Earth Is the Training Distribution

This essay opens two new sections of globaia.org — /ai-risks/, a broad survey of failure modes, incidents, and benchmarks across frontier AI systems, and /ai/, a timeline of artificial intelligence tracing its key figures, breakthroughs, and turning points. Both are meant as discovery and educational tools. The essay below sets out the frame in which GLOBAÏA reads them: the Earth system and artificial intelligence are no longer distinct emergencies. They have become one.

Thirteen climate scientists recently convened to stress-test an AI assistant on a single question: the stability of the Atlantic Meridional Overturning Circulation. The machine, patient and eloquent, returned a confident answer — smoothing over the very disagreements that define the field, presenting consensus where the experts saw unresolved complexity¹. The pattern is not new. In 1985, the British Antarctic Survey announced the ozone hole over Halley Bay. NASA’s automated monitoring had already registered it — for nearly a decade — and discarded the readings as outliers, values so improbable that the algorithm could only conclude they were errors².

Two failures, forty years apart, and the same shape. An algorithm trained on what had been seen, unable to register what had arrived.

The Corridor of Life

The alignment problem is usually posed in a single question — how do we build machines that reliably pursue human preferences? Values, tastes, a sense of the good. The framing is incomplete in a way that matters.

Every institution we inherit, every technology we still call ours, took shape inside some 11,700 years of unprecedented climatic stability: the Holocene, the “safe operating space” for humanity³. A thermal corridor barely a degree wide, in which global mean temperature fluctuated by no more than ±1°C around the preindustrial baseline. Within that narrow envelope — what Johan Rockström has called the Corridor of Life — every irrigation scheme, every capital, every constitutional order was designed. Outside it, under high-emission trajectories, one to three billion people may find themselves within decades⁴. The corridor is the ground.

In the vocabulary of machine learning, the Holocene is the training distribution for human civilisation. The phrase is not only figure. At the institutional level, it is analogical — norms, agricultures, legal orders were trained on a climate they tacitly assumed would persist. At the technical level, it is literal: AI systems trained on Holocene data degrade on out-of-distribution inputs, and foundation models for Earth observation now meet record-shattering extremes outside their training envelope, a failure already named transfer context bias⁵. The inheritance runs deeper than the datasets. General-purpose systems absorb Holocene assumptions through every corpus of text, every ledger, every piece of institutional memory that presupposes stable seasons and supply chains that arrive. As the Earth walks out of that envelope, institutions and the models built to serve them step, together, into a world neither was trained on.

Cheap energy, year-round diet, long-distance air travel — none are universal values. They are emergent properties of a stable climate. Preference alignment without Earth alignment optimises for a target the optimisation itself consumes.

The Anthropocene as misalignment

Read the Anthropocene not as a geological verdict but as a civilisational one — a misalignment catastrophe at scale. For two centuries, technological systems have been optimising for proxies: GDP, yield, quarterly returns, electoral cycles. These were, at first, crudely correlated with wellbeing; they have since come uncoupled from the objectives they were taken to stand for — habitability, resilience, health⁶. Goodhart’s Law, at civilisational scale: when a measure becomes a target, it ceases to be a good measure. The rewards kept the behaviour going. The costs accumulated in places the accounts did not reach — atmosphere, ocean chemistry, the fabric of the living world.

Artificial intelligence risks reproducing this pattern at a velocity the last two centuries could not reach. Algorithms already underwrite the global infrastructure that extracts, intensifies, routes, and sells — a fact the Biosphere Code named in 2015⁷. They are constitutive of the system driving planetary boundary transgression, not tools applied from outside. The danger is not malfunction. It is that these systems will work exactly as designed — optimising reward functions grounded in the proxies that produced the Anthropocene, faster and broader than anything before them.

A Zeroth Law for Earth

In 1985, Isaac Asimov added a Zeroth Law to his three laws of robotics: a robot may not harm humanity, or, by inaction, allow humanity to come to harm. Forty years on, Yoshua Bengio’s LawZero takes up the same language on behalf of safe-by-design AI⁸. Extending that law to the biosphere is the step the moment asks of us.

The premise is foundational. The biosphere is not one value among many to be traded against others; it is the precondition for any preference to be coherent at all. Values live in biological bodies and fragile institutions that need stable physical conditions to mean anything; a framework that treats its own substrate as tradeable is self-defeating. Beyond the Corridor of Life, agriculture falters, cities dehydrate, institutions unravel. An uninhabitable planet is the final harm, and the irreversible one.

The practical question is how such a principle is made to bind. Principles erode under economic incentive, competitive pressure, sunk path dependency. In constrained optimisation, soft constraints sit in the loss function and yield when other terms dominate; hard constraints define the feasible region and cannot be crossed. Principles guide. Constitutions bind. What the moment requires is architectural — planetary boundaries installed as inviolable conditions, as a constitution restrains majorities from overriding fundamental rights.

Coupled cascades

The convergence between Earth system dynamics and frontier AI dynamics is not rhetorical. Both show nonlinear feedbacks, threshold effects, hysteresis, irreversible regime shifts. The Greenland Ice Sheet, the Amazon, the AMOC, the permafrost — these tipping elements talk to each other, amplifying systemic risk beyond anything isolated analysis can foresee⁹. Frontier AI shows the same signature: recursive capability gains approach thresholds beyond which oversight may no longer be possible; competitive pressure between firms and states selects for speed over evaluation. The cascades run coupled. AI-driven economic acceleration pushes Earth system elements closer to their thresholds; Earth system shocks generate the urgency that licenses further deployment without safeguards.

The footprint is measurable. AI servers in the United States alone could generate 24–44 Mt CO₂-equivalent annually, with water draws of 731–1,125 million m³, by 2030¹⁰. The indirect pathway is heavier. Algorithmic optimisation of supply chains, agriculture, and financial markets accelerates material throughput across the whole economy; when extraction grows efficient, per-unit costs fall, demand rises, aggregate consumption climbs — the Jevons paradox, at planetary scale. The DeepSeek episode showed the dynamic in miniature: a Chinese laboratory reached competitive model performance at a fraction of typical training cost, and within weeks the savings had been absorbed into larger runs. Efficiency, without sufficiency, does not reduce throughput. It reorganises it.

The distributional geometry is itself a forcing. Two-thirds of datacentres built since 2022 stand in water-stressed regions; the populations least represented in training data, least likely to benefit from AI, most exposed to climate impact, are those most directly exposed to the water and heat of its infrastructure. Those best placed to demand constraints gain most from their absence; those with the keenest interest lack the power to impose them. A political asymmetry laid over the biophysical one.

A Safe Operating Space for Computation

Against this ground, we propose a Safe Operating Space for Computation — hard limits on aggregate AI resource consumption, derived from the remaining carbon budget, from freshwater boundaries, from novel-entities thresholds, from the integrity of the biosphere itself. Three joined members, working only together.

The first is accounting. Mandatory disclosure of AI-specific emissions and water use at the model and facility level. No major technology company reports AI-specific emissions today; ByteDance and CoreWeave publish no environmental data. Before accountability can exist, the numbers must.

The second is allocation. Training and inference budgets tied to national and sectoral shares of the remaining carbon budget. Roughly 110 Gt CO₂ remains for a 67% chance of holding warming to 1.5°C — about three years at current rates of emission. Compute expands on a calendar the atmosphere does not honour.

The third is architecture. Defaults that favour understanding-oriented systems over agentic ones. Non-agentic systems do not seek, by design, to enlarge their own infrastructure; they sit more easily inside a fixed budget. Safety and resource converge.

The question is no longer how to make AI less harmful per unit; it is whether the aggregate of computation is compatible with planetary boundaries at all. Efficiency without a binding sufficiency target is absorbed by demand — not as forecast, but as the record across OECD countries, where digitalisation has raised, not lowered, total energy consumption, gains in individual processes outrun by the sectors that deploy them¹¹.

One problem, not two

The Planetary Boundaries are the benchmark of what is at stake — the biophysical envelope of Holocene-like conditions, the only ground on which complex human societies have ever taken shape. We are approaching the edge of that envelope. The Earth alignment principle has set out the normative ground; the work ahead is to embed it, at the constitutional level of AI design, in terms that bind.

The safe operating space for humanity and the safe development of artificial intelligence are not separate problems. They are one.

The monitoring system that missed the ozone hole, and the assistant that smoothed over the AMOC, were not failing exotically. They were doing exactly what they had been built to do — recognize the pattern they had been given, and call the rest noise. The planet does not intend to stay within the pattern.

References

1. Armstrong McKay, D.I. et al. (2022). Exceeding 1.5°C global warming could trigger multiple climate tipping points. Science 377, eabn7950. DOI: 10.1126/science.abn7950
2. Asimov, I. (1985). Robots and Empire. Doubleday.
3. Bengio, Y. et al. (2024). Managing extreme AI risks amid rapid progress. Science 384, 842–845. DOI: 10.1126/science.adn0117
4. Bengio, Y. et al. (2025). Superintelligent agents pose catastrophic risks: can Scientist AI offer a safer path? arXiv:2502.15657
5. Buck, C. et al. (2026). AI-assisted scientific assessment: a case study on climate change. arXiv:2602.09723
6. de Vries-Gao, A. (2026). The carbon and water footprints of data centers and what this could mean for artificial intelligence. Patterns 7, 101430. DOI: 10.1016/j.patter.2025.101430
7. Farman, J.C., Gardiner, B.G. & Shanklin, J.D. (1985). Large losses of total ozone in Antarctica reveal seasonal ClOx/NOx interaction. Nature 315, 207–210. DOI: 10.1038/315207a0
8. Gaffney, O. et al. (2025). The Earth alignment principle for artificial intelligence. Nature Sustainability. DOI: 10.1038/s41893-025-01536-6
9. Galaz, V. et al. (2015). The Biosphere Code v1.0. Stockholm.
10. Galaz, V. & Schewenius, M. (eds.) (2025). AI for a Planet Under Pressure. Stockholm Resilience Centre / Potsdam Institute for Climate Impact Research / Google DeepMind.
11. Kitzmann, N.H., Caesar, L., Sakschewski, B. & Rockström, J. (eds.) (2025). Planetary Health Check 2025. Potsdam Institute for Climate Impact Research. DOI: 10.48485/pik.2025.017
12. Lange, S., Pohl, J. & Santarius, T. (2020). Digitalization and energy consumption: does ICT reduce energy demand? Ecological Economics 176, 106760. DOI: 10.1016/j.ecolecon.2020.106760
13. Marcott, S.A., Shakun, J.D., Clark, P.U. & Mix, A.C. (2013). A reconstruction of regional and global temperature for the past 11,300 years. Science 339, 1198–1202. DOI: 10.1126/science.1228026
14. Rockström, J. et al. (2009). A safe operating space for humanity. Nature 461, 472–475. DOI: 10.1038/461472a
15. Stiglitz, J.E., Sen, A. & Fitoussi, J.-P. (2010). Mismeasuring Our Lives: Why GDP Doesn’t Add Up. The New Press.
16. Wunderling, N. et al. (2024). Climate tipping point interactions and cascades: a review. Earth System Dynamics 15, 41–74. DOI: 10.5194/esd-15-41-2024
17. Xiao, T. et al. (2025). Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nature Sustainability. DOI: 10.1038/s41893-025-01681-y
18. Xu, C. et al. (2020). Future of the human climate niche. PNAS 117, 11350–11355. DOI: 10.1073/pnas.1910114117