The global conversation about artificial intelligence took a sharp turn this week at the AI Impact Summit in New Delhi, where a new analysis accused the world’s largest tech companies of obscuring the true environmental cost of the technology behind a fog of climate promises.
The report, commissioned by the non-profit groups Beyond Fossil Fuels and Climate Action Against Disinformation, scrutinised 154 corporate and institutional statements. It concluded that the industry is engaging in “diversionary” tactics by conflating two very different forms of AI: the lean, traditional machine learning used for forecasting and optimisation, and the explosively popular, energy-hungry generative AI that powers chatbots and image creators.
According to the analysis, not a single claim demonstrated that consumer-facing tools like Google’s Gemini or Microsoft’s Copilot have led to a “material, verifiable, and substantial” reduction in planet-heating emissions. Report author and energy analyst Ketan Joshi likened the strategy to fossil fuel companies overstating minor green investments. “These technologies only avoid a minuscule fraction of emissions relative to the massive emissions of their core business,” Joshi said. “Big tech took that approach and upgraded and expanded it.”
The Evidence Gap and a Pervasive Claim
The report found the evidence base for most green claims to be strikingly weak. Of the statements studied, only 26% cited published academic research, while 36% provided no cited evidence at all. Many of the scrutinised claims originated from an International Energy Agency (IEA) report—which had chapters on traditional AI’s potential—and from corporate sustainability reports by Google and Microsoft.
One prominent example of a pervasive yet weakly evidenced claim, identified in the report, is that AI could help mitigate 5-10% of global greenhouse gas emissions by 2030. Google repeated this figure as recently as April last year, citing a report it commissioned from the Boston Consulting Group (BCG). That BCG report, however, ultimately traced the statistic back to a 2021 blog post which attributed it to the firm’s “experience with clients,” a methodology critics have called insufficient.
Google, responding to the new analysis, stated that its estimated emissions reductions are “based on a robust substantiation process grounded in the best available science,” and that it has transparently shared its principles and methodology. Microsoft declined to comment. The IEA, which has separately predicted AI could cut global emissions by up to 5% by 2035, did not respond to requests for comment on the report.
The Soaring Cost of the AI Boom
The critique arrives amid mounting concern over the physical footprint of the AI revolution, which is largely driven by generative models. Data centres, the engine rooms of the digital economy, currently consume an estimated 1-2% of the world’s electricity, a figure projected to double by 2030. In the United States, data centres’ share of electricity demand is forecast to more than double to 8.6% by 2035, with AI workloads being a primary driver. The IEA predicts they will account for at least 20% of the growth in electricity demand in industrialised nations by 2030.
While a simple text query to a model like ChatGPT may use roughly the energy required to power a lightbulb for a minute—with recent estimates from researchers and companies ranging from 0.24 to 0.42 watt-hours per request—the cumulative scale is vast. A single training run for a model like GPT-4 can consume as much electricity as 40,000 US households use in a year. Goldman Sachs estimates AI could increase data centre power demand by 160% by 2030.
This growth has a significant secondary toll: water for cooling. A single data centre proposed by Microsoft near Phoenix was reported to be projected to consume 56 million gallons of fresh water annually. Furthermore, despite efficiency gains and corporate pledges to use renewable energy, the sheer speed of expansion is increasing reliance on fossil fuels; gas-fired power generation for data centres is expected to more than double by 2035.
“When we talk about AI that’s relatively bad for the planet, it’s mostly generative AI and large language models,” said Sasha Luccioni, AI and Climate Lead at the open-source platform Hugging Face, who was not involved in the report but said it added crucial nuance. “When we talk about AI that’s ‘good’ for the planet, it’s often predictive models, extractive models, or old-school AI models.”
Nuance in the ‘AI for Good’ Narrative
The report and experts do not dismiss the potential of traditional AI for specific climate applications. Examples cited include Google Maps’ fuel-efficient routing, which the company says has prevented millions of metric tons of CO2 emissions, and AI tools used for flood prediction, contrail reduction, and optimising energy use in buildings. Luccioni highlighted AI applications in conservation, such as detecting illegal logging or tracking endangered species, and in accelerating research for greener batteries and solar panels.
However, the central argument of the new analysis is that these beneficial, niche applications are being misleadingly bundled with the far more pervasive and resource-intensive generative AI tools to create an overall impression of environmental benevolence. The report argues this “muddling” of AI types allows the sector to present climate solutions and carbon pollution as a package deal.
Joshi warned that the discourse must be “brought back to reality.” “The false coupling of a big problem and a small solution serves as a distraction from the very preventable harms being done through unrestricted datacentre expansion,” he said. The release of his findings at the Delhi summit, an event positioned to focus on “People, Planet, and Progress,” underscores the growing global pressure to scrutinise not just AI’s risks, but the veracity of its promised benefits.
