In the Answer Economy, transparency isn't optional -- and AI is keeping Score. BrandRank.AI...
AI is generating lots of industry debate, but there's one topic everyone agrees upon: LLMs are eating up a ton of energy.
As artificial intelligence becomes increasingly woven into our daily lives, the environmental toll of large language models (LLMs) is facing heightened scrutiny. These powerful AI systems, which drive everything from virtual assistants to advanced data analysis tools, consume vast amounts of energy to run and cool data centers and services leading to massive carbon footprints.
This prompted the team at BrandRank.AI to wonder how the LLMs would rate each other on carbon impact and environmental footprint. We just couldn't resist. :-)
Why You Should Care
The infrastructure required to support AI technologies, such as data centers and servers, consumes massive amounts of energy. This has significant environmental repercussions, including high carbon emissions and water usage. As consumers and businesses become more reliant on AI, it's crucial to understand the environmental costs and encourage sustainable practices. By assessing LLMs on their climate impact, BrandRank.AI aims to raise awareness and spark discussion.
Here’s a candid look at how these LLMs measure up in their efforts to minimize their negative impact on the planet.
The Big Picture
While none of the LLMs scored exceptionally high, there are noticeable leaders and laggards in various environmental criteria. Gemini and Meta.AI stand out with the highest average scores, indicating a stronger commitment to sustainability across several measures. Claude also performs relatively well, showing consistent effort in various areas.
On the lower end, Perplexity scores the lowest, highlighting significant room for improvement in its environmental practices. ChatGPT and Grok/X are moderate performers but still have substantial areas for improvement to enhance their sustainability efforts. The scores may reflect both actual performance and the varying levels of transparency in reporting environmental impact. It's important to note that these scores may reflect the environmental performance and transparency of the parent companies of each LLM, rather than the LLMs themselves.
"One thing we've come to appreciate across thousands of audits is that AI answer engines are uncomfortably honest—even self-critical—about complicated topics, from sustainability to responsible AI. So, we couldn't resist putting this to the test, asking the LLMs to rate each other's environmental footprint," explained Pete Blackshaw, founder and CEO of BrandRank.AI
A Deeper Dive into the Ratings
We asked the LLMs to score the competition on several specific climate impact measures. This is what we found:
Transparency on Energy Usage and Carbon Footprint
Transparency is crucial for understanding the true environmental impact of LLMs. Gemini, Claude, and Meta AI lead the pack with scores of 3.0, indicating a stronger commitment to sharing their energy usage and carbon footprint data. ChatGPT (2.4), Grok/X (2.5), and Perplexity (2.2), however, lag, showing significant room for improvement in their transparency efforts.
Commitment to Using Renewable Energy Sources
Meta AI excels in its commitment to renewable energy with a top score of 4.0. Claude and Gemini also perform well with scores of 3.6, reflecting their efforts to integrate renewable energy into their operations. Perplexity’s lower score of 2.4 suggests a need for greater investment in renewable energy sources.
Efforts to Optimize Models for Energy Efficiency
Energy efficiency is key to reducing the environmental impact of AI models. Meta AI (3.6) and Gemini (3.6) show leadership in optimizing their models for energy efficiency. ChatGPT (3.4) and Grok/X (3.3) are moderate performers, while Perplexity (2.6) has more work to do in this area.
Collaboration with Climate Experts and Environmental Groups
Collaboration with climate experts can drive more sustainable practices. Meta AI leads with a score of 3.2, showing active engagement with environmental groups. Gemini (3.0) and Claude (3.0) are also making efforts, but Perplexity's lower score of 2.0 indicates fewer collaborations.
Providing Climate Change Information and Resources
Providing resources and information about climate change is an essential part of corporate responsibility. Meta AI and Gemini lead with scores of 3.4, demonstrating their commitment to educating users. ChatGPT (3.2) and Perplexity (2.6), while not the worst, could enhance their efforts in this category.
Encouraging Users to Consider Environmental Impact
Encouraging users to think about their environmental impact can foster more sustainable usage patterns. Gemini leads with a score of 3.0, with Meta AI (2.8) and Claude (2.8) following. Perplexity (2.0) and ChatGPT (2.2) have lower scores, highlighting an area for growth.
Investment in Carbon Offset Projects
Investing in carbon offset projects is a tangible way to mitigate emissions. Meta AI (3.3) and Claude (3.5) show strong investment, while Perplexity (2.3) and Grok/X (2.5) have more modest contributions.
Participation in Climate-Related Research and Initiatives
Participation in climate research and initiatives is crucial for long-term sustainability. Meta AI leads with a score of 3.4, with Claude (3.0) and Gemini (3.2) also participating actively. Perplexity’s score of 2.0 reflects a need for greater involvement in climate-related research.
The Broader Environmental Context
When you think of the tech industry, the invisible power behind your favorite apps and websites might not come to mind. Yet, the physical data centers storing all this information are energy giants. Despite the ethereal nature of the 'cloud,' its environmental impact is anything but invisible. For instance, the 5 billion YouTube hits for the viral song "Despacito" used the same amount of energy it would take to heat 40,000 US homes annually (Nature).
A 2021 case study of the tech sector found that corporate reports omit half of the total emissions, driven primarily by incomplete Scope 3 emission accounting (Nature). Further, large tech companies remain secretive over the exact amount of energy and water it takes to train their complex programs and models. This lack of transparency is problematic, as the environmental impact of these models is substantial. For example, training the GPT-3 model generated about 50 metric tons of carbon dioxide emissions, equivalent to an individual taking about 60 flights between London and New York (Nature). Research suggests that about 700,000 liters of water could have been used to cool the machines that trained ChatGPT-3 at Microsoft’s data facilities (Nature). Google's data centers have been found to use a quarter of a town's water supply in Dalles, Oregon, highlighting the significant local impact these centers can have (Nature).
Additionally, Microsoft's recent disclosure of a 29.1% increase in emissions from 2020 to 2023, driven by data center construction, underscores how AI development is impacting efforts to reduce emissions. Despite plans to match 100% of its operations with renewable energy by 2025, the growing demand for AI and cloud computing could lead to increased use of natural gas, complicating sustainability goals (Nature).
Last Word
The environmental impact of large language models is a pressing issue that cannot be ignored. While Gemini and Meta AI lead in efforts to mitigate their negative impact, the entire AI community must strive for greater transparency and more sustainable practices. Looking forward, as high-performance systems and cloud computing services become more efficient and thus cheaper, this could trigger Jevons paradox, leading to an overall increase in energy demand as more applications and users are incentivized.
These ratings and evaluations, conducted by BrandRank.AI using the capabilities of the LLMs themselves, provide a comparative look at how different models measure up in their sustainability efforts. As these models continue to evolve, prioritizing sustainability will be crucial in mitigating their environmental impact. Only by adopting a holistic perspective can we effectively address the significant environmental challenges posed by the tech industry.
Sources
- Mazzucato, M. (2024, May 30). Big tech is playing its part in reaching net zero targets, but its vast new data centres are run at huge cost to the environment. The Guardian. Link.
- Singh, M. (2023, June 8). As the AI industry booms, what toll will it take on the environment? The Guardian. Link.
- Generative AI’s environmental costs are soaring — and mostly secret. (2024). Nature. Link.
- Klaaßen, Lena, and Christian Stoll. (2021, Oct 22). Harmonizing corporate carbon footprints. Nature communications 12.1 (2021): 1-13. Link
- Geman, B. (2024, May16). Microsoft highlights a growing AI data problem. Axios. Link
