Main

Early scientific developments in potable water supply and wastewater disposal systems (encapsulated hereafter as ‘water systems’) enabled ancient societies to transform into urban metropolises beyond their riverside origins and build resilience to weather perturbations, including wet and dry spells1. For instance, the Nazcans constructed subterranean aqueducts to transport drinking water long distances while mitigating evaporation losses2, and the Indus Valley civilization constructed brick sewers to drain baths and latrines into isolated soak pits to mitigate exposure of people to sewage3.

While engineering feats have produced manifold benefits, some instances of technological innovation have resulted in ‘progress traps’: events where human ingenuity to solve a given problem inadvertently manifests unanticipated problems that outpace society’s—and technology’s—capacity to then solve them4. For instance, Ancient Rome’s lead plumbing was an engineering marvel, connecting its vast population to reliable water and wastewater networks, but its outflows have also been linked to contaminating harbour water with lead, potentially poisoning marine life and people5.

More recently, artificial agricultural irrigation has depleted groundwater aquifers6 and caused salination7. Wastewater treatment has inadvertently contributed to global warming, toxicity and acidification8. Desalination of sea water has caused air, marine and land pollution9. Innovations in adjacent sectors realizing short-term benefits have created longer-term problems for water resources, such as hydroelectric dams for energy production degrading aquatic ecosystems, biogeochemical dynamics and water quality10. Despite successful, and essential, innovations across water systems, our thirst for technology-based problem-solving has often locked us into chronic progress traps.

Today, some 25% of the global population lack access to clean water, 50% lack access to sanitation services and 30% lack access to hygiene facilities11. Anthropogenic climate change threatens to exacerbate these issues, with higher temperatures increasing water scarcity globally and extreme events, including storms, floods and droughts, damaging water systems infrastructure in developed nations and undermining water, sanitation and hygiene (WASH) efforts in developing nations12.

Against this backdrop, artificial intelligence (AI), and its subdivision of machine learning (ML), is the latest technological intervention proposed to solve problems across water systems by building climate resilience, enhancing performance of infrastructure and, in limited cases, assisting WASH efforts. However, burgeoning applications of AI may give rise to serious, and unexpected, problems that are underappreciated and must be responsibly, and pre-emptively, managed to avoid unintentionally undermining efforts to meet Sustainable Development Goal 6.

In this Perspective, we provide a balanced consideration of AI in water systems. We survey potential system-wide benefits of AI applications from catchment to end-user. Then we highlight potential systemic barriers, direct risks and exposures to cascading failures, which may prove catastrophic for communities. Finally, we propose a three-tiered risk mitigation approach, necessary to prevent proliferation of this currently nascent technology perpetuating the progress-trap phenomenon.

Here we define AI as a machine-based ‘intelligent agent’ capable of interacting with its environment with the aid of sensors, interpreting information for decision-making and autonomously taking actions to achieve goal-oriented outcomes via a human or robotic actuator, while ML refers to the subset of algorithmic models that learn and predict outcomes through passive observation of the environment13.

Benefits from catchment to end-user

Given the current lack of wide-scale deployment in the ‘real world’, we highlight presumed benefits from AI applications across three levels covering water systems end-to-end: (1) water supply (catchment level), including enhanced supply insights, catchment management and emergency response; (2) water distribution and disposal (network level), including efficient treatment and network infrastructure design, operations and maintenance; and (3) water demand (end-user level), including improved service availability, demand management and water justice (Fig. 1).

Fig. 1: Example benefits of AI for solving problems across water systems.
figure 1

AI has the potential to yield system-wide benefits ranging from enhanced catchment insights to optimized network efficiency to improved service for end-users.

Full size image

Beyond component-specific applications detailed below, advanced AI may also eventually be used to simulate, inform and optimize operating policy for whole water systems in line with integrated water resources management principles14.

Enhanced insights at catchment level

Over 10% of people worldwide are exposed to high and critical water stress, and climate change is expected to worsen this exposure in urban and rural areas alike15. As such, complete, high-resolution and reliable analysis of Earth’s natural water resources, hydrologic cycles and anthropogenic perturbations is essential to monitor and manage water supply16.

ML models may process big datasets, such as interferometric synthetic aperture radar imagery, and (re-)construct missing data17 to provide precise quantitative estimates of historical freshwater location and persistence, including withdrawal and replenishment, which aids forensic identification of water stress and scarcity drivers18. Complementary algorithms analysing satellite, drone, terrestrial and reservoir data may support real-time observation, anomaly detection, and swift short-term predictions of the hydrological cycle and weather patterns19. This includes quantity parameters, such as evapotranspiration20, condensation, precipitation21, infiltration, surface run-off, streamflow22, subsurface flow and soil moisture23, as well as quality factors, for example, nutrients such as phosphorus and nitrogen24 and minerals such as flouride25.

Such AI applications may be used for optimizing aquifer drawdown schedules to maintain water tables within sustainable limits26 and dam-filling schedules to minimize harm to aquatic ecosystems related to upstream and downstream hydrology alterations27. These enable automated detection of public health hazards, including pollution plumes, waterborne disease pathogens, such as protozoa (for example, giardia), bacteria (for example, dysentery), viruses and parasitic worms28, as well as eutrophication and harmful algal blooms29. Similarly, they may help detect illegal and harmful accidental activities, such as dumping or discharging hazardous chemicals into reservoirs or recreational water bodies30.

In emergency prevention, preparedness and response, by integrating real-time rainfall data with early warning systems and control technologies, AI may monitor reservoir inflows and communicate with dam telemetry to manage safe spillway releases31. Such technologies may intervene in human mismanagement of dams, mitigating events like the 2011 Brisbane flood, which resulted in damages of over AU$2 billion32. Meanwhile, intelligent ‘rain cloud to stormwater’ monitoring systems, utilizing remote-sensing and community observations, could improve flood disaster response33. Smart groundwater management, leveraging borehole sensors, satellite data and ML, can also improve resilience through early warning action in drought prone regions, such as Kenya34.

The drive for integrated catchment management necessitates understanding of water cycle dynamics within Earth system models (ESMs) to forecast short-period weather and long-period climate variability and associated influences on drought, desertification, storm surge, and water insecurity prevalence and intensity35. While still in its infancy, neural ESMs36 may improve understanding of underpinning physics, uncover hidden parameters and expand simulation options37.

On the basis of such forecasts, optimization algorithms could support sustainable, long-term catchment watershed and infrastructure planning. For instance, AI-enabled ESM outputs paired with geographic information systems could efficiently examine climate risks to dams and downstream damages associated with dam failure38. It may inform expansion of artificial water sources, such as desalination or recycled water39, where water scarcity is predicted. Furthermore, AI-enhanced hydraulic models, characterizing drainage basin water-flow pathways and velocities, flooding footprints and tidal levels, can hone river engineering, dam weir and wall upgrades, and storm surge barrier implementation40.

Optimized efficiency at network level

Considering growing population demands on water systems, AI may support development of new potable water, stormwater and sewerage infrastructure shaped by engineering innovations33 alongside effective management of ageing critical assets41.

Goal-driven AI systems, paired with virtual testing environments, may accelerate prototyping and testing of more sustainable materials42, such as graphene-based nanomaterial membranes for desalination43 or metal–organic frameworks for desert water harvesting44.

Optimization algorithms could be implemented to enhance reliability, longevity and expenditure minimization—critical for public utilities—in the design, construction and upgrade of treatment and distribution facilities45. AI-powered digital twins of cities46 may also help to rapidly scale water-sensitive urban design47, including prioritized placement of bioretention systems, buffer strips and swales, infiltration trenches, porous paving, sedimentation retention, artificial wetlands, rainwater harvesting systems, and aquifer storage and recovery systems.

Together, AI, Internet of Things devices and robotics may enhance operational efficiency across water and wastewater facilities. For instance, coagulation, flocculation, sedimentation, filtration (for example, reverse osmosis) and disinfection (for example, chlorination) processes in water treatment plants could be intelligently fine-tuned to meet drinking water standards by leveraging sensor data on microbial and contaminant content of inflows and outflows at any given time48.

Similarly, the performance of wastewater treatment plants could be advanced through self-adaptive unit processes, including preliminary screening and grit removal, primary phase-separation (for example, clarification), secondary (for example, fixed film) and tertiary treatment (for example, activated carbon) and disinfection (for example, ultraviolet light), based on the real-time organic and inorganic content of sewerage inflows and effluent discharge requirements49. In addition, intelligent anaerobic digesters may boost biogas and electricity production from by-product sludge50, while intelligent classification and sorting could maximize the efficacy and safety of biosolids for agricultural reuse51.

Smart distribution systems also provide advantages over traditional supervisory control and data acquisition systems52. ML models utilizing real-time data from network sensors could measure, monitor and optimize flow pressure and velocity to improve energy efficiency and operating costs by autonomously controlling and configuring water pump stations without human oversight53. Advanced computational systems may help to prevent harmful sewage overflows during wet weather events by fine-tuning utilization of storage in wastewater pump stations, pipes and manholes, and expedite alerts to clean-up crews where unplanned discharges do occur54.

Intelligent technologies may transform routine maintenance activities and reduce downtime. Network leakage results in the loss of 45 billion litres of potable water per day in developing countries, which is equivalent to hydrating 180 million people, and major pipeline leaks can short circuit high-intensity cables, posing a lethal threat to people55. Predictive analytics, supported by sensors and cloud computing, can detect anomalies, pinpoint locations and prioritize the severity of leaks to accelerate isolations and repairs in real time56, with efforts to forecast pipe deterioration57 and resolve algorithm transferability across heterogeneous pipelines58 already improving the accuracy of AI-enabled leak identification applications. ML models paired with traditeional CCTV data, used for image classification, object identification and semantic segmentation, may similarly be implemented to predict, diagnose and fix wastewater network defects and blockages59.

Furthermore, AI could extend asset life and optimize capital expenditures by automating maintenance operations, such as the cleaning of ultrafiltration membranes in treatment plants, and designing predictive upgrade schedules based on historical and real-time asset condition assessments60.

Improved services at end-user level

At the community level, computational intelligence could contribute to more sustainable, resilient and equitable access to water systems. For instance, AI-based analysis of historical, smart meter, satellite imagery and water consumption forecast data may inform management of conflicting sectoral and transboundary demands with precise allocations as well as monitor withdrawal compliance61.

The agriculture sector is responsible for 70% of annual freshwater withdrawals, of which 60% (that is, 42% of global total) is wasted62. Targeted AI applications could help reduce this unnecessary consumption. AI may enable rapid experimentation in ‘virtual farms’ to determine minimum irrigation volumes and schedules to maximize crop yields under various conditions63. Such programmes implemented alongside digital twin and robotic technologies could enable precision farming with smart irrigation systems64. Autonomous processing of satellite or drone hyperspectral imaging, enabled by computer vision and ML algorithms65, may provide detailed maps of soil moisture and crop conditions, which water authorities could use to monitor irrigation shortfalls or excesses and adjust supply allocations accordingly66.

At the household level, smart water-saving devices, such as intelligent toilets, taps and sprinklers, may curtail household water consumption, while smart meters coupled with predictive demand and pricing analytics could provide incentives to drive behavioural change towards water conservation67. Furthermore, AI may control safe, decentralized potable water, stormwater and sewerage systems, such as automated rainwater tanks, domestic water recycling and home biodigesters68. Household units comprising real-time fluorescence sensors coupled with ML can accurately predict and intervene faecal contamination of drinking water in line with World Health Organization risk levels to prevent disease outbreaks common in both high- and low-income countries69.

While most of these applications rely on established water systems infrastructure, AI also has potential to improve water justice. Neural ESMs and optimization algorithms could support international development agencies and governments in determining where to prioritize investment in WASH efforts to effectively address the most pressing problems while building climate resilience.

Intelligent water technologies, including off-grid facilities, such as solar-powered ‘water ATMs’70, and portable devices, such as ‘smart handpumps’71, could be distributed and monitored remotely to improve safe water access, especially for women and girls. The proliferation of personal smart phones in developing nations72 could also enable mass communication of drinking water contamination or educational information about menstrual and hygiene practices similarly to that during the COVID-19 pandemic73. Portable AI systems may be trained to evaluate drinking water quality, based on free residual chlorine content, to prevent outbreak of waterborne diseases in humanitarian settlements74.

Barriers and systemic risks

As evidenced throughout history, technological problem-solving can bring about unintended consequences, which may prove more challenging than the original problem. Given the potential proliferation of AI across water systems, as highlighted above, it is important to understand the risk landscape. To this end, we highlight issues that may undermine potentially beneficial applications of AI in water systems, including: barriers related to infrastructure and human capital; direct risks related to design errors and malicious use; and indirect exposure to cascading failures (Fig. 2).

Fig. 2: Example risks of AI across water systems that may lead to progress traps.
figure 2

Infrastructure and human capital barriers, direct risks related to design errors and misuse and indirect risks related to cascading system failures may undermine the potential benefits of AI if not managed responsibly.

Full size image

Infrastructure and human capital barriers

AI is only as good as the systems into which it is integrated and the people responsible for its development. Many of the potential AI applications outlined above require established water systems infrastructure, supporting information and communications technology (ICT) infrastructure and domain expertise. In this vein, we highlight below anticipated instances of how infrastructure and human capital requirements may create technical and socioeconomic barriers that limit the deployment of, and give rise to unintended problems associated with, AI in the water sector.

The lack of foundational and safely managed infrastructure, including dams, treatment plants, pipes, toilets, showers and taps—currently leaving one in four people without clean drinking water and two in four people being without adequate sanitation services—will undermine the capacity of AI to address these water systems deficiencies in low-income regions, thereby precluding the most vulnerable populations from its associated benefits75.

Even in developed countries with well-established water systems infrastructure, the complexity and cost associated with integration of advanced digital technologies, which are necessary to support AI applications, across the water sector may limit the feasibility of deploying AI in the short term76. Indeed, the water industry brands itself as slow and painful, more so than other sectors, when it comes to innovation owing to its long project timelines, investment constraints and conservative nature77.

Currently, AI applications must be tailor-made to the specific context, and further algorithmic development is needed for accuracy in most situations. While there may exist successful real-world applications in other sectors, or demonstrations of water-system-specific algorithms in the literature, these are unlikely to be readily transferable to water systems in practice.

Wide-scale deployment of AI will require human capital with both AI and water sector subject matter expertise. As such, shortfalls in digital literacy among water sector workers and consumers, and lack of human capital at non-governmental organizations (NGOs) in WASH contexts, may further impair access to the potential benefits of AI in the water sector78.

Beyond limiting the deployment of AI in the water sector, particularly in developing regions, the unequal distribution of these barriers may also give rise to unexpected problems as AI begins to proliferate in more digitally capable hubs around the world. For instance, ‘digital divides’ in human capital, where high-skilled labour forces are advantaged while low-skilled labour forces are disadvantaged, could reinforce global inequalities76.

In addition, deployment of AI where barriers only partially exist but have not been fully resolved, such as where water sector practitioners develop skills sufficient to implement AI systems but insufficient to effectively identify and correct errors and malfunction, could undermine any potential benefits by giving rise to serious consequences, as outlined below.

Direct risks from design errors and misuse

Technical robustness, governance and ethics of AI79, which are increasingly explored in other sectors such as agriculture80, engender a distinct risk landscape in the water systems context. We provide below possible examples of how errors and biases in data and algorithmic models, including goal misalignment81, as well as increased exposure to misuse by malicious actors, may see potential applications of AI in the water sector cause substantial social, economic and environmental harm.

At the catchment level, AI applications necessitate thorough knowledge of highly complex Earth system processes, including the water cycle and climate change. Extraction of flawed hydrological data from satellite feeds or weather forecast data from a neural ESM by an unsupervised ML model optimizing water allocations in high-competition regions could result in unexpected shortfalls of water supply for human consumption. An AI algorithm tasked with minimizing damages in the event of a dam failure could inadvertently prioritize reduction of economic losses at the expense of human life if it were accidentally programmed to optimize for wrong or overly narrow goal ranges81.

At the network level, mistakes in the programming of intelligent wastewater treatment plant models, such as for automated maintenance of biological secondary treatment units, could cause a process crash, leading to downstream discharge of untreated effluent or upstream network overflows82. Such an event could expose human and marine life to untreated sewage as well as result in environmental discharge fines for the operator. Meanwhile, goal-oriented AI for optimizing water pipe maintenance based on failure probability and damage prediction could inadvertently undermine the water security of low-income populations by prioritizing infrastructure in more affluent areas83.

At the end-user level, errors in training datasets, a sensor fault or failures in algorithmic generalizations of an ML model directing recycled water to different end-uses based on real-time quality data could result in a public health crisis if non-potable water, or worse, water contaminated with pathogens, is distributed to households for consumption. AI tools collecting household water consumption data may raise privacy concerns over use profiling, while AI-enabled water-demand-reduction applications may inadvertently employ biased ‘micro-nudging’ leading to undemocratic water access, undermining the dignity and autonomy of at-risk populations83.

Incorporation of AI into water systems collectively heightens the risk of network-wide failures. Specifically, overreliance on AI, either in critical components of water systems or through high coupling between water systems components, could lead to systemic risks, where the isolated risks described above could potentially result in compromising an entire utility’s assets and services.

Furthermore, water is already the subject of intra- and international geopolitics and corporate competition—where appropriation of fresh water is associated with agricultural land grabbing estimated at 310 billion cubic metres of green water (that is, rainwater) and 140 billion cubic metres of blue water (that is, irrigation water) per year84—and cybersecurity is of growing concern given the recent increase in events of system compromise across the sector85. While previous generations of cyberattacks, including distributed denial-of-service (DDoS), ransomware, structured query language (SQL) injection and Trojan horse, were disruptive, the presence of embedded AI with minimal human oversight may provide hackers the opportunity to take full control of highly interconnected systems86.

Such network-wide failures may put entire communities at risk of water insecurity and could quickly translate to humanitarian crises and conflicts due to the relatively localized nature of water resources that—unlike the established global supply chains of energy and food resources—are not easily substituted or traded en masse internationally87.

Cascading system failures

The direct failures of AI within water systems highlighted above may indirectly cascade into local and regional catastrophes outside the water industry. In highlighting such instances below with presumed examples, we note that these failures could occur independently but that flawed AI, alongside lack of human oversight, may exacerbate the frequency and severity of such indirect risks.

Heavy reliance on AI could create fragile interdependencies between critical infrastructure systems. Scenarios could include three-way coupling where cloud computing underpins AI in water systems and energy systems, while water cooling is needed for cloud computing and electricity generation, and grid power is required for water systems and data centre operation. Such tight connectivity amplifies the risk of accidental failures or malicious cyberattacks cascading across systems and makes recovery from otherwise isolated events substantially more challenging.

Notwithstanding the above, seemingly successful applications of AI in water systems may have unexpected negative repercussions. Inadvertent socio-ecological consequences may occur where an AI-enabled digital twin optimizes the processes of a seawater desalination facility but does not accurately account for brine effects on ecosystems at the discharge point, resulting in damage to the marine environment and biodiversity88. Similarly, energy and food security problems could arise where AI models implemented by the water industry are biased such that water–energy–food nexus trade-offs are not appropriately represented89.

Furthermore, while ML advances can reduce computational energy demand90, expanded use of inefficient AI systems may increase the power intensity of data centres, thereby increasing water usage in liquid cooling technologies and greenhouse gas emissions that feedback to undermine water security91.

Responsible AI to prevent progress traps

To ensure that potential applications of AI in water systems realize intended benefits and do not unintentionally perpetuate progress traps, we make three sets of recommendations for the water industry to safely deploy this nascent digital technology. Detailed in turn below, the first addresses gaps in infrastructure and digital literacy, the second outlines technical mechanisms for trustworthy AI, and the third proposes a six-layer framework to guide benefit and risk assessment of AI applications across water systems in practice.

Address gaps in infrastructure and literacy

Where foundational catchment, treatment and distribution assets and hygiene facilities are lacking, there is little potential for AI to solve current water systems deficiencies. Governments, development funds, philanthropists and start-ups seeking to further WASH efforts in developing countries must consider social equity and economic efficiency when evaluating AI applications in place of, or complementary to, ‘brick and mortar’ projects92.

In developed countries, water utilities should ensure that adequate ICT infrastructure, such as sensors and cloud computing capabilities, are accounted for in AI application planning. Furthermore, water utilities must develop clear strategies and architectures, such as application programming interfaces, for systems integration and interoperability, given the need to maintain legacy infrastructure, including both physical structures and electromechanical equipment as well as heterogeneous data siloes containing operations critical information, alongside new digital technologies93.

Where AI applications are deemed appropriate, the water industry will need to manage the upskilling, reskilling and new skilling of its workforce, to ensure that the sector is equipped with the human capital necessary to design, operate and manage AI systems. Professional associations and trade unions should engage with academic institutions and NGOs with expertise in AI to develop new educational courses and certifications.

To this end, it is important that AI in water systems is explainable. On the one hand, explainable AI is necessary for domain experts to meaningfully confirm, challenge and transfer knowledge across use-cases94. Meanwhile, the move from research-based settings to practical applications makes deployment and communication of white box (that is, algorithms providing understandable results), rather than black box (that is, algorithms that can hardly be understood even by domain experts), algorithms by AI experts essential to ensure that water industry practitioners and civilian end-users lacking expertise can trust and interact with AI in water systems by understanding its functionality95.

Finally, the water industry must develop and enact sectoral-specific legislation, regulations and policies fit for purpose in handling the nuances of AI in water systems, especially when it comes to matters of technical standards and transparency, human agency and oversight, safety and security, accountability and liability, and diversity and inclusion. Where such governance frameworks are lacking, utilities looking to implement AI across their water systems should give due consideration to evolving issues such as insurances and liabilities related to failure of AI systems.

Establish mechanisms for trustworthy AI

At the current time, our understanding of AI is in a state of rapid development, so the water industry must keep abreast of issues related to technical robustness, governance and ethics as the field of AI safety evolves (for example, see the European Union’s ethics guidelines for trustworthy AI)96. Here we outline institutional, software and hardware mechanisms that the water industry should develop and maintain in its ‘toolbox’ to ensure responsible deployment of AI across water systems.

Institutional mechanisms shape knowledge, incentives and accountability97. Routine red team exercises, where cybersecurity experts are internally engaged to find vulnerabilities in AI systems, should be conducted by water utilities to keep ahead of malicious actors seeking to compromise critical water systems infrastructure. Bug, bias and safety bounty schemes can also be used to incentivize external stakeholders and beneficiaries to disclose problems with AI systems in practice. Such schemes may be especially useful when it comes to mapping systemic exposures to cascading socio-ecological consequences and water–energy–food nexus trade-offs. Beyond these exercises, as the water industry gains real-world experience with AI, collaborative cross-sector knowledge banks of implementation best practices, safety incidents and lessons learned should be maintained.

Software mechanisms address specificities, understanding and oversight of AI systems themselves97. The water industry should work with academic experts to establish design standards, interpretability manuals and user-testing methods to ensure reproducibility, privacy preservation and verification. Processes of human-centred design, safe by design and secure by design could help mitigate several of the risks outlined above. Cross-sector knowledge banks of validated standard AI source codes, for typical applications such as pipe leakage detection and treatment process optimization, could be maintained to accelerate best practice. Water utilities must also maintain audit trails of problem definition, design, development and operation of AI systems and should have these traceable logs analysed by expert third-party auditors to maximize the capture of incidents and lessons learned.

Hardware mechanisms confront the capability, accessibility and reliability of physical resources98. Water utilities must ensure that AI applications incorporate fail-safes enabling automated or human-initiated shutdown and workarounds to mitigate potential catastrophes caused by malfunctioning or compromised systems. Quality control inspectors should be engaged to regularly measure and report on performance of AI and smart cyber-physical systems on a case-by-case and water-system-wide basis. Water sector stakeholders should also consider establishing open-access research and development partnerships with academic experts, who generally lack access to commercial scale hardware, to accelerate collaborative cross- and intersector advances in trusted AI for water systems.

Framework for benefit and risk assessment

The water industry must establish a transparent framework for responsible deployment of AI in relation to control of its infrastructure and services that provides balanced assessment of benefits and risks99. AI is not an end goal itself and should be treated as a technological response to clearly defined problems. ‘Don’t start with moon shots’: a holistic approach must include thorough understanding of water systems deficiencies, evaluation of typical AI systems that will safely address such deficiencies, and staged prototyping, pilot and roll-out processes100.

To this end, we propose an exemplar six-layer framework (Table 1), which elaborates general concepts from the trustworthy AI guidelines on technical robustness, governance and ethics, that addresses theoretical screening, proof of concept and practical scale-up considerations for the deployment of AI in water systems. The example considerations provided here are intended to inspire water sector practitioners with the basis of a ‘live scorecard’ to qualify the net value proposition of a given AI application pre-, mid- and post-implementation in the real world.

Table 1 Framework for holistic benefit and risk assessment to support water utilities in the responsible deployment of AI across water systems
Full size table

Closing reflections

The world is not on track to meet Sustainable Development Goal 6. Over 1.6 million people are dying annually from unsafe and inaccessible drinking water, stormwater and sewerage services, and climate change is expected to exacerbate water-related issues. In response, AI has been proposed as the latest technological innovation to help address water systems deficiencies. However, technology alone cannot solve water supply and wastewater disposal problems. Poorly managed proliferation of AI across water systems may give rise to progress traps that could further undermine and complicate water security.

As such, this Perspective sought to cross-pollinate domain-siloed knowledge, and contextualize synthesized insights, from the technical water systems and AI safety literature, to raise awareness among academics, water and AI industry practitioners and layman end-users of the need to prioritize ‘responsible deployment’ of AI in water systems to mitigate risks.

Given that AI has not yet proliferated, and its deployment is particularly nascent in the water sector, empirical data on real-world applications is relatively scarce. As such, the example water-system-wide AI applications outlined here, to highlight potential system-wide reach of AI, are based on demonstrations of AI algorithms or isolated case studies in the academic literature. Similarly, the examples of AI risks provided are speculative, albeit informed by cutting-edge AI safety literature.

Notwithstanding the above, one real-world example of a failed AI application causing harm to people was reported in November 2022101. In this instance, the public health department of Toronto, Canada, replaced its traditional method of using day-old laboratory tests with an ML-based predictive water quality assessment tool to determine whether the water quality at local beaches was safe for swimming. Rather than being more accurate as purported, the ML tool identified only about 30% of unsafe beach water days, resulting in 50 instances of public bathers being exposed to dangerous bacteria levels over the summer. This highlights the very real harm that AI failure may cause if responsible deployment principles are not prioritized and effectively executed.

As empirical data on successful and unsuccessful applications of AI across water systems becomes more readily available, we encourage researchers and practitioners to rigorously database and evaluate such information to build a thorough understanding of the real benefits and real risks, necessary to evolve risk management practices as AI proliferates. We hope that the framework conceptualized here provides a basis from which multidisciplinary academics and water and AI industry practitioners can develop proactive and critical risk management practices, informed by participatory approaches that engage and educate end-users.

Finally, the water industry must take a tiered approach to risk anticipation and mitigation to ensure responsible deployment of AI in water systems, including addressing barriers related to infrastructure and digital literacy, establishing institutional, software and hardware mechanisms for trustworthy AI, and prioritizing applications based on rigorous benefit and risk assessment. With US$6.3 billion projected investment in AI water technologies, we urge the water sector, particularly the larger and more developed utilities driving the foray into digital water, to allocate a substantial portion of this funding away from purely technical capacity building to AI safety initiatives that will help ensure potential benefits of AI in water systems are realized at scale.