Artificial intelligence has documented uses in water work, mainly in flood forecasting and in finding leaks in pipe networks, but it does not supply the pipes, treatment plants, or toilets that most of the unserved population lacks. The World Health Organization and UNICEF reported on August 26, 2025 that 2.1 billion people still lacked safely managed drinking water and 3.4 billion lacked safely managed sanitation. Evidence for AI exists for specific tasks, and each result comes with limits that matter before anyone calls it a solution.
Key takeaways
- The global gap is large: 2.1 billion people without safely managed drinking water and 3.4 billion without safely managed sanitation, in the WHO and UNICEF report covering data through 2024.
- A peer-reviewed flood model published in Nature in 2024 reached reliable forecasts up to five days ahead in watersheds without river gauges.
- A Bangkok field trial of an acoustic leak-finding tool reported accuracy above 90 percent, with novice operators matching experts.
- In Canada, the 39 active long-term drinking water advisories on reserves are tracked as construction and design projects, which software alone does not complete.
The size of the gap AI would have to help close
The WHO and UNICEF joint monitoring program publishes the standard global count. Its 2025 report puts the number of people without safely managed drinking water at 2.1 billion, about one in four, including 106 million who drink straight from untreated surface water. For sanitation, 3.4 billion lack safely managed services, and 354 million practice open defecation. Another 1.7 billion lack basic hygiene services at home. See the WHO and UNICEF fast facts release for the full figures.
These are access gaps, and closing them mostly means building and operating infrastructure: wells, treatment, pipes, latrines, sewers, and the trained staff and budgets to run them. A forecasting model can warn a district of a flood, but it cannot treat the water afterward.
Flood forecasts where no river gauge exists
Flood warnings normally depend on a long record from a gauge in each river. Many basins in low-income regions have none. Nearing and colleagues, working with Google, published a model in Nature in 2024 (DOI 10.1038/s41586-024-07145-1) that predicts extreme river flows in ungauged watersheds. The authors report reliability at up to five days of lead time, comparable to or better than same-day nowcasts from the Copernicus GloFAS system, and accuracy for five-year return period floods similar to or better than existing accuracy for one-year events. The model feeds an early warning service that publishes free forecasts in more than 80 countries.
Limits are worth stating. A forecast helps only if a warning reaches residents and an evacuation or protection plan exists, and the paper evaluates prediction skill, not lives or property saved.
Finding leaks before they drain a utility
Water lost from pipes before it reaches a customer is called non-revenue water, and utilities hunt for leaks by listening to pipes, a slow task that depends on operator skill. A 2022 paper in Results in Engineering by researchers working with Thailand's Metropolitan Waterworks Authority describes a cloud system that classifies leak sounds with a neural network and guides a mobile app. The authors compared three model types, found a deep neural network matched a convolutional one with a simpler structure, and report field-trial accuracy above 90 percent, with results from novice operators similar to those of experts.
The paper does not report how much water a utility saved over time, and a classifier trained on one city's pipes and soil may behave differently elsewhere.
What the Canadian record shows
Indigenous Services Canada tracks long-term drinking water advisories on public systems on reserves. As of October 21, 2026, its advisory tracker showed 39 active advisories in 37 communities and 158 lifted since November 2015. In 2026 to that date, eight advisories were added and eight lifted. The tracker sorts active cases by project stage, from feasibility study and design to construction and completed work awaiting lifting. That framing locates the bottleneck in capital projects, operators, and maintenance. No source reviewed for this article shows AI shortening those steps.
Reading claims about AI and water
| Use | What is documented | What is not shown |
|---|---|---|
| Flood forecasting | Peer-reviewed skill at up to five days in ungauged basins; operational in more than 80 countries | Outcomes for people after warnings |
| Acoustic leak finding | Field accuracy above 90 percent in one Bangkok study | Water saved, cost, results in other cities |
| Access to safe water and sanitation | The gap is measured in billions of people | Any AI tool that closes it without construction and operation |
A useful test for any claim is to ask what was measured, where, against what comparison, and who paid to keep the system running afterward. The table is a reading aid for this article, not a ranking of technologies.
FAQ
Can AI provide clean drinking water by itself?
No. Clean water depends on a source, treatment, and a distribution system. AI can help with tasks around them, such as forecasting floods or locating leaks, but the physical system has to exist and be maintained.
What is non-revenue water?
It is water a utility produces but does not bill, through leaks, metering errors, or unauthorized use. Leak detection tools target the leakage part.
Does the flood model work everywhere?
The authors report that it extends skill to ungauged basins, which was the aim, and the system covers more than 80 countries. Performance still varies by region and event, and warnings need local communication channels to be useful.
Is there evidence that AI has lifted Canadian drinking water advisories?
None was found in the official sources reviewed. The tracker describes advisories moving through engineering and construction stages.
Related reading
- AI in agriculture and its measured results
- AI, sustainability, and the environment
- Timefold as a planning and scheduling solver
- How automated Telegram bots work
More in AI and Society.
Last verified: October 2026
Sources
- 1 in 4 people globally still lack access to safe drinking water, WHO and UNICEF, August 26, 2025
- Progress on household drinking water, sanitation and hygiene 2000-2024, WHO and UNICEF, 2025
- Nearing G. et al., Global prediction of extreme floods in ungauged watersheds, Nature 627, 2024
- Vanijjirattikhan R. et al., AI-based acoustic leak detection in water distribution systems, Results in Engineering 15, 2022
- Long-term drinking water advisories, Indigenous Services Canada, data as of October 21, 2026