Green AI? What the True Cost of AI Means for Sustainability

Will artificial intelligence advance sustainability or hinder it? Does AI consume a lot of water and energy, increase surveillance of people, and is it trained under exploitative conditions? Can there be such a thing as ethical AI?

---

Artificial intelligence (AI) is the game-changer of our time. It writes texts, generates images, and optimizes complex industrial processes. But while we—fascinated by its speed—stare at our chat windows, a massive machine is churning away in the background, its insatiable appetite for resources seemingly incompatible with climate goals and human rights. It’s time to take a look behind the digital facade: What does a prompt really cost us? And to what extent can artificial intelligence (AI) contribute to greater sustainability?

Water Consumption by AI

People rarely associate technology with water. But hardware needs to be cooled, usually in what are known as evaporative cooling rooms.

• Training: Training GPT-3 alone consumed 5.4 million liters of water.
• Operation: Google’s data center in Iowa consumed a staggering 3.785 billion liters in 2024.
• Projection: For U.S. data centers, consumption is estimated to reach 257.38 billion liters by 2028.

Every single query contributes to this. Consumption is estimated at 5 to 50 milliliters per chat. That may not sound like much at first, but when multiplied by the 2.5 billion daily queries at OpenAI alone, that drop turns into a tsunami. Added to this is “virtual water”: The production of a single modern microchip, for example, requires up to 100 liters.

Energy Consumption and Carbon Footprint of AI Data Centers

The energy consumption of AI data centers is enormous. A single data center often requires several hundred megawatts of connected load. That’s as much as tens of thousands of households.

• Carbon footprint: AI data centers already account for 2.5 to 3.7% of global greenhouse gas emissions. This puts them ahead of the entire aviation industry (2%).
• Future projection: By 2030, AI energy demand will rise to approximately 945 terawatt-hours (TWh). This is roughly equivalent to the current electricity consumption of all of Japan.

A comparison in everyday life: A conventional Google search requires about 0.3 watt-hours. An AI query uses nearly ten times that amount (2.9 Wh).

Energy Sources for AI: Nuclear Power vs. Renewable Energy

To satisfy the AI giants’ hunger for baseload electricity, a surprising turnaround is taking place:

• Google is investing in mini nuclear power plants (molten salt reactors).
• Microsoft is restarting the “Three Mile Island” reactor in the U.S. state of Pennsylvania—the very site that once became infamous for a near-meltdown.

But is this sustainable? While the Intergovernmental Panel on Climate Change (IPCC) estimates the median CO₂ emissions from nuclear power at just 12g per kWh (similar to wind power at 10g), the problem of nuclear waste disposal remains unresolved worldwide, and the new form of energy generation—nuclear fusion—is expected to remain uneconomical well into the 2030s. To date, there is still no such power plant that generates more energy than it consumes.

The AI Climate Hoax

Major tech companies engage in greenwashing to obscure the true costs of AI.

They often claim that AI can combat climate change and keep water and energy consumption low. A study by Algorithm Watch exposes these lies:

• Many claims about positive climate impacts refer to “conventional” AI, which requires comparatively small amounts of data for tasks such as statistical predictions or image recognition.
• Generative AI, in particular, has the computational capacity to combat climate change. However, this type of AI requires vast amounts of data and consumes massive amounts of energy and water.
• The positive environmental impacts of “conventional” AI are only weakly substantiated. Only 26% of the claims about it are based on scientific publications. 36% are unsubstantiated.

The social downside: exploitation and human rights violations

Sustainability is more than just ecology. The seemingly magical AI owes its training to precarious “click work,” primarily in the Global South.

• Trauma included: Workers review violent and traumatic content for less than $2 an hour to train filters—often without psychological support. This can lead to depression, PTSD, and other mental health conditions.
• (Mass) surveillance: People are less likely to exercise their fundamental democratic rights—such as freedom of speech—in public, monitored spaces.
• Restriction of individual autonomy: People feel more uncomfortable visiting places associated with a particular religion or sexual orientation.

How fair is artificial intelligence? Is there such a thing as ethical AI?

When AI systems make decisions based on existing gender inequality—as is the case, for example, when hiring new employees in a male-dominated industry—these inequalities are reproduced.

AI systems reflect existing power dynamics and perpetuate them. Combined with AI surveillance, this can lead to racist conclusions being drawn about a person’s characteristics based on their phenotype.

To avoid discrimination, it is therefore recommended to actively incorporate gender-sensitive considerations and diversity into the programming process—for example, by having a diverse team. To this end, it also makes sense to generally promote women in the digital industry.

From Problem to Solution: How AI Can Become Sustainable

Despite its downsides, AI offers enormous opportunities for ecological and social transformation. Here’s what it takes:

• Assess necessity: Prioritize human or analog solutions and use AI only if it offers added value over the conventional approach.
• Sufficiency & Precision: AI models that function with small amounts of data, minimal training and pattern recognition time, and low hardware requirements.
• Training: To rule out discrimination, developers take specific care to avoid biases when training the AI.
• Technical Efficiency: The hardware and software should deliver high performance with relatively low energy consumption.
• Energy Consumption: Minimize negative environmental impacts by relying on renewable energy, circularity, and the recovery of waste heat.
• “Build-Measure-Learn Loop”: Continuous monitoring of the AI’s activities and improvement of energy consumption in a repetitive process.
• Disclosure & Burden of Proof: Disclosure of all data in an understandable form. The positive impacts of the application must outweigh the negative environmental consequences.

Conclusion: AI Needs Guidelines from Policymakers and Businesses

To prevent AI from becoming a climate killer, a water waster, an exploiter, a racist, or a surveillance tool, it must be “Sustainable by Design.”

For policymakers, this means:

• Political Regulation: The EU AI Act and international treaties such as the Council of Europe’s Convention on Artificial Intelligence (CAI), in line with legal standards for human rights, democracy, and the rule of law, must establish legally binding minimum social and environmental standards.
• Energy Assessment: The use of AI must be mandated to align with regional energy capacities and 100% renewable energy.

For businesses, this means:

• Transparency: Companies must disclose the resource consumption of their models.
• Efficiency over scale: Prioritize small, specialized models over gigantic “jack-of-all-trades” models.
• Use of alternatives: Companies (and individuals) should use eco-friendly alternatives like Ecosia for simple queries and critically assess the necessity of AI use in advance.

Cover photo: Purchased from Adobe Stock, edited by Team Tina Teucher

---

Tina Teucher is an expert in sustainable transformation and has been working with companies on future-proof solutions since 2008. One of her areas of focus is Green AI. She delivers keynote speeches on topics including sustainability and artificial intelligence (AI).

---