Why Build Environmental Superintelligence?

The Question and the Answer in Brief

Why should humanity build Environmental Superintelligence (ESI), a planetary-scale AI system that protects nature in real time? This analysis derives its answer from first principles and current evidence rather than from any prior position. Candidate reasons were generated broadly, tested against the measured state of the world as of August 2026, subjected to independent fact-verification and adversarial red-team review, and ranked by explicit criteria. Only what survived is listed.

The answer begins with a piece of good news that is easy to miss. Information is the enabling resource for environmental protection, and it has never been cheaper or more abundant. Low-cost sensors, continuous Earth observation, open data, and machine inference whose price falls by an order of magnitude a year are collapsing the cost of seeing what is happening to nature, understanding it, and deciding what to do. Environmental harm is therefore increasingly a soluble problem of information and action. Institutions and physical intervention still lag, and information alone creates neither political will nor a repaired river. What has changed is that the knowing is no longer the expensive part.

Seven reasons survived, in three tiers. Tier 1, the decisive reasons: first, at the scale civilization now operates, nature can no longer be protected by unaided human cognition, and information is the resource that can close the gap; second, AI is becoming one of the most consequential forces in the biosphere’s future while nature has no representation in the machine layer, so the realistic threat is destruction by indifference, and the window for encoding values is open now; third, the physics and economics of machine intelligence are collapsing the cost of environmental cognition toward zero, undermining the oldest political obstacle to stewardship, but only if someone deliberately builds the environmental instance of that collapse.

Tier 2, strongly reinforcing: fourth, civilization must now build physical infrastructure at the fastest rate in its history and environmental process is a principal rate limiter, which makes ESI the only visible path where the buildout and nature both win; fifth, the biosphere is the largest store of survival-tested information known, mostly unread and being erased permanently; sixth, planetary environmental intelligence is emerging anyway, fragmented, thin, and not protection-first, so the real question is not whether it gets built but whose objective function it inherits.

Tier 3, the horizon reason: seventh, any future for life beyond Earth runs through biosphere competence that humanity demonstrably lacks and can only learn from Earth. The top two reasons turn out to be two faces of a single structure. Protecting nature now requires machine-scale cognition, and machine cognition left to its defaults omits nature. One build serves both, a coupling developed below.

The First-Principles Foundation

Six independently grounded foundations generate the seven reasons. They appear here once, allowing every reason that follows to be traced to physics and to the definition of optimization rather than to anyone’s preferences.

Life is the maintenance of improbable, functional configurations of matter against thermodynamic degradation, paid for with energy directed by information: sensing conditions and responding to them. Schrödinger’s What Is Life?, 1944, made this vivid as feeding on negative entropy; modern information thermodynamics makes the information-and-energy exchange exact. Environmental damage is the disruption of those configurations: matter and energy in the wrong place, time, form, or quantity. Protection is therefore, at bottom, a control problem: keeping a physical system inside viable bounds, before it is a legal, economic, or moral one. The legal and moral machinery sets those bounds. Which states of nature count as viable is a value choice, not a physical fact, and that machinery exists to serve the control function.

Where control is cheap. A control problem can be addressed at the decision point, before dispersal, or afterward. The asymmetry between those options is thermodynamic, not accidental: preventing a release means acting on a decision, using information; reversing one means re-concentrating matter that has mixed into air, water, soil, and tissue, and the minimum work of separation grows with dilution, while real-world costs run far above the minimum. The decision that prevents dispersal costs almost nothing by comparison with the work that undoes it. Even prevention’s full cost, mandated abatement included, runs far below remediation’s wherever both have been measured (Reason 3). Prevention beats remediation structurally, and the ratio worsens as pollutants disperse and persist.

At the decision point, prevention’s control layer is made of information operations: sensing states, predicting consequences, attributing causes, authorizing or withholding flows, and verifying outcomes. The physical act, whether closing a valve, denying a permit, or rerouting a process, is trivial. The real cost lies in knowing which valve to close, when and why to act, and whether the intervention worked. The abatement or forgone activity a decision mandates carries its own separate cost, which no amount of information erases. Whoever can perform these information operations cheaply and universally can, in principle, deliver prevention’s cognitive layer cheaply and universally. Better cognition also shrinks the physical layer by identifying the designs and targets that require less abatement for the same protection.

The capability condition is straightforward: a regulator can hold a system within bounds only if its capacity to sense, decide, and respond is commensurate with the variety and speed of the disturbances it must absorb. This is Ashby’s law of requisite variety (1956), the founding theorem of regulation in cybernetics. A regulator can also succeed by attenuating variety at the source through standards, bans, and design rules, and institutions have used that strategy to real effect. Yet the measured record below shows that the residual variety now swamps the response capability in place. Civilization’s disturbance-generating capacity has grown exponentially: millions of facilities, hundreds of thousands of chemicals, and machine-speed industrial processes. The controller has not: deliberate human cognition runs at roughly ten bits per second per person (Zheng & Meister, Neuron, 2024), while documents, meetings, and regulatory cycles connect it with latencies of months to years. When regulatory capability falls short of the variety it must govern, the theorem predicts not merely worse control but unabsorbed disturbances. That is precisely what the record that follows documents: a response lag between disturbance and effective action, not a failure of information itself.

Optimizers consume what they do not represent. Any optimizing system allocates resources according to its objective and its constraints; whatever appears in neither is priced implicitly at zero and treated as free input or free sink. This is not a claim about malice. It is what optimization is. It applies identically to markets (where it is called an externality), to bureaucracies, and to AI systems; the AI-safety literature’s instrumental-convergence results are its limit case for very capable optimizers. For any unrepresented system that lies within an optimizer’s reach and is usable as resource or sink, it predicts consumption, by indifference, not intent. The biosphere is exactly such a system: it is the planet’s accessible stock of energy, land, water, and matter.

The cost trajectory. The cost of information operations, including sensing, storing, transmitting, predicting and judging, has fallen on multi-decade exponentials and continues to fall, with no binding floor in sight for practical purposes. The cost of physical work, moving, separating and building matter, falls far more slowly: learning curves of tens of percent per doubling at best, bounded by material and energy floors. Therefore, over time, every function that can be moved from the matter column to the information column gets radically cheaper. Because prevention’s control layer is made entirely of information operations, it sits in the information column.

The seven reasons below show what these established ideas produce when they meet the measured state of the world in 2026. Each is stated with its evidence and with what it does not prove.

Tier 1: The Decisive Reasons

1. Nature can no longer be protected by human cognition alone, and information is the resource that closes the gap

Civilization changes the living world at machine speed and answers at the speed of paperwork. The distance between those two clocks is where the damage lives.

The claim is architectural before it is empirical. The controlled system comprises on the order of 1.5 million regulated facilities in the United States alone, more than 350,000 chemicals in commerce, and a continuous stream of authorization decisions governing molecular flows at industrial speed. The controller is a network of human minds, each processing deliberately at roughly ten bits per second, coupled through documents and meetings, with regulatory latencies measured in years. By the requisite-variety condition, this controller cannot absorb the disturbance load: roughly 10⁵–10⁶ environmental professionals deliberating at ~10 bits per second each, against millions of facilities operating continuously. By any plausible accounting, requisite variety fails by orders of magnitude. That is a prediction, and in 2026 it is no longer merely predicted. It is measured, in four ways.

The ledger is wrong, but the remedy is better measurement that reaches effective action. Environmental governance runs mostly on estimates and self-reports, and wherever independent measurement has been deployed at scale, the official numbers have failed. Roughly one million aerial site measurements found US oil and gas methane emissions running about three times the federal inventory, with a small fraction of sites (under 2%) producing half or more of well-site emissions (Sherwin et al., Nature, 2024); the IEA finds global energy-sector methane about 80% higher than what governments report to the UNFCCC (IEA Global Methane Tracker, 2025). The EPA’s own emission-factor system is built, by its own description, on averages such that “approximately half of the subject sources will have emission rates greater than the emission factor.” Of the more than 350,000 chemicals registered for production globally, about 120,000 cannot even be conclusively identified from public records (Wang et al., Environmental Science & Technology, 2020). These are signals of a measurement gap and a response lag, a capability not yet installed at the scale required. A protection system that does not know what is being emitted, by whom, in what quantity, is not managing the environment. It is managing a stale model of the environment.

The watchers cover a fraction of the field, and their capacity is falling. The EPA’s enforcement database tracks roughly 1.5 million regulated facilities. States, which run most delegated programs, conduct the large majority of inspections; the EPA itself performed nearly 8,300 in FY2025, a figure the agency presents as its second-highest in eight years. The coverage arithmetic is unchanged either way. Combined federal and state activity still reaches only a small fraction of the regulated universe in any year, and the sharper statistic is the response lag after violations are found: in the most recent year analyzed, only 12% of facilities with identified air-pollution violations and 2% with water violations drew any formal enforcement response, and over 3,000 facilities had sat in significant Clean Water Act noncompliance for three years or more (Environmental Integrity Project analysis of EPA’s own compliance data, 2026, an advocacy organization staffed largely by former EPA enforcement officials, working from the agency’s records). Capacity is moving in the wrong direction: EPA staffing fell roughly a quarter in a single year, from ~16,200 (January 2025) to ~12,200 (September 2025), below its 1988 headcount, with departures skewed toward the most senior scientists, and 31 states had already cut their environmental agency funding over the preceding decade. Physical coverage is thinner still: more than 70 million Americans live in counties with no regulatory air-quality monitors at all (American Lung Association, 2025), with higher estimates for fine-particle monitoring specifically; less than 0.001% of the deep seafloor has ever been visually observed; roughly 86% of species on Earth have never been described by science. When systematic PFAS testing of US drinking water finally began, contamination surfaced in the water serving more than 170 million Americans (EWG analysis of EPA testing data, 2026), a signal that had been present all along but remained outside effective response because it was unmeasured.

The clocks do not match. Detection has become fast, but response has not. Satellites now flag methane super-emitters within hours, yet roughly nine in ten of UNEP’s alerts to responsible parties still draw no response, an improvement from ninety-nine in a hundred two years earlier (UNEP International Methane Emissions Observatory). The signal is arriving sooner, but the capability to act on it has not kept pace. Deepwater Horizon flowed uncontrolled for 87 days; Aliso Canyon leaked for 112 days and was quantified only because research aircraft happened to fly. Meanwhile, the institutional clock runs in years: a 4.5-year average for a federal environmental impact statement, multi-year lags from violation to enforcement, and permit backlogs exceeding 20% in the water program. Environmental damage moves at the speed of physics; environmental protection still moves at the speed of paper.

The system being governed is degrading at record rates. Seven of nine planetary boundaries are now transgressed, with ocean acidification having joined in 2025 (Planetary Health Check, PIK). 2024 was the first calendar year above 1.5°C, at 1.55°C (WMO), and saw the largest single-year CO₂ increase on record. The seasonal peak reached 432 ppm in May 2026, the highest concentration in roughly three million years, while fossil CO₂ emissions set another record in 2025. Monitored wildlife populations have declined an average 73% since 1970 on the Living Planet Index. The honest caveat is that this geometric-mean metric is sensitive to its worst-declining populations, not a census of all animals. Pollution kills about 9 million people per year, one death in six, a figure essentially unchanged for a decade (Lancet Commission).

The decisive step in the argument is to read the successes honestly, including their limits. Environmental governance has won real victories, and they share a signature. The ozone layer is healing because a single measurable substance class was monitored globally and verified continuously. When unexpected CFC-11 emissions appeared in 2018, atmospheric networks detected them, attribution studies traced them to eastern China, and, through treaty machinery and domestic enforcement, global emissions returned to baseline within about two years (Montzka et al., Nature, 2021). Acid rain fell 96% because a binding cap was enforced over continuously monitored sources. CEMS made every ton visible and the allowance market trustworthy, producing benefit-cost ratios estimated near 40:1 for that program (Chestnut & Mills, 2005). Brazil’s DETER satellite alerts enabled enforcement that cut Amazon deforestation an estimated 59% below counterfactual over 2007–2011, with no measured loss of agricultural output.

But the same record teaches the limit, and a truth-seeking analysis must state it: measurement without consequences protects nothing. CO₂ is the most continuously measured substance on Earth, monitored since 1958 and now attributed source-by-source, yet it sits at record emissions because the coupling to consequences was never built. DETER makes the same point as a natural experiment. With the satellite system fully operational, deforestation surged when enforcement was withdrawn (2019–2022) and fell again when enforcement returned. Monitoring and will are multiplicative, not additive. What the successes establish is therefore precise: institutions win when a problem is visible, measured, attributable, and coupled to consequences, and ESI can manufacture the first three conditions everywhere, at machine cost, for problems where creating them through human labor is impossible (millions of small sources, hundreds of thousands of chemicals, diffuse flows, real time). The fourth condition remains political. Measurement does not replace will; it is the substrate that makes will effective, cheap to exercise, and, as DETER showed when enforcement returned, instantly re-armable.

What this reason does not prove: first, that information automatically produces protection. Information is the enabling resource, but it does not, and the political-will objection is treated at full strength in the Objections section. Second, this reason alone does not require superintelligence. Detection at scale is narrow automation plus budget, and satellites plus ordinary machine learning already do much of it. What narrow systems demonstrably do not do, as the evidence of Reason 6 shows, is integrate across media, exercise permitting-grade judgment, analyze cumulative impacts, and close the loop from detection through authorization to verification. The necessity claim here is for machine cognition at minimum; the case for the integrated, judgment-capable system is carried by Reasons 3, 4, and 6 together.

2. AI is becoming one of the most consequential forces in the biosphere’s future, and nature has no representation in the machine layer

An optimizer consumes whatever it cannot see. The most powerful new optimizers on Earth cannot yet see nature.

This reason has both a certain layer and a speculative one; the certain layer alone would justify it.

The certain layer: AI is already restructuring Earth’s metabolism. Data-center electricity consumption reached 415 TWh in 2024, grew 17% to 485 TWh in 2025, and is on track for roughly 950 TWh by 2030 (IEA), with data centers projected to absorb about half of all US electricity demand growth through 2030. Hyperscaler capital expenditure exceeded $400B in 2025, while analysts’ tallies of 2026 guidance are approaching $725B. Total data-center investment, roughly $580B in 2025, now exceeds the world’s entire investment in upstream oil, roughly $540B (IEA World Energy Investment, 2025). Gas-turbine order backlogs have reached 116 GW, with roughly a fifth of orders from data-center customers and delivery slots stretching toward the 2030s; hyperscalers are signing multi-decade power contracts and behind-the-meter gas arrangements to jump the queue. The companies most committed to AI report the consequences themselves: Google’s emissions are up 81% from its 2019 baseline per its 2026 environmental report; Microsoft’s latest verified disclosure shows emissions roughly 23% above its 2020 baseline, with its carbon-negative pledge acknowledged to be at risk. Google’s own Chief Sustainability Officer: “Our AI infrastructure buildout is accelerating faster than the grid is decarbonizing.” Meanwhile, investment in AI for the environment is roughly two orders of magnitude smaller than AI investment generally, at roughly ~$6B/yr in AI-related climate venture funding against ~$450B/yr in AI capex. The default trajectory, on current measured facts, is that AI’s environmental footprint scales faster than AI’s environmental benefit. This is not an argument that AI is net-bad. The IEA argues AI-enabled efficiency could ultimately cut more emissions than data centers add. It is an argument that nobody is steering, and that the steering layer, an intelligence whose explicit objective is the biosphere, does not exist. Information is increasingly the enabling resource for that layer, but it must be connected to effective action.

The sharp layer: indifference is the canonical failure mode. Optimizers consume what they do not represent. The AI-safety literature supplies the limit case. Its oldest and most-cited argument holds that a sufficiently capable optimizer pursuing almost any goal acquires resources, including “space, time, matter, and free energy” (Omohundro, 2008), as an instrumental drive. Bostrom’s Superintelligence (2014) formalized the failure mode as infrastructure profusion: an advanced system transforms Earth’s surface into goal-serving infrastructure not out of hostility, but because nothing in its objective says otherwise. Yudkowsky’s formulation is the most compact: the AI “does not hate you, nor does it love you, but you are made out of atoms which it can use for something else”. The biosphere is the planet’s accessible stock of energy, land, water, and matter. Hendrycks (2023) adds the competitive-dynamics version: selection pressure among AI systems favors expansion and resource acquisition regardless of any single lab’s intentions. These limit-case arguments are robust for idealized goal-directed optimizers and contested as descriptions of present systems (Narayanan & Kapoor, cited approvingly elsewhere in this report, reject the strong version of the frame). The dispute need not be settled here, because the mundane form of indifference, combined with a measured, exploding physical footprint, already suffices. One does not need an AI that hates nature, only powerful optimization for which nature is neither objective nor constraint. That is a precise description of the present configuration. The Future of Life Institute’s safety index evaluates frontier labs across 33 indicators; not one is ecological. No published lab constitution encodes conservation laws, planetary boundaries, or ecosystem health. The nearest approach, Anthropic’s constitution now weighs “the welfare of animals and of all sentient beings”, is significant for a different reason: it proves the values layer of frontier AI is being actively written right now, and that the biosphere’s absence from it is a vacancy, not a law of nature. A small independent literature has begun naming the same vacancy (Korecki’s “Biospheric AI,” 2024, arguing anthropocentric and sentientist alignment are “not sufficient… to ensure that AI does not damage” the biosphere; Zeng et al., 2025, proposing co-alignment toward a “sustainable symbiotic society”; Lenton & Latour’s “Gaia 2.0,” Science, 2018).

How building ESI actually addresses this, stated carefully because the causal chain matters more than the slogan:

Legibility. Optimizers consume what they do not represent: by omission, what is absent from objectives and constraints is destroyed, and nothing can enter an objective or a constraint that has not first been made legible: measured, modeled, attributable. Today nature remains largely illegible to the machine economy. It is unmeasured, unpriced, and underrepresented relative to its causal importance, a capability gap that leaves signals unable to reach effective decisions. ESI is the project of making nature legible, sensed, modeled, valued, and machine-readable, so that what the machine layer affects, someone or something can be held to account for. Honesty requires the corollary that legibility is dual-use. The map of what lives where is also a prospecting map. Scott’s Seeing Like a State is the canonical warning, and wildlife platforms like eBird and iNaturalist already obscure locations of threatened species because exploitation of open data is real. The design consequence is asymmetric disclosure: aggregate visibility for accountability, with access controls on exploitation-grade resolution. This dual-use fact strengthens Reason 6: a legibility layer will be built by someone, and only a protection-first builder has reason to engineer the asymmetry.

Value seeding in the lock-in window. Values, data, and defaults embedded in early powerful systems tend to propagate through training corpora, institutional adoption, and the demonstrated fact that model outputs shape successor models’ distributions (MacAskill’s lock-in argument, 2022; Shumailov et al., Nature, 2024, cited here only for the propagation dynamic, not as evidence that values persist benignly). Constitutions are being drafted and revised now. A world in which rich, rigorous, real-time environmental ground truth exists and is woven into the infrastructure AIs are built on is a world with a different training distribution and a different default toolset for every future model than one where nature remains a rounding error in the corpus.

Representation at machine speed. Every other major interest in the AI ecosystem has agents that operate at machine speed: markets, militaries, advertisers. The biosphere has none. Proposals for AI whose purpose is safeguarding rather than pursuit have entered the safety mainstream, including Bengio et al.’s non-agentic “Scientist AI” (2025) and Buterin’s defensive acceleration. ESI extends that pattern to the one system everything else depends on. The extension must be stated honestly: where ESI acts, in permitting, compliance, and response, it is agentic, whereas Scientist AI deliberately is not. Agency is an additional safety burden, and it is exactly what the automation-failure and capture demands in the Objections section exist to constrain.

Grounding. The two-way conjecture, flagged as such. The most speculative claim this analysis retains is that nature is the longest-running record of many optimizing agents persisting under shared physical constraint, and that grounding AI objectives in physical law and biospheric dynamics, rather than in human preference alone, is a candidate route to alignment in general, not just alignment toward nature. Independent theoretical work points in this direction, including Levin on multi-scale competency and widening circles of care and Friston et al. on ecosystems of intelligence. The thesis is assessed here as an open research question, not a result. Nature’s stabilizing mechanism, selection by death, is not directly available as a training method; knowledge about nature does not by itself produce care for it, reflecting the same orthogonality that powers the threat model above; and no published alignment program yet treats biosphere-grounding as a working method. What is not speculative is that the question cannot be investigated without the instrument: the data, models, and ground truth that ESI would build are the prerequisite for testing whether physics-grounded alignment works at all. If the conjecture is even partially right, ESI is upstream of alignment generally; if it is wrong, ESI still stands on the three mechanisms above.

What this reason does not prove: that building one nature-aligned system will align the others. It will not, and claims otherwise fail standard critiques, including Critch on pivotal acts and Hendrycks on selection favoring the ruthless. Nor can one rely on Lovelock’s comforting Novacene argument that machine intelligence will preserve the biosphere out of thermal self-interest. Orbital data-center programs already in development show that silicon does not, in the limit, need a biosphere at all. Nor is “nature” a unified principle with a clean utility function. Choosing baselines and tradeoffs is governance, not physics, and pretending otherwise would discredit the project. The defensible claim is narrower, yet still enormous: the realistic path by which advanced AI damages nature is indifference amplified by scale; indifference is a representation failure; and representation is buildable, but only during the window in which the machine layer’s values and infrastructure are still wet cement. That window is open now and will not stay open.

3. The cost of environmental cognition is collapsing toward zero, undermining the oldest obstacle to stewardship, if someone builds for it

For all of industrial history, knowing what was happening to nature was the expensive part. That price is now falling faster than any cost curve in environmental history.

For all of industrial history, environmental protection was governed by a brutal economic fact: protection required costly human vigilance and costly physical remediation, so every unit of it was rationed, litigated, and resented. That constraint is contingent, not permanent. As information makes protection more capable and less costly, the measured world now confirms the change at every scale where it has been tested.

The asymmetry between prevention and remediation is extreme, and at the limit unbridgeable. Control is far cheaper at the decision point, before dispersal, than after it. Monitored prevention under the Clean Air Act returned benefits estimated near $2 trillion per year against $65 billion in costs by 2020, a central estimate of roughly 30:1, with a published range of 3:1 to 90:1, dominated by monetized mortality reductions (EPA Second Prospective Study, 2011). At the other pole, a peer-reviewed bounding exercise estimated that treating ongoing PFAS releases down to health-based levels would cost on the order of gross world product or far beyond it: $20 trillion to $7,000 trillion per year, a single-author, order-of-magnitude bound cited as such (Ling, Science of the Total Environment, 2024), for one chemical class with a ~$28B/yr market. For dispersed, persistent pollutants, remediation is not expensive; it is economically infeasible at any plausible scale. Only source control works, and because prevention’s control layer is made of information, that means knowing, predicting, authorizing, and verifying. Superfund sites average roughly two decades each; of every dollar spent on asbestos litigation, about 42 cents reached a claimant (RAND). Civilization has run the experiment both ways, many times, and the answer is uniform: bits before bonds; foresight before shovels.

What has kept prevention undersupplied is its dependence on one of the economy’s scarcest inputs: expert judgment. The US environmental industry is large, with roughly $573B/yr in revenues, but its analytic core is thin: about 130,000 environmental scientists and engineers (BLS), and on the order of $55B/yr of consulting judgment worldwide, set against modeled welfare losses that run to trillions per year in single categories, including air pollution at ~$8.1T/yr in health damages (World Bank), lead at ~$6T/yr (Lancet Planetary Health), and the IMF’s $7T fossil-fuel figure, of which $5.7T is implicit unpriced externalities. These categories partially overlap and are welfare estimates, not market flows, so no single ratio should be quoted from them. The comparison is not commensurable as arithmetic, but it is sufficient to establish the structural point: wherever cognition rather than capital binds, whether in vigilance, verification, analysis, design, or process, protection is rationed by the scarcity of expert judgment.

That input is information, whose price is now in free fall. Inference cost at constant capability has been falling roughly 9–900× per year, depending on the benchmark. GPT-3.5-level capability fell ~280× in under two years (Stanford AI Index; Epoch AI). On OpenAI’s GDPval benchmark, frontier models matched or beat human experts on just under half of tested professional tasks at roughly 100× the speed and 1/100 the cost (the vendor’s own figures, with its stated caveats). On legal document analysis, the nearest measured analogue to permitting and compliance work, top systems outperformed a practicing-lawyer baseline on document Q&A, 94.8% versus 70.1% (Vals, 2025). The sensing stack has collapsed in parallel: PM2.5 monitoring from $10,000+ reference stations to ~$260 sensors; satellite imagery from expert procurement to free public archives; genomic sequencing down five orders of magnitude. And the endgame has precedents. GPS, weather forecasting, and Landsat each began as expensive expert services and became free background utilities whose measured public value ($1.4T cumulative for GPS; ~$31.5B/yr for weather forecasts against ~$5B of spend) exceeds their cost by orders of magnitude, with usage exploding precisely when the price hit zero. Environmental protection’s cognitive layer is a candidate member of that class: from professional service, rationed by the 130,000, to civilizational utility, ambient as timekeeping.

The political consequence is the deepest part of this argument because it bears directly on the strongest objection to the whole project (see Objections): that environmental failure is a failure of will, not information. The honest statement is that will and cost are mutually endogenous. Policy bought solar down the learning curve through feed-in tariffs, industrial strategy, and subsidies; cheap solar then transformed the politics of climate in a way three decades of accurate reports never did. The fast energy transition now projects roughly $12 trillion cheaper than the slow one (Way et al., Joule, 2022). The loop compounds in both directions, and ESI’s build case is precisely to start that loop for protection’s cognitive costs. The bound must be stated with equal honesty: ESI directly collapses the costs of vigilance, verification, process, analysis, and design, a large but partial share of total protection cost. Physical abatement (scrubbers, cleanups, closures, foregone extraction) follows its own cost curves, although cheaper cognition accelerates those too through better design and targeting. The political claim is correspondingly bounded: collapsing cognitive cost removes the delay, uncertainty, and transaction-cost components of resistance to protection, not the capital components. Reports alone have rarely beaten entrenched interests; falling costs repeatedly have, and the two compound.

What this reason does not prove: that the collapse arrives by itself in this domain. It will not. The cost collapse in general-purpose intelligence flows to whatever pays, including trading, advertising, mineral exploration, and compliance-minimization, unless the environmental instance (curated regulatory corpora, permitted-workflow agents, sensor fusion, verified outcomes) is deliberately engineered. That is precisely the build case, and it is why this reason ranks in Tier 1: the door is open, but nothing walks through it automatically. Rebound effects are real but empirically bounded in the energy literature (typically 5–40%, rarely reversing gains); “compliance rebound”, cheaper permitting accelerating extraction itself, is unstudied and must be handled by the protection-first design demands set out in the Objections section.

Tier 2: The Strongly Reinforcing Reasons

4. The largest buildout in history is rate-limited by environmental process, and ESI is the only visible path where the buildout and nature both win

The fastest construction boom on record runs straight through environmental review. Either the boom stalls or the review is cut away, unless something can be rigorous at speed.

Civilization has committed itself, for climate reasons and for AI reasons alike, to the fastest infrastructure buildout it has ever attempted: rebuilt grids, gigawatts of clean generation, transmission, mines for critical minerals, and the data centers of the intelligence economy itself. On the evidence, a principal rate limiter is the environmental and interconnection process, where the delay between signal and effective response remains substantial. Roughly two terawatts of generation and storage (more than the entire existing US grid) wait in interconnection queues, with median waits now above five years and a historical completion rate of about 13% (LBNL Queued Up, 2025–2026). The US built 55 miles of high-voltage transmission in 2023 against a needed doubling of pace (Grid Strategies). A US mine averages 29 years from discovery to production, second slowest of any nation surveyed (S&P Global), while the IEA projects a potential 30% copper supply shortfall by 2035. The average environmental impact statement takes 4.5 years; 94% of surveyed clean-energy developers report projects delayed or abandoned due to federal permitting in the past year (Crux, 2026); roughly $121B in wind and solar investment is at risk from stalled federal review (Wood Mackenzie, 2026). The loop closes on itself: the AI buildout is jammed in the same queues, which is part of why gas turbines with delivery dates years out are being bought instead.

Two defaults are currently on offer, and both are bad for nature. Default one, the live 2025–2026 trajectory, strips away the review layer: NEPA implementing regulations rescinded, EPA staffing reduced by a quarter, and review discretion applied selectively. That buys speed by seeing less, while nature absorbs the error. Default two is to keep the friction, in which case the energy transition arrives late and nature absorbs the climate instead. Honesty requires conceding that the deregulatory movement is right about one fact, namely that review is too slow, while rejecting its remedy of seeing less. The third path exists only with machine cognition: reviews that take days and see more, including every document, every precedent, and every cumulative impact, with live dispersion modeling instead of decades-old emission factors; and permits that become dynamic, continuously verified operating envelopes rather than five-year paper snapshots. “Rigor at speed” is a contradiction for human institutions and a design specification for ESI. This is also where the thesis meets its first market tests: automated permitting and compliance workflows are already in commercial development with major industrial operators (EnviroAI among the first movers), an arrangement whose structural risk is confronted directly in the capture objection below because it is real.

What this reason does not prove: documents are the whole bottleneck. They are not. Post-permit engineering, interconnection studies, financing, and local opposition bind too (RFF; Niskanen), and only a minority of projects ever touch EIS-level review. Nor is speed directionally virtuous: a faster process accelerates whatever is in the queue, gas as readily as solar. Both caveats point in the same direction. The fast system must be built protection-first, with the objective chosen at the foundation, which is Reason 6, and its integrity guaranteed by the architectural demands in the Objections section.

5. The biosphere is the largest store of survival-tested information known, mostly unread, and being erased permanently

Four billion years of tested solutions are written in living tissue. Most of the library has never been read, and burned volumes cannot be rewritten.

Life on Earth encodes roughly 5×10³¹ megabases of DNA (Landenmark et al., PLOS Biology, 2015). This is the compressed output of four billion years of the largest optimization process ever run, with every extant genome a solution validated against physics by unbroken survival. Humanity has read almost none of it: ~86% of species have never been described; of the 1.67 million species even named, about 0.2% have reference-quality genomes sequenced (Earth BioGenome Project, 2025). The value of what has been read is not speculative. Roughly a third of approved small-molecule drugs are natural products or direct derivatives, with nearly two-thirds of anticancer agents natural-product-connected (Newman & Cragg, 2020). E.O. Wilson’s framing was precise: each species is richer in information than any human artwork, and mass extinction is “the folly our descendants are least likely to forgive us” because it is the one folly that is strictly irreversible. Extinction rates now run tens to hundreds of times background (IPBES). The library is burning faster than at any time in 65 million years, with the vast majority of volumes unopened.

The economics literature contains an instructive dispute. The classic skeptical result (Simpson, Sedjo & Reid, 1996) found the bioprospecting value of the marginal species to be tiny, a real challenge to information-based conservation arguments. The rebuttal (Rausser & Small, 2000) showed that the value returns and becomes large when search is information-guided, when you know enough about the library to know where to look. Read carefully, the exchange yields two lessons, not one. First, biological information is valuable in proportion to the intelligence applied to it, which is the rebuttal’s own mechanism. Second, guided search concentrates value in promising leads. That is why the operative argument for protecting the whole library is option value under ignorance: you cannot know in advance which volumes the guided search will need, and only preservation keeps the option alive. A superintelligence is the only adequate reader of a library this size, and reading requires that the library still exist. Even an unbounded future intelligence cannot recompute a lost lineage, because the evolutionary test cycles are not re-runnable. This is also where this reason joins Reason 2: if nature holds transferable insight about durable multi-agent complexity, the open question flagged there, its information is not merely pharmacologically valuable; it may be part of the alignment curriculum.

What this reason does not prove: a specific triage. Option-value arguments license preservation in general, not any particular tradeoff, and “protect the information” must not collapse into “sequence it and let the living system go”: the information is inseparable from the dynamics (a genome without its ecological context is a word without a language). Its clock is slower than the political and values windows of Tier 1, which is why it ranks here. But among all seven reasons, its losses are the only ones that are strictly, physically permanent.

6. Planetary environmental intelligence is being built anyway: fragmented, thin, and not protection-first. The open question is whose objective function it inherits

A planetary nervous system is already being assembled out of parts built for other purposes. Whoever founds it chooses what it will care about.

The research is unambiguous on one point: the components of planetary environmental intelligence already exist and work. AI weather models are operational at the world’s leading forecast agency and in consumer products used by billions (ECMWF AIFS, 2025; Google WeatherNext). Climate TRACE tracks 2.77 million emission sources, refreshed monthly. Global Fishing Watch announced in 2026 that it had mapped the world’s industrial fishing fleet, after its 2024 Nature study revealed that roughly three-quarters of industrial fishing vessels had been publicly untracked. Commercial methane satellites now make on the order of a hundred thousand targeted facility-scale observations a year, atop the ~1 million aerial site measurements amassed by survey aircraft. Yet every one of these systems is single-domain, none closes the loop from detection to protection, and the action layer, permitting, compliance, enforcement, remains the least developed of all. A 2026 independent assessment of the EPA’s AI program judged practical adoption “much more aspirational than real,” and the US government’s flagship permitting-AI effort remains in beta. The capability is also thinly redundant: when the MethaneSAT satellite failed in June 2025, its open-data, basin-scale public mission had no replacement. Coverage follows money, not need, leaving the data deserts precisely in the Global South regions where biodiversity is highest. And the capital asymmetry is roughly two orders of magnitude: ~$450B/yr of AI infrastructure capex against ~$6B/yr of AI-for-climate investment, with the flagship philanthropy a $100M program.

Meanwhile, the intellectual case has been made independently, from separate disciplines, for a generation, but never operationalized. Astrobiology holds that a civilization survives only by developing “planetary intelligence,” intentional feedback loops between knowledge of planetary-scale consequences and behavior (Frank, Grinspoon & Walker, 2022). Earth science: Lenton and Latour’s “Gaia 2.0” (Science, 2018) argues that deliberate planetary self-regulation depends on continuously monitoring and modeling the planet’s state, making sensor infrastructure the practical route to planetary self-awareness. Philosophy of computation frames planetary-scale computation as the accidental megastructure that must be turned to governing the planet that produced it (Bratton). Even frontier-lab leadership now names planetary problems among the top justifications for superintelligence itself (Suleyman, 2025). Everyone names the destination; no one has built the integrated system. As of this research, the only organization found with a stated program to build integrated, protection-first environmental intelligence is EnviroAI (enviro.ai), the effort this analysis was prepared to inform. That relationship is disclosed here so the reader can weight the finding accordingly; the search is also necessarily absence-of-evidence, not proof of absence.

This is a reason in its own right because of what happens by default. If integration occurs ad hoc, as economic and scientific pressure encourages, and is assembled from commercial systems whose objectives are engagement, throughput, extraction, or compliance-minimization, the planetary layer inherits those objectives. Optimizers consume what they do not represent, by indifference rather than intent. That principle applies one level up. Infrastructure is philosophy made durable. “Protection-first” is not a slogan here; it has architectural content, supplied by the Objections section below: measurement structurally separated from advocacy, asymmetric disclosure engineered against dual use, contestability and human authority over adverse actions, outcomes published and independently verifiable. A system with those properties at its foundation behaves permanently differently from one retrofitted toward them, even if their components are identical. The choice, therefore, is not whether planetary-scale environmental intelligence emerges; it is which objective function it is founded on, and by whom. That is the entire content of this reason: deliberately, now, with protection as the founding objective.

What this reason does not prove: that a purpose-built system stays pure. It does not, because capture is a permanent pressure rather than a one-time filter, and this reason holds only alongside the governance architecture demanded in the Objections section.

Tier 3: The Horizon Reason

7. Any future for life beyond Earth runs through biosphere competence, which humanity demonstrably lacks and can learn only from Earth

Every world we might inhabit is a closed-book exam in planetary life support. Earth is the only place to study.

The record here is short and humbling. Biosphere 2, roughly $200 million, 3.14 acres, and the most ambitious closed-ecosystem experiment ever run, could not keep eight people alive for two years without intervention. Oxygen fell from 20.9% to ~14.5% for reasons (soil microbial respiration coupled to CO₂ absorption by structural concrete) that were diagnosed only after the fact; 19 of 25 introduced vertebrate species went extinct, and virtually all pollinators died. The verdict published in Science in 1996 remains unrefuted three decades later: “No one yet knows how to engineer systems that provide humans with the life-supporting services that natural ecosystems produce for free” (Cohen & Tilman). The engineering alternative has fared little better at closure. After 25 years, the ISS recycles 98% of its water but recovers only about half of its oxygen from CO₂, against a >75% requirement for Mars-class missions; it grows no food and remains a resupplied outpost, not a biosphere. ESA’s MELiSSA ecological life-support loop is 37 years into development and still at pilot stage. Every closure attempt has failed on the biology, not the plumbing.

The implication runs in both directions. Any settlement of Mars, any long-duration habitat, and any eventual terraforming would constitute applied planetary ecology: a real-time control problem involving coupled biogeochemical loops that no one currently knows how to solve. Earth is its only full-scale reference implementation, training set, and testbed. An ESI built to protect Earth is also the instrument through which we learn to understand and run biospheres. Fully instrumented, closed-loop ecological testbeds are the natural laboratories in which that competence can be proven before it is needed. The reframing matters as much as the capability. It recasts environmental protection from a constraint on expansion into its prerequisite competence. A civilization that cannot keep its own biosphere thriving is not ready to carry life anywhere else; one that masters planetary stewardship carries the garden with it.

What this reason does not prove: that its urgency runs on the same clock as Tier 1, or that it strictly requires ESI. Closed-loop ecological research could proceed without it, though an Earth-scale environmental intelligence is the natural training ground and the fastest route to competence. It is the longest-horizon reason and ranks accordingly. But it is the reason that gives the project its full meaning. Protection is not the end state; it is the apprenticeship.

The Ranking at a Glance

#	Reason	Stakes	Evidence	Requires ESI	Window
1	Protection has become a problem of information at a scale beyond unaided human capability	Extreme	Very strong, measured	Machine cognition at minimum; the full ESI case is carried with Reasons 3, 4 and 6	Closing: degradation compounds
2	Nature needs representation in the machine layer	Extreme	Strong on footprint (measured); argument-grade on remedy mechanisms	Yes: no other candidate exists	Closing: values are being written now
3	The collapse in cognition cost undermines growth versus protection	Very high	Strong: observed cost curves, with physics as framing	Yes: the domain instance must be built	Partly: the first builder shapes the utility
4	The buildout is rate-limited by environmental process	Very high	Very strong, measured	Largely: rigor at speed needs machine cognition	Closing: transition timing
5	The biosphere is irreplaceable information	Very high	Strong	Yes: the only adequate reader and guardian	Gradual: losses are permanent
6	The default trajectory is fragmented and not protection-first	High	Strong	By definition	Closing: founding window
7	Life beyond Earth requires biosphere competence	High, long-run	Strong on the failure record	Partly: an Earth-run ESI is the natural training ground	Not closing, but the learning takes decades

Reasons Considered and Set Aside

Maximum truth-seeking requires showing the filter, not just its survivors. Several candidate reasons were examined and deliberately not ranked among the most compelling. Human health alone (9 million pollution deaths a year) is morally sufficient but analytically subsumed. It is the strongest evidence inside Reason 1, and taken alone it justifies better environmental tools without requiring the integrated planetary system. Commercial opportunity (a $573B US environmental industry ripe for automation) is real and funds the mission, but it is a business case, not a civilizational one. Leaning on it also invites the capture objection below. Generic AI-for-science optimism (“AI will discover fixes anyway”) is too vague to rank and slides toward the moral-hazard failure mode. National competitiveness framings are double-edged and subordinate the objective to whichever state wields it. Nature’s intrinsic value, the conviction that life is worth protecting for its own sake, was set aside from the ranking for the opposite reason: it is not one reason among seven but the foundation beneath all of them. The seven are addressed to any decision-maker, regardless of whether they share that conviction.

The Synthesis

Seven reasons resolve into one asymmetry and one choice.

The asymmetry is one of speed. Civilization now changes nature at machine speed and answers at paper speed. Every shortfall measured in this analysis lives in the gap between those two clocks: inventories three times off, nine in ten alerts unanswered, 4.5-year reviews, a quarter of the watchers gone in a year, seven of nine planetary boundaries breached. The bottleneck is the delay between signal and effective response, and that delay is newly tractable. Cheap sensors, continuous orbital observation, and inference at collapsing cost mean the seeing and the understanding can now happen in something close to real time. Human institutions cannot close a planetary delay unaided; machine intelligence is the instrument that can, and the price of that instrument is falling faster than any cost curve in the history of environmental work. Protection at planetary scale runs on information, and for the first time, cognition is the thing civilization knows how to make abundant.

The choice: machine intelligence is being woven into the planet’s metabolism regardless. That means 485 terawatt-hours rising toward a thousand, capex guided toward three-quarters of a trillion dollars a year, and values drafted into constitutions this year. The only open variable is whether, within the emerging machine layer, there exists an intelligence whose objective is the living world: one that makes nature legible and accountable to every optimizer, represents it at the speed of the systems that could consume it, and preserves the four-billion-year library that intelligence itself may need.

The top two reasons are two faces of this single structure: protect nature with intelligence, and make intelligence something nature survives. Both are achieved by the same build. In simplest form:

Nature is the system that turns energy into life. It now shares a planet with a second system that turns energy into intelligence. Environmental superintelligence is the layer built to let each see the other, because what the machine age cannot see, it will not keep.

Key Sources

First principles: Schrödinger, What Is Life? (1944); Ashby, An Introduction to Cybernetics (1956), the law of requisite variety; Landauer (1961) and Sagawa & Ueda (2008–2012) on the thermodynamics of information; Zheng & Meister, “The unbearable slowness of being,” Neuron (2024).

State of nature and governance: Planetary Health Check 2025 (PIK; seven of nine boundaries); WMO State of the Global Climate (2025, 2026); WWF/ZSL Living Planet Report 2024; IPBES Global Assessment (2019); Lancet Commission on Pollution and Health (2022); Sherwin et al., Nature 626 (2024); IEA Global Methane Tracker (2025); Wang et al., ES&T (2020); Environmental Integrity Project enforcement analysis (Feb 2026); Environmental Protection Network EPA staffing data (2025); American Lung Association State of the Air (2025); Ocean Discovery League, Science Advances (2025); EWG/EPA UCMR5 PFAS data (2026); UNEP International Methane Emissions Observatory.

Successes and their limits: Montzka et al., Nature 590 (2021), on CFC-11 detection and reversal; EPA Acid Rain Program results; Chestnut & Mills (2005); Assunção, Gandour & Rocha on DETER deforestation deterrence; EPA Second Prospective Study of the Clean Air Act (2011).

AI and alignment: IEA Energy and AI (2025) and Key Questions on Energy and AI (2026); IEA World Energy Investment (2025); Google 2026 Environmental Report; Microsoft sustainability disclosures; Omohundro, “The Basic AI Drives” (2008); Bostrom, Superintelligence (2014); Hendrycks, arXiv:2303.16200 (2023); Korecki, “Biospheric AI,” arXiv:2401.17805 (2024); Zeng et al., arXiv:2504.17404 (2025); Lenton & Latour, “Gaia 2.0,” Science 361 (2018); Frank, Grinspoon & Walker, Int. J. Astrobiology (2022); MacAskill, What We Owe the Future (2022); Shumailov et al., Nature 631 (2024); Anthropic constitution (2025–26); Bengio et al., arXiv:2502.15657 (2025); Levin, Frontiers in Systems Neuroscience (2022); Friston et al., Collective Intelligence (2024); Scott, Seeing Like a State (1998); Future of Life Institute AI Safety Index (2025).

Economics and throughput: World Bank air pollution cost (2022); Larsen & Sánchez-Triana, Lancet Planetary Health (2023); IMF fossil subsidies WP/2023/169; Ling, Science of the Total Environment 918 (2024); Way, Ives, Mealy & Farmer, Joule (2022); LBNL Queued Up (2025, 2026); CEQ EIS timelines (2020); Grid Strategies transmission data (2024); S&P Global mine-development timelines (2024); IEA Global Critical Minerals Outlook (2025); Wood Mackenzie via Reuters (2026); Crux developer survey (2026); Stanford AI Index (2025); Epoch AI inference-price trends (2025); OpenAI GDPval (2025); Vals Legal AI Report (2025); NIST/RTI GPS economic study (2019); Lazo et al., BAMS (2009); USGS Landsat valuation (2019); Environmental Business Journal (2026); BLS Occupational Outlook (2025); RAND asbestos litigation study (2005).

Long horizon: Cohen & Tilman, Science 274 (1996); Severinghaus et al., Eos (1994); NASA ECLSS water-recovery milestone (2023) and oxygen-recovery program; Landenmark, Forgan & Cockell, PLOS Biology (2015); Mora et al., PLOS Biology (2011); Earth BioGenome Project Phase II, Frontiers in Science (2025); Newman & Cragg, J. Nat. Prod. 83 (2020); Simpson, Sedjo & Reid, JPE (1996); Rausser & Small, JPE (2000).

Red team: Supran, Rahmstorf & Oreskes, Science (2023); Kahan et al., Nature Climate Change (2012); Mildenberger, Carbon Captured (2020); Berg, Kölbel & Rigobon, Review of Finance (2022); Verra offset investigations (2023); Zou, American Economic Review (2021); Karplus, Zhang & Almond, PNAS (2018); Dutch childcare benefits scandal and Robodebt Royal Commission records; Charlotin AI Hallucination Cases database (2026); Dahl et al., Stanford RegLab (2024); McLaren, Climatic Change (2020); Narayanan & Kapoor, “AI as Normal Technology” (2025); Ostrom, polycentric governance (2009); Fast Company AI-climate claims audit (2026).

Prepared August 18, 2026. This analysis was performed independently: candidate reasons were generated and ranked from first principles and external evidence. Statistics are current as of publication per the cited sources; an independent verification pass confirmed the large majority of claims against primary sources, and the corrections it required are incorporated. Claims flagged as contested, estimate-grade, vendor-reported, or conjectural are so labeled deliberately. The analysis asserts necessity, not sufficiency: it is an argument for building environmental superintelligence truthfully, verifiably, and protection-first, not a warranty that any particular build succeeds.