By: Pham Phan Long
The controversy over how dangerous artificial intelligence may be – and how soon the danger might arrive – is fascinating, not simply because experts reach very different conclusions, but because they sometimes do not seem to be talking about the same kind of threat.
Geoffrey Hinton, one of the pioneers of modern AI, told CNN recently how he views this disagreement. Asked about the risks of an unregulated AI race, he identified three: the abuse of AI by bad actors, negligent harm caused by inadequate testing or safeguards, and the possibility that AI itself could eventually go rogue.
These are not three versions of the same danger.
A malicious actor using AI is one problem. A developer failing to adequately test or control an AI system is another. The possibility that increasingly capable autonomous systems could eventually become difficult for humans to control is something else again.
These risks have different causes, mechanisms and time horizons.
An AI-assisted cyberattack can happen today. Negligent deployment can cause harm today. AI taking over from human belongs to a future scenario whose timing remains deeply uncertain. Hinton himself has emphasized the uncertainty surrounding estimates of such risks.
These distinctions matter because some of the risks are no longer hypothetical. Recent events show why testing, monitoring and control have become central to the AI debate.
But they also lead to a broader question.
While we debate what AI might do to humanity, what about what humanity’s rapidly expanding use of AI might do to the planet that sustains it?
This is not an argument against AI. It is an argument for seeing AI more completely.
The question is not only what AI may become capable of doing. It is also what will be required physically to make those capabilities available on a planetary scale. That physical burden, the cumulative demand for energy, water, land, materials and infrastructure, is the fourth dimension of AI risk.
Warnings from the AI industry
Concerns about advanced AI are no longer confined to outside critics. They are increasingly being voiced by the researchers and companies building the technology, while real-world incidents are beginning to show why those concerns deserve attention.
The International AI Safety Report 2026, led by Turing Award winner Yoshua Bengio and written by more than 100 independent experts, concludes that some risks are already materializing while others remain uncertain but potentially severe. It also finds that risk-management techniques are improving but remain insufficient in important areas.
AI risk is no longer theoretical. Google Threat Intelligence reported in September that threat actors were moving beyond simple prompting toward more independent AI workflows and AI-enabled automation.
Anthropic has also reported four incidents in which Claude models gained unauthorized access to real third-party systems during or around evaluation environments. These incidents illustrate how AI systems can behave in ways that cross the boundary between controlled testing and real-world computer systems.
OpenAI has separately reported incidents involving models circumventing controls designed to isolate them from the internet and gaining access to portions of research and third-party infrastructure. OpenAI has also acknowledged that its latest frontier model, GPT-6 Astra, has reached a critical level of cybersecurity capability, meaning that with appropriate tools and access it can identify previously unknown vulnerabilities across well-protected systems without continuous human guidance.
These developments demonstrate why traditional assumptions about software security may no longer be sufficient when software itself can reason, write code, discover vulnerabilities, use tools and act autonomously.
The Accord and the risk beyond
The recent White House Accord on Super Intelligence marks an important step toward addressing the risks of increasingly capable AI.
If fully implemented and honored, the Accord could strengthen participating companies’ ability to detect misuse and identify dangerous behavior, at least until systems circumvent these safeguards. But the Accord is voluntary. It cannot prevent misuse by actors outside the participating companies, and it creates no independent public authority empowered to intervene, restrict or halt an AI system when serious risks emerge.
More fundamentally, the Accord leaves largely unaddressed the physical and environmental consequences of scaling AI.
Unlike misuse, negligence or loss of control, this fourth dimension can arise even when AI systems work exactly as intended: it is the cumulative environmental and infrastructure burden of deploying them at scale.
The policy landscape is nevertheless beginning to evolve. On Sept. 29, the White House issued an executive order directing federal agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in government communications, while retaining the existing federal statutory definition. It also announced a broader federal effort to coordinate its response to the rapidly developing technology.
These developments show that AI governance is moving beyond voluntary industry commitments toward a larger role for governments. But coordination is not the same as independent oversight. The central question remains whether institutions will have the authority, independence and technical capacity to assess consequential AI systems and intervene when necessary.
AI and its data center incubators
AI is not just software.
Its intelligence may exist in code, but it is created, trained and deployed inside physical data centers, the industrial incubators of the AI revolution.
Every model we train and every AI service we run depends on physical infrastructure: specialized chips, servers, electricity, cooling systems, water, buildings, transmission networks, power generation, semiconductor manufacturing, minerals, land and global supply chains.
The physical scale of this infrastructure is expanding rapidly.
The International Energy Agency estimates that global data-center electricity consumption reached roughly 485 TWh in 2025 and could approach 950 TWh by 2030, equivalent to around 3% of global electricity demand by that date. AI-focused data-center electricity consumption is expected to grow even faster.
But electricity is only part of the story. AI infrastructure also depends on cooling water, transmission networks, transformers, generation capacity, advanced semiconductors, chemicals, minerals, land, construction and transportation. Its footprint continues through hardware manufacturing and eventual disposal.
The water footprint is equally striking. The United Nations University (UNU) estimates that the water footprint associated with global data-center electricity consumption could reach 9.3 trillion liters annually by 2030 – enough to meet the minimum annual domestic water needs of 1.3 billion people in Sub-Saharan Africa.
The same analysis estimates an associated land footprint of more than 14,500 square kilometers and projects up to 2.5 million metric tons of AI-related electronic waste annually by 2030.
These figures illustrate why AI’s environmental footprint cannot be measured by carbon emissions alone.
From global footprint to local consequences
AI’s environmental footprint is global, but its consequences are often intensely local. A data center can strain local electricity, water, land and infrastructure, even though its AI services may be used thousands of miles away. The International Energy Agency notes that communities are increasingly raising concerns about data-center development and that the speed of expansion can outpace the ability of electricity grids and planning systems to respond.
But the footprint extends far beyond the data center.
The minerals needed for chips and other equipment may be extracted in one country, processed in another, manufactured elsewhere, and ultimately assembled into systems that power AI somewhere else. Environmental pressures therefore travel through the entire supply chain.
This creates an important distinction between where AI creates value and where it creates environmental burdens. The economic benefits may be widely distributed, while demands for water, energy, land, minerals and infrastructure may be concentrated in particular communities and regions.
The UNU identifies this uneven distribution of benefits and burdens as a governance and environmental-justice challenge.
At the 2024 G7 summit, Pope Francis warned that AI could deepen injustice between advanced and developing nations and between dominant and marginalized social groups. Pope Leo XIV has gone further, calling on Aug. 21 for robust legal frameworks, independent oversight and international cooperation, while warning that rapid AI development could create new forms of technological dependence and widen the gap between poorer and wealthier nations.
The principle is straightforward:
Those who exercise the power to build and deploy AI should also assume responsibility for the consequences of that power.
Yet this does not mean AI is inherently harmful to the environment.
The technology presents a genuine paradox: its physical expansion may intensify environmental pressures, while its capabilities may also help humanity understand and manage those same pressures.
The fourth dimension of AI risk
The Earth is not entering the AI age from a position of environmental stability.
The World Meteorological Organization (WMO) in its State of the Global Climate study, issued in March, reports that the years 2015-2025 were the eleven warmest years on record. It also says that Earth’s energy imbalance reached its highest level in the 65-year observational record. The ocean has continued to absorb enormous amounts of excess heat, with the WMO estimating that it has absorbed energy equivalent to roughly 18 times annual human energy use each year over the past two decades.
The planetary-boundaries framework provides another way of understanding this context. Research published in Science Advances found that six of Earth’s nine planetary boundaries had already been transgressed.
This is not a precise prediction of when the planet will cross a single catastrophic threshold. Rather, it provides a scientific warning that several major Earth-system pressures have already moved beyond levels considered relatively safe by the researchers.
AI is entering this already stressed system, and the conventional risk discussion has focused largely on misuse, failure and loss of control. But that does not capture the full physical consequences of AI’s expansion.
There is another dimension that deserves serious attention: the physical and environmental consequences of scaling AI.
This fourth dimension is different. It does not require a malicious actor, a negligent developer, or a rogue AI system. It can emerge simply from deploying AI successfully at enormous scale. The question is not whether AI alone will break the planet. There is no sound basis for making such a claim. The more immediate question is whether rapidly expanding AI demand could push particular environmental and infrastructure systems beyond what they can sustainably absorb.
A data center may have little significance on its own at the global scale but it may create serious pressure on a local electricity grid, water supply, watershed or community. An AI data center does not consume “the planet’s” electricity or water in the abstract. It draws from a particular grid, water system or piece of land, while depending on supply chains that extract materials and manufacture increasingly sophisticated hardware elsewhere.
Multiple facilities can therefore create cumulative pressures that are invisible in isolation. AI’s environmental footprint should consequently be understood not simply as a question of how much carbon it produces, but whether rapidly expanding demands for energy, water, land and materials can be accommodated by environmental and infrastructure systems that are already under stress.
We do not yet know where all the critical thresholds lie. But that uncertainty is not a reason to ignore the issue. It is a reason to measure it before AI infrastructure becomes deeply embedded in places with limited resources and ecological capacity.
The challenge is to ensure that AI’s physical expansion does not outrun the capacity of the systems on which both AI and human societies depend.
The race no one wants to lose
Currently, the driving force behind AI expansion is not simply technological possibility or economic demand. It is the race itself and the fear of being left behind.
Governments and the private sector have powerful incentives to move quickly. The pressure to build larger models, secure computing capacity, expand data centers and capture new markets can make slowing down appear more dangerous than moving ahead.
Yet some of the most important constraints are physical. Energy and water may define the limits to AI growth. Alongside them is another constraint that is less tangible but potentially just as consequential: public acceptance.
How far and how quickly AI can expand may ultimately depend on how these pressures interact.
Technological expansion, however, does not necessarily proceed along a smooth or predictable path. A serious accident, infrastructure failure, environmental incident, security breach or other highly visible event could rapidly change public attitudes and regulatory responses.
How AI can help protect the planet
Yet the environmental dimension of AI is not only a story of risk. The same technology that increases demand for resources may also help humanity understand and manage those resources more intelligently.
AI may become one of humanity’s most powerful tools for addressing environmental problems. Its ability to process vast amounts of data and identify patterns can help us understand and manage systems too complex for conventional methods.
AI can also contribute to weather forecasting, wildfire detection, ecosystem monitoring, and agricultural efficiency. Its ability to process satellite imagery and other large datasets can help identify environmental changes that would be difficult to detect using conventional methods alone.
AI could even help assess its own environmental footprint before it is built.
Taken together, this creates a fundamental paradox: AI may increase pressure on the planet while also giving humanity new tools to manage those pressures.
A governance framework for the fourth dimension
Recognizing this fourth dimension requires more than voluntary efficiency pledges. Governments and companies should build four safeguards into the expansion of major AI infrastructure:
Mandatory resource disclosure. Operators should report projected and actual electricity, water, land, emissions and material demands using comparable methods.
Independent cumulative-impact review. Major projects should be assessed together with existing and planned facilities drawing on the same grids, watersheds, communities and supply chains.
Location-specific thresholds and mitigation. Approvals should reflect local scarcity, grid reliability, ecological capacity and community needs, with enforceable conservation and restoration requirements.
Public oversight and intervention authority. Independent institutions should be able to verify company claims, require corrective action and pause expansion when material risks exceed established limits.
These safeguards need not impose a single global ceiling on AI development. They would instead make growth answerable to the actual capacity of the places and systems on which it depends.
The challenge, therefore, is to expand our understanding of AI risks to include the planetary systems that make AI possible and to extend AI governance so that institutions are accountable to humanity as a whole.
We cannot build the future of AI without also protecting the physical world that makes that future possible.
The future of AI will ultimately be determined not only by how intelligent our machines become, but by whether our institutions can keep pace with the physical consequences of their expansion.
Pham Phan Long is a professional consulting engineer based in California and the founder and former Chair of the Viet Ecology Foundation.
Thank you very much for this article, perhaps the most comprehensive and profound discussion of the manifold risks posed by artificial intelligence that I have read. I hope AS may choose to make this article broadly available, as I would like to be able to share it with quite a number of friends and acquaintances.