Does AI's future promise outweigh its future environmental risks?

The promise of how artificial intelligence (AI) can improve our lives is often framed in environmental terms, better climate modelling and adaptation, smarter energy grids, precision conservation, and pollution prevention. Yet AI systems depend on energy-, water-, and mineral-intensive infrastructures that are scaling rapidly.

A fair assessment must follow AI end-to-end software and hardware, direct and indirect effects, and the higher-order societal shifts they induce. A recent UNEP policy brief offers a comprehensive whole-of-lifecycle lens and a sober set of recommendations.

Where the benefits are real, but fragile

On the positive side of “promises”, AI can accelerate environmental science translation and help to support pro-environmental policy, for example, by summarising complex evidence for non-specialists. Yet the same generative affordances can also amplify misinformation, undermining collective action.

UNEP’s framing distinguishes such higher-order effects (behavioural, institutional, epistemic) from direct impacts (energy, water, minerals, e-waste), and indirect effects (e.g., "rebound effect" in autonomous mobility leading to more travel), as a useful way to guide policy prioritisation and research design.

Direct impacts: some uncomfortable facts

Large Language Model (LLM) complexity and volume of queries are creating more demand of energy and carbon. Early models were estimated to require approximately ten times the electricity of a traditional Google search to respond to user queries. Data centres required for AI expansion, already in the millions globally, are estimated to use 9% of U.S. electricity generation by 2030.

Ultra-pure water is needed to make the semiconductors for AI models and data centres, which require water directly for cooling and indirectly via electricity generation. Emerging analyses project 4.2–6.6 billion m³ of additional global water demand attributable to AI by 2027, which is more than half the UK’s 2023 annual water use. Locational choices also matter, as many chip and data-centre hubs sit in water-stressed regions, intensifying trade-offs.

AI’s hardware cycle relies on copper, rare earths, and other sought-after minerals with significant local impacts (habitat loss, contamination) and low end-of-life recovery. With only 22% of global e-waste currently recycled in environmentally sound ways, scaling AI without design-for-disassembly and circular flows risks compounding an already fast-growing waste stream.

Measurement – the missing backbone

UNEP’s report puts it bluntly: a comprehensive, transparent, and standardised accounting for AI’s footprint does not yet exist. Estimates vary, data access is limited, and methods for indirect and higher-order effects lack maturity.

In the near term, UNEP argues for focusing on the most tractable direct metrics, energy, GHGs, water, minerals, e-waste, paired with mandatory disclosure by AI providers and user-facing transparency.

Do the benefits outweigh the risks? It depends.

On current trajectories, with opaque accounting, rapidly escalating demand, and infrastructure sited in stressed watersheds, the environmental risks can easily overtake incremental efficiencies. But the balance can flip if the sector internalises environmental constraints as strict design requirements rather than afterthoughts.

Dr Ian Tennant, Lecturer in Biomedical Science, ARU Peterborough