Despite alarming headlines about AI's environmental impact, the actual water and energy consumption of tools like ChatGPT is far smaller than most people believe. A single ChatGPT query consumes approximately 0.3 milliliters of drinkable water, according to OpenAI, while a basic text prompt requires just one kilojoule of energy, equivalent to running a microwave for a single second. The confusion stems from how researchers measure and report these figures, often conflating water "withdrawal" with actual water "consumption," inflating the apparent impact by a factor of 1,000 in some cases.

Why Are AI Water Consumption Estimates So Wildly Different?

The disconnect between alarming reports and actual numbers reveals a measurement problem rather than an environmental catastrophe. A 2024 Washington Post article claimed that writing an email with ChatGPT uses an entire bottle of water, while a bestselling book on AI overstated a data center's water usage by a factor of 1,000 before requiring a correction. These errors stem from confusion between water "withdrawal," which includes recycled water, and water "consumption," which measures only water that is evaporated. The University of California paper frequently cited in these discussions projects that by 2027, global AI water withdrawal will reach 4.2 to 6.6 billion cubic meters, but actual consumption will be only 0.38 to 0.6 billion cubic meters, roughly one-tenth of the withdrawal figure.

Even more important is the distinction between drinkable water and water used in electricity generation. Of the approximately 17 milliliters consumed per basic ChatGPT request to a US data center, only about 2 milliliters could actually be used for drinking. When researchers account for these nuances, the global AI water consumption in 2027 is projected to reach between 0.12 and 0.22 billion cubic meters of drinkable water, representing just 1.5 percent of Britain's drinking water consumption.

How Much Energy Do AI Tools Actually Require?

Energy consumption varies dramatically depending on the task. For text-based queries, both ChatGPT and Google Gemini require approximately one kilojoule of energy per prompt, equivalent to running a microwave for one second or consuming about 2 percent of an iPhone battery charge. Image generation is more demanding, requiring 2 to 4 kilojoules according to research from MIT, but remains modest compared to video generation. OpenAI's Sora, a text-to-video tool, consumed approximately 3,600 kilojoules to generate a single video, equivalent to an hour of continuous microwave operation, which contributed to the tool being shut down after a few months due to energy efficiency concerns.

When zoomed out to the global scale, the energy picture becomes even clearer. Data centers, including those supporting the entire internet infrastructure, consume merely 2 percent of global electricity, with approximately one-quarter of that dedicated to AI. This proportion is expected to rise to 3 percent by 2030, but the additional power needed for AI this decade remains dwarfed by demand from air-conditioning, electric vehicles, and industrial processes.

Steps to Understanding AI's Real Environmental Impact

• Distinguish between withdrawal and consumption: Water withdrawal includes recycled water and can be 10 times higher than actual consumption, which measures only evaporated water that cannot be reused.
• Separate drinkable from non-drinkable water: Most water used in AI data centers cools equipment or supports electricity generation, not human consumption, making direct comparisons to drinking water misleading.
• Compare AI energy use to other technologies: AI's 3 percent of global electricity use by 2030 is modest compared to air-conditioning, transportation, and food production, which collectively demand far more power.
• Consider the task type: Text queries consume minimal energy, while video generation can be orders of magnitude more intensive, so usage patterns matter significantly.

The real problem with AI's energy footprint is not the total amount consumed but rather where that consumption is concentrated. National power grids were designed a century ago to distribute coal power to homes and workplaces, not to handle the localized demand spikes created by massive data centers. In the United States, some households near data centers are already experiencing higher electricity bills due to this concentrated demand. The concern should focus on infrastructure planning and grid preparation rather than AI's absolute energy appetite.

Historical context illustrates how efficiency improves over time. In 1999, according to Forbes, it took one kilowatt-hour of energy to transmit 2 megabytes of data to purchase a book on Amazon. Today, that same amount of energy can generate a studio-quality video, demonstrating that technological progress typically brings efficiency gains alongside increased capability. As AI demand grows, industry will likely find ways to supply the necessary energy, and better technology will help use existing resources more efficiently.

The takeaway is not that AI has no environmental impact, but rather that the narrative of an impending "AI energy apocalypse" oversimplifies a complex infrastructure challenge. A single ChatGPT query consumes less water than a toilet flush, and the energy required for text-based AI tasks is negligible compared to other modern technologies. The real discussion should center on how societies choose to build and power the data center hubs that support AI, not on whether AI itself is inherently unsustainable.