The Real Cost of Progress Is Not What We Count

What if the most misleading question about a new technology is, “How much does it consume?”

That question sounds responsible. It asks us to measure water, energy, emissions, money, or lives, then compare the cost with the benefit. But measurement alone cannot tell us whether a risk is tolerable. A golf course may consume more water than a data center. A hamburger may require hundreds of gallons of water. A laboratory accident may kill one person, while a research program could theoretically prevent millions of deaths.

Yet these comparisons do not settle the moral question. They may not even address the right one.

The deeper issue is not simply the size of a technology’s footprint. It is what kind of future the technology makes possible, who gets to authorize that future, and whether the damage can be reversed once the experiment has escaped its boundaries.

This is why arguments about artificial intelligence infrastructure and dangerous biological research unexpectedly belong in the same conversation. Both reveal a recurring failure in modern society: we treat costs as if they were interchangeable, even when one cost is a gradual draw on a shared resource and another is a low probability of irreversible catastrophe. We compare quantities while ignoring qualities.

A better approach begins with a different unit of analysis: not consumption, but permission.

The Fallacy of the Single Score

Public debates often search for one decisive number. How many gallons of water does one AI query use? How many emissions does a data center produce? How likely is a laboratory accident? How many people might benefit from a scientific discovery?

The appeal is obvious. A single number appears to transform a complicated judgment into an accounting exercise. If one activity uses less water than another, perhaps criticism is hypocritical. If a research program might someday prevent a pandemic, perhaps present risks are justified. If a system increases productivity, perhaps its environmental cost is acceptable.

But numbers become deceptive when they flatten fundamentally different kinds of harm.

A cotton shirt that requires more than 700 gallons of water is not automatically equivalent to a data center that consumes millions of gallons per day. The comparison may be useful for scale, but it leaves out location, timing, scarcity, and control. Water drawn from a wet region is not the same as water drawn from an aquifer serving a drought stressed community. A one time cost is not the same as a permanent demand. A private recreational use is not the same as infrastructure designed to become a basic layer of the economy.

The same problem appears in biological research. A laboratory accident that exposes a worker to a deadly pathogen is not morally equivalent to a statistical risk spread across millions of people, even if the numerical probabilities could be made to look similar. A mistake involving an organism that disappears quickly is not the same as releasing one capable of sustained human transmission. The relevant facts include scope, speed, reversibility, and the ability to contain failure.

The central mistake is treating risk as a single vertical scale, from small to large. In reality, risk is multidimensional.

Consider five dimensions:

• Magnitude: How much harm could occur?
• Probability: How likely is the harm?
• Reversibility: Can the damage be repaired, or does it permanently alter the future?
• Distribution: Who receives the benefits, and who bears the costs?
• Governance: How strong are the institutions responsible for preventing and responding to failure?

Two activities can have the same expected cost while being radically different in every dimension that matters. Expected value is useful for some decisions, but it is not a substitute for judgment. A one in a million chance of a civilization altering event is not simply one million times less important than a certain inconvenience.

The question is not only how much harm a system causes on average. It is what happens if the system fails in the one way we cannot undo.

From Footprints to Thresholds

Environmental criticism of AI often gets trapped between two bad positions. One side treats every water statistic as proof that the technology is uniquely destructive. The other responds by pointing to industries that consume more water, as if greater waste elsewhere settles the matter.

Both positions miss the significance of thresholds.

An individual AI interaction may have a small footprint. But individual actions aggregate into infrastructure. A request is not merely a request when millions of people make it every day. It becomes a reason to construct data centers, secure electricity contracts, draw water, expand transmission networks, and reorganize the labor market around a new computational layer.

This changes the question from “What does my query consume?” to “What kind of system does widespread use authorize?”

Golf courses offer a useful contrast. Their water use may be enormous, but golf is not generally presented as an unavoidable foundation for every sector of the economy. It is also geographically bounded and, at least in principle, easier to reduce without redesigning education, medicine, government, and work. AI infrastructure is being built with a different ambition. Its advocates often describe it as a general purpose technology, one that will become embedded in nearly every institution.

That ambition creates a higher burden of proof. The more a technology claims to be inevitable, the more carefully society must examine its externalities. Inevitability is not a scientific finding. It is a political request for permission.

The same principle applies to laboratories working with ancient or highly dangerous viruses. The argument for such research may be intellectually compelling. Studying pathogens preserved in permafrost could reveal something about viral evolution or help prepare medicine for future threats. But the benefits are often uncertain and distant, while the risks can be immediate and difficult to reverse.

This is not an argument against curiosity. It is an argument against confusing the existence of a possible benefit with a sufficient justification for taking any risk in pursuit of it.

A threshold model helps clarify the distinction. Before adopting a technology or permitting an experiment, ask whether it crosses one of four thresholds:

1. Resource threshold: Does it consume a scarce resource at a scale that changes local conditions?
2. System threshold: Does it become so embedded that failure would disrupt essential institutions?
3. Containment threshold: Could an accident escape the boundaries where it can be managed?
4. Legitimacy threshold: Have the people bearing the consequences meaningfully consented to the risk?

A technology does not need to be the largest consumer or the most dangerous activity in existence to cross one of these thresholds. It only needs to change the structure of dependence around it.

Why Moral Reactions Become Data

There is another layer to these disputes. Public outrage is often dismissed as irrational when it focuses on a particular technology while ignoring larger sources of harm. If people condemn AI water use but continue eating meat, wearing cotton, or playing golf, critics say the concern is inconsistent.

Sometimes it is. But inconsistency does not make the concern meaningless.

People are not reacting only to gallons or emissions. They are reacting to what a technology symbolizes. AI is associated with a culture that may alter employment, education, creativity, truth, and political power. Its environmental footprint becomes a concrete handle for a more diffuse anxiety: the fear that institutions are rapidly adopting systems before anyone has decided whether they should be adopted.

Likewise, concern about a dangerous virus laboratory is not merely fear of one accident. It reflects distrust in the assumption that technical expertise automatically guarantees control. Reports of past accidents, inadequate safety conditions, or weak oversight make the risk feel less like an abstract probability and more like evidence that the system may not deserve confidence.

The emotional response is therefore often a form of governance feedback. It tells us that people are not evaluating a technology in isolation. They are evaluating the history of promises, secrecy, incentives, and institutional behavior surrounding it.

This does not mean every public fear is correct. Fear can exaggerate vivid events and ignore invisible harms. But dismissing fear because its arithmetic is imperfect is also a mistake. The public may be using a particular statistic to express a broader question: Who is deciding, and what happens if they are wrong?

A society that answers only with better public relations will deepen the distrust. A society that answers with transparent data, independent oversight, and genuine limits may discover that opposition becomes more specific and more manageable.

The Asymmetry of Irreversible Mistakes

The strongest connection between environmental infrastructure and dangerous biological research is the problem of asymmetric failure.

Many modern systems reward visible benefits and discount remote harms. The benefits arrive on a quarterly report, in a product launch, or in a promising research result. The harms are dispersed across communities, delayed into the future, or assigned a tiny probability. This creates an institutional bias toward action.

But some errors are not ordinary expenses. They are option destroying events. Once a pathogen spreads beyond containment, the world cannot simply return to its previous state. Once a region’s groundwater is degraded, recovery may take decades. Once society becomes dependent on a system that concentrates power and infrastructure in a few firms, reversing course becomes politically and economically difficult.

This suggests a principle of asymmetric caution:

The more irreversible the downside, the less acceptable it is to justify action through average benefits alone.

Asymmetric caution does not require freezing innovation. It requires changing what counts as evidence. For ordinary consumer goods, it may be reasonable to compare costs and benefits using market prices. For technologies with systemic or irreversible consequences, society should demand additional conditions:

• The benefit must be concrete rather than merely possible.
• The risk must be independently measured rather than reported only by the operator.
• The activity must have a credible containment and shutdown plan.
• The people exposed to the downside must have representation.
• The institution must be capable of surviving scrutiny, including disclosure of failure.

These conditions apply to a data center as much as to a research facility, though the hazards differ. A data center may not release a virus, but it can lock communities into water demand, energy demand, and economic dependence. A laboratory may not consume vast quantities of water, but it can create a biological risk whose consequences cannot be recalled.

The common issue is not “technology is dangerous.” The common issue is whether the institutions deploying technology can be trusted with consequences that exceed their own control.

A Practical Framework for Responsible Permission

The most useful response is neither blanket enthusiasm nor blanket rejection. It is a more demanding form of permission.

Before supporting a new project, ask four questions.

1. What is the benefit, precisely?

Replace broad claims with measurable outcomes. “This will transform society” is not a benefit. “This will reduce diagnostic errors by a documented amount in settings where clinicians are scarce” is closer. “This may teach us something about ancient viruses” should be accompanied by a clear account of what knowledge would change and why it cannot be obtained through safer methods.

2. Who pays the physical cost?

Identify the resource and the community. Do not stop at total water use. Ask whether the water comes from a stressed basin, whether local residents have alternatives, and whether the operator pays the true social cost. Do not stop at a laboratory’s scientific value. Ask who bears the exposure risk, whether workers can refuse unsafe tasks, and whether nearby populations are informed.

3. What is the failure mode?

Imagine not the average day, but the bad day. For a data center, this might involve water shortages, grid stress, or a concentration of critical services in an unaccountable provider. For a laboratory, it might involve a procedural mistake, equipment failure, a natural disaster, or a pathogen whose behavior differs from expectations.

Then ask the most important follow up: can the system be paused without causing a larger crisis?

4. What would change our minds?

A responsible institution defines stopping rules before it begins. These might include contamination events, breaches of safety protocols, local resource thresholds, or evidence that the claimed benefit is not materializing. If no evidence could ever justify slowing or stopping the project, then the project is not being governed. It is being defended.

This framework also improves individual choices. People do not need to calculate the exact water footprint of every AI query or abandon every product with an environmental cost. They can instead ask whether their use reinforces a system whose benefits are clear, whose costs are disclosed, and whose scale is accountable.

Key Takeaways

• Do not compare technologies using a single number. Evaluate magnitude, probability, reversibility, distribution, and governance together.
• Distinguish individual consumption from system authorization. A small personal action can help build a large infrastructure when repeated millions of times.
• Treat public unease as information. It may express distrust of incentives and oversight, not merely confusion about statistics.
• Demand stronger evidence for irreversible risks. Distant and speculative benefits should not automatically justify immediate, difficult to contain dangers.
• Ask for stopping rules. Any institution seeking permission should explain what evidence would lead it to pause, redesign, or end the project.

The mature question is not whether AI is worse than golf, whether a laboratory is more valuable than a potential accident, or whether one lifestyle choice makes another criticism hypocritical. Those comparisons can provide context, but context is not a verdict.

The real question is whether a society is capable of granting permission deliberately. Can it distinguish a manageable cost from a systemic dependency? Can it recognize that some harms are not interchangeable? Can it require powerful institutions to disclose uncertainty before the public is forced to absorb the consequences?

Progress is often described as the expansion of what humans can do. That is only half the story. The harder achievement is expanding what humans can do without surrendering the ability to say no.

A gallon of water and a new virus cannot be placed on the same scale. Neither can a useful tool and an irreversible mistake. The task of a responsible society is not to pretend they are comparable. It is to build institutions wise enough to know when comparison itself has become a way of avoiding judgment.