Why the AI safety field is in demand for better questions

June 28, 2026

I spoke currently with an economist who specialises in “artificial intelligence” (AI). Thoughtful, well-read, deeply engaged with the question of how “artificial intelligence” will reshape economic activity. At some point the conversation turned to “economic growth”. While I mentioned physical constraints, his answer came “Like physical limits? But they at least allow for multifold increases in GDP, no?”. Well, no! Or only if you look at the concept of “Gross Domestic Product” (GDP) how it is defined, a function of capital and workforce – you accumulate one, you deploy the other, and the “economy” measured based on the concept of “GDP” expands.

I asked where energy came from in that model? A puzzled pause.
Where minerals came from? Another pause.

The conversation moved on.

This is not a criticism of one economist, extremely thorough on multiple fronts. It is a description of an entire intellectual tradition. The dominant framework for thinking about “economic growth” — and by extension, about the value and risk of AI — treats the economy as a financial system that happens to use some physical inputs. Thermodynamics does not appear in the model, nor does the depletion of our bodies, nor the absorption capacity of forests and oceans. These are “externalities“, in both the technical and the colloquial sense: things that happen outside the system being studied.

The AI safety field has inherited this blind spot wholesale. It asks serious questions about model behaviour, about misuse, about long-term alignment between machine intelligence and human values. What it does not ask is what world this technology is being built for, and whether that world has the physical capacity to support it.

The economy is a physical system

Nicholas Georgescu-Roegen, a Romanian mathematician and economist working in the 1970s, made an observation so obvious in retrospect that its continued absence from mainstream economics is itself a kind of data point:

“Every economic process transforms low-entropy resources into high-entropy waste.”
Nicholas Georgescu-Roegen

You cannot run it backwards. A barrel of oil burned is gone. A tonne of copper dispersed across ten million circuit boards is, for all practical purposes, gone. The laws of thermodynamics do not negotiate with financial instruments.

The standard production function — GDP as a function of “capital” and “labour” — omits the variable that does the actual physical work. Robert Ayres and Benjamin Warr, working across INSEAD and Chalmers University, spent years tracking the relationship between useful work — energy successfully converted into economic output — and GDP growth across the twentieth century. The correlation came out at close to 0.9. Not price, not capital investment. The availability of useful energy.

Global material extraction tells the same story from a different direction. The “International Resource Panel” from the “UN Environment Programme” (UNEP) puts annual extraction of materials at 104 billion tonnes in 2023, up from 22 billion in 1970. The economy grew because we found and consumed vastly more physical stuff. Capital and labour organised that consumption. They did not replace it.

There is a financial dimension to this that sharpens the point. Global debt stood at $348 trillion at the end of 2025, roughly three times global GDP of around $115 trillion, according to the “Institute of International Finance” (IIF).

More troubling than the absolute figure is the direction of travel: each dollar of new debt generates progressively less GDP than the last, a decline in what economists call the “marginal productivity of debt” that has been visible in “US Federal Reserve” (FED) data since the 1980s. In the early post-war decades, a dollar borrowed produced more than 70 cents of GDP growth. By the time of the 2008 financial crisis, that figure had fallen below 10 cents. The money being created is not disappearing. Rather, it is flowing into financial assets and real estate, inflating valuations rather than building productive capacity. The financial system, in other words, is pricing the future as if the physical constraints described above do not exist. That gap between financial abstraction and physical reality is not a footnote to the argument. It is the argument.

The data sits in plain sight, published by credible institutions, cited in specialist literature. It does not appear in the models that guide investment decisions, technology deployment, or safety research. We have built very sophisticated maps of a terrain we have not looked at.

Three things happening at once

(1) Start with energy: Conventional “crude oil” production peaked somewhere between 2005 and 2008, according to successive editions of the IEA’s World Energy Outlook. The 2008 edition was candid about it: non-OPEC conventional production was already at plateau, and most non-OPEC countries had already passed their peak. What kept total supply figures rising afterward was not new conventional discovery but the addition of unconventional sources — fracked tight oil, oil sands, deepwater — each more expensive to extract, each with steeper decline rates than the fields they supplemented. Eighty percent of global primary energy still comes from “fossil fuels” (e.g., coal, gas, oil). The era of cheap, abundant, high-quality energy has been ending for two decades. We have been covering the gap with increasingly costly substitutes and calling it “stability”.

(2) Meanwhile, the ore grades being mined are falling. In Chile’s copper mines — which hold between 40 and 55 percent of known global reserves — the average grade dropped 28.8 percent in a single decade. At today’s average of around 0.5 percent copper content, roughly 99.5 percent of everything blasted out of the ground is waste rock. Energy consumption in those Chilean operations rose 46 percent between 2003 and 2013; copper output rose 30 percent. More effort, less return. S&P Global puts gold head grades down 13.4 percent since 2012, copper down 7.5 percent. These are not cyclical dips. The direction has been consistent for decades, across multiple metals, across multiple continents.

(3) The third pressure is harder to see because it operated quietly for so long. Forests, soils and oceans have been absorbing roughly half of human “Carbon Dioxide” (CO₂) emissions for decades — an enormous subsidy, provided free of charge by systems we did not build and cannot replace. That capacity is eroding. In 2023, the land carbon sink collapsed to its lowest level since systematic measurement began in 1958, absorbing between 1.5 and 2.6 billion tonnes of CO₂ against a recent average of 9.5 billion. Atmospheric CO₂ grew at 86 percent above the previous year’s rate, despite fossil fuel emissions rising by less than one percent. The “10 New Insights in Climate Science 2025“, produced by more than 70 scientists across 21 countries, concluded that land and ocean systems are approaching the limits of what they can absorb.

What connects these three pressures is that they are not independent. Lower ore grades require more energy per tonne of useful metal. Higher energy use accelerates the degradation of carbon sinks. A destabilised climate damages the water systems and agricultural soils that mining and food production depend on. Each constraint tightens the others. The problems do not queue up politely. They arrive together.

Into this world, we are building AI infrastructure at speed

The “International Energy Agency” (IEA) 2025 report on energy and AI contains a number worth sitting with. Data centre electricity consumption is growing at 15 percent per year — more than four times faster than all other sectors combined. AI is the primary driver. Electricity consumed by AI-focused infrastructure alone surged 50 percent in 2025. By 2030, US data centres are projected to consume more electricity than the production of aluminium, steel, cement, and chemicals put together.

That demand will not be met cleanly. Natural gas and coal together are expected to cover over 40 percent of the additional load from data centres through the end of the decade. The AI buildout is, in material terms, partly a fossil fuel buildout.

The hardware side compounds this. AI infrastructure requires copper, lithium, cobalt and rare earth elements — the same materials already under pressure from grade decline, geographic concentration, and the water and energy costs that extraction increasingly demands. The energy transition and the AI buildout are competing for the same depleting resource base, on the same timeline.

The efficiency argument is worth taking seriously on its own terms. Computational efficiency has improved by roughly an order of magnitude annually in recent years, and the electricity consumed per AI query has fallen sharply. A “large language model” (LLM) query today uses less power than a few seconds of television. But efficiency at the task level has not slowed total consumption at the system level. The number of tasks, the size of models, and the scale of deployment are all growing faster than efficiency gains can offset. The “rebound effect” — whereby lower cost per unit enables more units — is a well-documented pattern in “energy economics“. It is operating here at scale.

The result is a technology sector whose physical footprint is expanding rapidly into a resource base that was already under stress before the current buildout began.

What “safety” currently means

The AI safety field has produced work that matters. Alignment research — how to build systems that reliably pursue intended goals rather than proxy measures of them — is a hard problem and people are working on it seriously. Interpretability research is trying to make the reasoning inside large models something other than a black box. These are the problems good to be questioned.

But the field’s definition of safety is almost entirely about model behaviour: will the system say harmful things, pursue misaligned goals, be turned to destructive use, or eventually develop objectives that conflict with human welfare. These are risks located inside the technology and its immediate social context. They take the pace and scale of deployment as a given and ask how to make the thing less dangerous as it spreads.

Whether the deployment trajectory itself is compatible with the physical world it is accelerating into — that question does not appear in the research agenda. It does not appear because the research agenda is shaped by what funders will fund, and the funders are, in the main, the institutions driving the deployment. It is an architecture, and it produces exactly the result you would predict: a field that asks how to make the engine safer without asking where it is going or what it is burning to get there.

The questions that would require slowing down — whether this much AI, deployed this fast, consuming these resources, serves any purpose proportionate to its cost — sit outside the funding perimeter. Not because they are unreasonable. Because they are inconvenient to ask.

The questions that are not being asked

Any serious engineering discipline asks, before deployment at scale, what the thing is for and what systems it depends on. You do not build a dam without a hydrological survey. You do not design a power grid without modelling demand across the network it will join. AI is being deployed across the global economy with no equivalent accounting of the physical systems it consumes and transforms. That is a clear observation about the methodology involved.

Some of the relevant work does exist, at the margins. Research on AI energy efficiency, on the mineral footprint of hardware, on use cases where AI substitutes for more resource-intensive activity — this is real and occasionally valuable. But this is not what the field calls “safety”. It lives in a separate silo, funded differently, weighted differently, never brought into the central conversation about risk.

If it were, safety research would have to engage with questions it currently ignores:

At what deployment pace can the energy system absorb new AI demand without locking in fossil dependence for another decade?
Which use cases generate value commensurate with their physical cost, and which simply add load to a grid already under strain?
How do you coordinate AI development with the governance of the physical systems it is reshaping — energy grids, mineral supply chains, water tables, land use?

These are harder than alignment. They do not yield to technical solutions alone. They require the kind of institutional architecture that Western systems have consistently failed to build.

Safe for what world?

An AI that does not say harmful things, does not pursue misaligned goals, and resists weaponisation is not safe in any meaningful sense if its deployment is accelerating the depletion of the systems that economies and societies depend on. Safety defined only at the boundary of the model is not safety. It is a perimeter drawn to avoid the harder question.

The world AI is being built for is not stable. Its energy supply is transitioning under pressure, with that transition constrained by the mineral and water limits of the places where transition materials are found. The carbon absorbers that have cushioned decades of emissions are weakening faster than the models forecast. Soils are eroding faster than they form. The most critical mineral deposits are getting harder and more expensive to work with every passing year. None of this is speculative. It is in the data, published by the “IEA”, the “UN Environment Programme”, the “Food and Agriculture Organization” (FAO), and the scientific bodies that track atmospheric carbon.

None of it makes AI development wrong. It makes the absence of these questions from the safety agenda a choice, not an oversight. A technology deployed without reference to the physical world it runs on and draws from is not a solution to the problems of that world. It is one more claim on systems that are already overdrawn.

The safety field is serious. It has simply been asked to guard a smaller room than the one we are actually in. Until the physical world appears inside the frame — as constraint, as context, as the system within which AI either makes things better or accelerates their unravelling — the research will keep producing answers to a question that is not quite the healthy one.

Safe for what world, exactly? That is not a rhetorical question. It is the one that has not been asked.

Sources

Energy-GDP correlation:
- Ayres & Warr (2003, 2009), INSEAD/Chalmers University
- IEA World Energy Balances Database (1960–2023)

Global material extraction:
- UNEP International Resource Panel, Global Material Flows Database (1970–2023)

Global debt and debt-to-GDP ratio:
- Institute of International Finance, Global Debt Monitor (February 2026)

Marginal productivity of debt:
- Federal Reserve Bank of St. Louis, FRED Database; analysis in Weiner, K., Monetary Metals (2017–2024).

Conventional oil peak:
- IEA World Energy Outlook 2008
- Aleklett et al. (2010), Energy Policy, vol. 38.

Ore grade decline:
- S&P Global Market Intelligence (2024)
- Calvo, Mudd et al., Engineering Conferences International (2016)
- Escondida case study, Earth Resource Investments (2025)

Carbon sink weakening:
- Ke, Ciais et al. (2024), National Science Review, vol. 11
- 10 New Insights in Climate Science 2025, Future Earth / The Earth League / WCRP