The AI Sustainability Trap: Why Researchers Say Society Must Ask Whether It Needs AI at All

Artificial intelligence is often presented as a technological ally in the fight against climate change, capable of optimizing electricity grids, reducing industrial waste, improving transport systems and helping governments manage scarce resources. But a new analysis argues that this optimistic story may be concealing a deeper problem: many systems promoted as “sustainable AI” are being developed within the same economic structures that drive environmental destruction and social inequality. Rather than asking only how artificial intelligence can be made safer, fairer or more energy-efficient, the authors say policymakers should confront a more fundamental question: is a particular AI system necessary in the first place?

Writing in AI & Society, Paul Schütze of Osnabrück University and Benedetta Brevini of the University of Sydney argue that current responsible AI governance often creates the appearance of action without imposing meaningful limits on the technology’s environmental and social costs. Their paper describes this phenomenon as “placebo change”: policies, voluntary commitments and ethical principles suggest that risks are being managed, while the underlying forces driving those risks remain intact. The authors propose replacing the dominant model of responsible AI with what they call “critical AI governance,” an approach that would judge AI against democratically defined social and ecological goals before allowing systems to be developed or deployed.

The controversy begins with the phrase “sustainable AI,” which contains two very different ideas. The first is AI for sustainability: using machine-learning systems to reduce emissions, improve energy efficiency, monitor ecosystems or optimize transport. The second is the sustainability of AI itself: accounting for the electricity, water, minerals, data infrastructure, labor and electronic waste required to build and operate these systems. The paper argues that public discussion has focused heavily on the first idea while often treating the second as a secondary technical problem. That imbalance can make an AI application appear environmentally beneficial even when its full lifecycle imposes substantial costs elsewhere.

Modern AI systems depend on a large physical and social infrastructure that is easy to overlook when the technology is described as software. Training and running large models requires data centers filled with specialized processors, cooling equipment and networking hardware. Those facilities consume electricity and, in many regions, significant quantities of water. The hardware depends on mined materials and global supply chains, while the production and maintenance of datasets may involve poorly paid or invisible labor. At the end of a device’s useful life, discarded servers and electronic components add to the growing stream of e-waste. From this perspective, the authors contend, AI’s environmental footprint is not simply a design flaw that engineers can eliminate through better code. It is connected to the economic system in which the technology is produced and expanded.

This distinction challenges a powerful assumption known as technological solutionism: the belief that complex political and social problems can be solved primarily by introducing a more advanced tool. An AI system can sort waste, detect equipment failures or forecast energy demand, but it cannot decide what a city should value. For example, a traffic-management system could be optimized for faster car journeys, safer cycling, cleaner air, lower fuel consumption, shorter commutes or improved public health. These objectives may conflict, and selecting among them requires political judgment and public agreement. An algorithm can calculate how to pursue a chosen goal, but it cannot provide the democratic decision about which goal is worth pursuing.

The paper describes the result as a form of displacement. An AI-based efficiency gain in one location may shift environmental burdens to another through energy use, mining, water consumption or waste disposal. Even when an application has a relatively small direct footprint, it may help preserve “business as usual” by creating confidence that technological progress is addressing the climate crisis. This is the authors’ central meaning of placebo change: an intervention appears transformative because it produces visible activity, reports, dashboards or efficiency improvements, but it does not alter the economic and political causes of unsustainability. In the most troubling version of this argument, the language of sustainability becomes a form of ethics washing, allowing institutions to advertise responsibility while continuing to expand resource-intensive systems.

The researchers direct particular attention to the role of soft law in AI governance. Soft law includes voluntary company commitments, industry standards, codes of conduct, best-practice frameworks and certification programs. Unlike hard law, it generally lacks one or more of three features: legally binding obligations, precise requirements and an independent authority empowered to interpret and enforce the rules. Such instruments can be useful when organizations genuinely want to cooperate, but they cannot reliably compel action when compliance conflicts with commercial interests. The authors argue that soft law can therefore create the social impression that AI is under control, reducing pressure for stronger regulation even when companies face no meaningful penalties for failing to meet environmental targets.

The European Union’s AI Act illustrates the tension, according to the analysis. The Act is a binding legislative instrument in many areas, but the paper argues that its treatment of AI’s environmental impact retreats into voluntary governance. The legislation refers to codes of conduct for measuring and minimizing environmental effects, while certain provisions require documentation of energy and computational resource use for general-purpose AI systems. Yet these requirements do not establish binding limits on emissions, water use, material consumption or total lifecycle impacts. Reporting what a system consumes is not the same as requiring companies to reduce that consumption. The authors describe this gap as a regulatory vacuum: environmental sustainability is acknowledged in legal language, but companies are not given enforceable minimum standards or sanctions for failing to lower their footprint.

Corporate sustainability reporting illustrates why disclosure alone may be insufficient. The paper cites a 2024 statement by a member of the European Parliament that Google would not provide a specific AI disclosure in a recent sustainability report, despite having done so two years earlier. The authors also point to Google’s 2025 environmental report, which recorded a 51 percent increase in total carbon dioxide emissions in 2024 compared with 2019. Microsoft, they note, reported emissions 23.4 percent higher than in 2020. These figures do not establish that AI alone caused the increases, and the paper does not present a new emissions dataset. They do, however, highlight the limits of governance that emphasizes documentation without requiring absolute reductions or making companies disclose comparable information about particular AI systems.

Critical AI governance would begin before the conventional regulatory process, the authors argue. Instead of assuming that AI development is inevitable and then attempting to mitigate its consequences, governments would first ask whether a proposed system advances human well-being, social justice and ecological stability. That could mean prohibiting applications judged unnecessary or harmful, rather than merely requiring them to publish risk assessments. It would also mean challenging the idea that innovation and economic growth should automatically take priority. The authors emphasize that this is not a call for rejecting every form of automation or withdrawing from technology. It is a demand to treat technological development as a political choice rather than as an unavoidable future to which society must adapt.

Their proposal rests on three broad principles. The first is to question the ideologies and economic imperatives that drive AI expansion, including the pressure to become an “AI leader” and the assumption that more computational power necessarily represents progress. The second is to prioritize social and political solutions over technological fixes, addressing the roots of climate change, inequality and resource consumption rather than delegating those problems to algorithms. The third is to invest in public digital infrastructure and independent research. Publicly controlled infrastructure could reduce dependence on a small number of powerful technology companies, while independent universities, civil-society groups and researchers could support scrutiny that is not shaped by corporate incentives.

This model would also shift who gets to decide which technologies are built. Communities affected by data centers, extraction, automated workplaces and algorithmic decisions would have a stronger role in defining acceptable uses. Democratic participation would not be limited to commenting on systems after companies had already designed them. It would include deciding whether particular systems are desirable, what social goals they should serve and whether the environmental cost is justified. The authors argue that questions about ownership, labor, supply chains, data control and the relationship between technology companies and states must be part of AI governance, because these conditions determine who benefits from AI and who absorbs its costs.

The paper is conceptual rather than experimental: no datasets were generated or analyzed, and it does not offer a complete legislative blueprint. Instead, Schütze and Brevini identify what they see as a structural weakness in responsible AI frameworks. If governance starts from the assumption that AI will exist and should be made more trustworthy, they argue, it may never reach the prior question of whether a system should be developed at all. Their warning is deliberately provocative because it targets the most familiar promise in the AI debate—that better design and better oversight will always be enough. As the technology spreads into energy, transport, medicine, government and everyday communication, the authors say, the crucial test of responsibility may not be whether AI can be made more efficient, but whether society is willing to say no when its costs outweigh its benefits.