recognized a prediction problem—a sequential decision-making process was needed to indicate where to dig next, under uncertain conditions. The known variable was that lead pipes were most likely to be found in postwar homes and least likely to be found in newer homes. So the team created a machine-learning model to help narrow down the search for homes that were most likely to have lead pipes. The results of each new dig could be fed back into the model, improving its accuracy.

The artificial intelligence was designed to help city crews dig only where pipes were expected to need replacement in order to expedite the process. And it worked. According to an Atlantic article, workers inspected 8,833 homes, and of those, 6,228 homes had their pipes replaced—a 70% rate of accuracy.

Flint residents complained, however, that their neighbors’ pipes were excavated and replaced while their pipes were left in the ground. It seemed inequitable. Mistrust of the machine-learning model fed this concern. And project managers could not simply tell property owners to trust the AI modeling program with the health of their families at stake.

After the city contracted AECOM, to speed up the project in 2018, citizens began noticing a decline in accuracy. The new contractor had scrapped the AI program and was struggling to locate the sinister pipes efficiently. “As of mid-December 2018, 10,531 properties had been explored and only 1,567 of those digs found lead pipes to replace. That’s a lead-pipe hit rate of just 15 percent, far below the 2017 mark,” reported The Atlantic.

As political pressure mounted, city administrators and Flint Mayor Karen Weaver demanded that the AECOM dig across the city’s neighborhoods and replace the infrastructure of every house on selected blocks, rather than picking out the homes likely to have lead because of age or property type. While the decision has undoubtedly increased the replacement program’s cost exponentially, it may also offer peace of mind to the beleaguered residents of Flint.

Negotiations are also currently underway to incorporate the machine-learning modeling program in AECOM’s future work. But this complex issue raises an interesting question: how confident are we in AI technology? If your family’s health were at stake, as is the case for many Flint residents, would you trust a machine learning model? WE_bug_web

Related Posts


  1. One obvious explanation for how the AI performed would be that when they started, there were sufficient houses with lead pipes that the tool simply couldn’t miss – but that as the effort proceeded it was harder and harder to guess which ones had lead?

    I am being deliberately extreme here, to make a bigger point. AI tools tend to be “black box” in nature which means that when equity and fairness of public services are at issue, the inability of the tools to explain themselves, or indeed of their operators to explain them satisfactorily, is a major problem. AI has a big role in government and in infrastructure management, for sure, but it needs to address this!

  2. Wow, what happened to good ole sleuthing- check the inside service line coming into the property- if its lead, replace it. If it’s suspect then trace the line using good ole line locating talent and pothole the corp at the main looking for a copper line or lead- if lead then dig it up. There are many companies with the talent to do this- maybe even the city has the staff with the know how. Both of the methods discussed in the article seem expensive and way too complicated.

  3. “Both of the methods discussed in the article seem expensive and way too complicated.”
    I think it must be global warming!

  4. This is a Two Tier problem. Lack of trust in everyone from Governors office to the DEQ to the DHHS, the town council the mayor and water quality officials from Flint Water. This has nothing to do with technology. If you had over a 90% success rate . You still have a chance to miss some. This created the mistrust with this program. Case in point if your neighbour had his water feed line replaced you would think you would get yours replaced. There is no record on your house that shows your waterline was replaced in the 60’s or 70’s from lead to copper. So when you moved in after the fact you have no idea what your incoming water line is lead or copper. They even went after some failures to check at the curb line and then to check up by the home. That way they did not have to dig up the whole waterline to your house. But for the Mayor that was not good enough so she opted to go the more costly route because the city was not paying for this the state was. She was quick to spend others money.

Leave a Reply

Your email address will not be published. Required fields are marked *