Every Drop a Story

 

Life, Liberty, and the Pursuit of Responsible AI

Every decision to entrust AI with greater responsibility carries a human obligation: to understand what could go wrong, who could be affected, and who will answer for the consequences. The same obligation applies to the money, land, electricity, and water we commit to its development. AI’s promise is immense. Realizing that promise responsibly requires us to consider its benefits, risks, and costs together, while preserving our ability to question, intervene, and change course.

This is the starting point for a more useful conversation about AI.

Debates about its future often settle into two familiar positions. One anticipates increasingly powerful systems escaping human control and threatening our survival. The other dismisses such concerns as speculation and assumes society will adapt as it has to earlier technologies.

Neither position provides an adequate basis for making decisions today. Catastrophe is not inevitable, and successful adaptation is not guaranteed. Both outcomes depend, in part, on choices about how AI is developed, deployed, financed, and governed.

The central question is whether our capacity to exercise responsibility can grow alongside the power we create.

The responsibility we delegate

A serious AI failure would not require a machine to become conscious or develop hostile intentions. Harm could arise from a system pursuing a poorly defined objective, acting on incomplete information, or being used in circumstances for which it was never adequately tested.

Consider an AI system assisting with reservoir operations. Its recommendations might improve efficiency under familiar conditions. But what happens during an unusual combination of extreme rainfall, equipment failures, uncertain forecasts, and competing demands? Would it recognize the limits of its analysis? Would operators have enough information and time to challenge its recommendations? Would there be a workable alternative if the system became unavailable?

These questions become more consequential as systems receive greater authority and access to critical infrastructure. A flawed recommendation can sometimes be corrected before anyone acts. An automated decision affecting water releases, electricity delivery, or financial transactions may have consequences before anyone notices the error.

Human oversight must therefore mean something practical. It requires people with the expertise, information, time, and authority to intervene. Assigning someone to approve an output offers little protection if that person cannot evaluate it or is expected to accept it routinely.

Delegating a task to AI does not eliminate the responsibility of the organization using it.

There is also the risk of deliberate misuse. Systems that make useful work easier can make harmful activity easier as well. Fraud, manipulation, surveillance, and cyberattacks deserve attention alongside questions about the behavior of future autonomous systems.

These are different risks, requiring different responses. Treating them as one undifferentiated threat makes it harder to identify where safeguards will do the most good.

A risk management approach

My work in water resources planning offers a useful way to think about this challenge.

We do not wait for certainty that a catastrophic flood will occur before considering protective measures. Nor does the possibility of a flood establish that catastrophe is unavoidable. We examine hazards, exposure, vulnerability, consequences, and uncertainty. We consider who is at risk, what protection is feasible, and how much additional protection is justified.

We also ask what happens when our assumptions fail.

The same discipline should guide AI. What decisions will a system support or make? How serious could its errors be? Who would be exposed? Could a failure spread across connected systems? How quickly could people detect it, interrupt it, and recover?

The answers should determine the level of scrutiny. An assistant drafting a routine email does not warrant the same safeguards as a system influencing medical treatment, operating infrastructure, or supporting military decisions. Oversight should reflect both the potential harm and the difficulty of reversing it.

For consequential uses, responsible deployment calls for testing under adverse conditions, clear operating limits, independent review where warranted, and reliable alternatives. Monitoring must continue after deployment because real operating conditions can reveal weaknesses that testing missed.

Uncertainty also deserves an honest place in the decision. It can justify proceeding cautiously, restricting a system’s authority, or delaying a particular application while evidence improves. The appropriate response depends on what is at stake.

Risk management makes that judgment explicit. It gives us a way to pursue benefits while preparing for failure.

When incentives reward speed

Technical safeguards operate within economic and institutional pressures.

A company may take safety seriously while worrying that additional testing will allow competitors to secure customers, capital, and talent. Governments may face similar pressures when they view AI development as a source of national advantage.

This creates a collective-action problem. Participants can recognize the value of caution and still face strong incentives to move faster.

The distribution of benefits and costs compounds the difficulty. A developer may capture much of the financial return from a successful product, while workers, consumers, communities, or public agencies bear some of the consequences when it fails. Markets cannot be assumed to account fully for costs that fall on people outside a transaction.

Water economics offers a familiar parallel. Decisions that make sense for individual groundwater users can collectively deplete an aquifer. The problem requires rules and incentives that connect individual choices to shared consequences.

AI governance faces a related task. Clear responsibilities, credible evaluation, incident reporting, and proportionate accountability can help make responsible conduct part of the competitive environment.

Good governance should also allow useful applications to proceed. Requirements that are vague, unnecessarily burdensome, or insensitive to differences in risk can discourage beneficial innovation. The aim should be rules that are understandable, enforceable, and responsive to evidence.

The resources behind the intelligence

AI also has a physical and financial footprint.

Computing depends on data centers, semiconductor manufacturing, electricity, transmission networks, cooling, land, equipment, and capital. Water requirements vary with location, cooling technology, and the energy sources supplying a facility. These differences matter when assessing local impacts.

An investment can be attractive to its developer while creating costs elsewhere. Additional electricity demand may require infrastructure upgrades. Water demand may affect a constrained supply. Land and public funding committed to one project become unavailable for other uses.

That makes the AI buildout a matter of public planning as well as private investment.

What benefits will a proposed facility create? What resources will it require over its lifetime? Who will pay for supporting infrastructure? How will its demands interact with drought, heat, population growth, and existing commitments? What happens if anticipated computing demand or financial returns do not materialize?

These questions are especially important when infrastructure investments are costly to reverse. A community may remain responsible for maintenance, financing, or environmental consequences long after the assumptions behind a project have changed.

Financial success and social value must both be examined. A profitable project can impose uncompensated costs on others. A useful technology can also attract investments whose expected returns prove unrealistic.

Responsible planning therefore requires credible demand forecasts, consideration of alternatives, transparent allocation of costs, and attention to cumulative effects. A facility that appears manageable on its own may become part of a much larger regional demand on water, energy, and land.

The people affected should have a meaningful opportunity to understand and influence those choices.

What responsible AI could make possible

The reason to take these responsibilities seriously is also the reason to remain hopeful: AI could help us solve difficult and consequential problems.

It could support scientific discovery, help clinicians interpret complex information, expand educational assistance, and improve the analysis of public investments. In each setting, value depends on how well the system performs a useful task and how reliably people can assess its output.

In water resources planning, the opportunities are substantial.

Planners must consider climate uncertainty, hydrology, groundwater conditions, infrastructure limitations, agricultural demands, ecosystems, economics, and community needs together. AI could help organize these sources of information, identify patterns, and explore a wider range of alternatives.

It could assist with drought detection, flood forecasting, infrastructure maintenance, and the evaluation of water investments. It could help analysts examine combinations of conservation, recycling, managed aquifer recharge, reservoir operations, and ecosystem restoration under different futures.

Used carefully, it could also strengthen distributional analysis. A project’s overall benefits tell us little about whether vulnerable households can afford the resulting rates, whether a rural community receives reliable service, or whether environmental costs fall disproportionately on particular groups.

AI may help make those differences more visible. The judgments that follow still require public reasoning.

An analytical model can estimate tradeoffs. It cannot confer democratic legitimacy on a decision about whose needs take priority. Those choices require accountable institutions, professional judgment, and participation by the people affected.

Making oversight meaningful

Responsible AI should be visible in how an organization makes decisions.

Before adopting a consequential system, an organization should be able to explain the problem it intends to solve, the evidence supporting the proposed use, the alternatives considered, and the consequences of failure. It should identify who has authority to approve, monitor, restrict, and suspend that use.

People affected by consequential AI-assisted decisions should have appropriate ways to obtain explanations, challenge errors, and seek correction. Transparency is useful when it enables understanding and action.

Organizations must also preserve the capacity to operate when AI fails. That may require backup procedures, retained expertise, and regular practice making decisions without the system. Efficiency gains become fragile when the ability to question or replace a tool disappears.

Safeguards must evolve as systems, applications, and operating conditions change. Approval at one point in time cannot substitute for continuing responsibility.

The choices that remain ours

The future of AI will be shaped by decisions about authority, incentives, investment, and public purpose.

We decide which tasks to delegate and which judgments require direct human responsibility. We decide how much uncertainty is acceptable in a particular use. We decide what resources to commit, whose interests to consider, and what evidence is sufficient to proceed.

These choices deserve the same ambition we bring to technological development.

Responsible innovation means building the ability to learn from errors, correct course, and protect people as capabilities expand. It means treating public trust as something earned through conduct. It means recognizing that benefits are more durable when the institutions supporting them are competent, accountable, and open to challenge.

For me, this is what life, liberty, and the pursuit of responsible AI bring together: protecting human well-being, preserving people’s ability to understand and contest decisions affecting them, and directing innovation toward broadly shared benefits.

The power we create carries obligations. Meeting them is how we turn technological possibility into lasting public value.