Artificial intelligence (AI) and its environmental impacts have been increasingly under discussion since ChatGPT, and other similar AI models have become accessible to everyone. Today’s job markets are more and more focused on efficiency, which AI provides, but what is the real cost of this efficiency?
First of all, what is AI?
Artificial intelligence (AI) is the ability of a machine to use skills traditionally associated with human intelligence, such as reasoning, learning, designing or creating. Large language models are a type of AI that can generate and understand human language by processing large amounts of text data. ChatGPT is an example of an AI-powered large language model. As we know, ChatGPT is able to generate impressively well written text in different languages.
It is important to note that AI is not a new invention. The first uses of AI date back to the beginning of the 1900s, although those AI models were very simple compared to AI models used today. Some examples of the uses of AI in our everyday life today are popular search engines (e.g., Google), social media platforms (e.g., Instagram, Facebook and TikTok), direction apps (e.g., Google Maps and Apple Maps) and autocorrect and other text editing software. It is more or less impossible to not use AI at all in our everyday life. However, in this blog post we are focusing on AI-powered large language models which have the greatest effect on the environment.
Environmental impacts of AI
D-mat estimated the carbon footprint of the use of AI for the updated version of the Climate Puzzle. By reducing the use of AI (e.g. ChatGPT) for 20 enquiries per day a person can reduce their carbon footprint by 32 kg in a year which is about 5 grams per query.
However, there are also other environmental impacts such as excessive water usage, about 1-10 litres per 100 enquiries, to cool down the machines in data facilities. The water consumption depends on, for example, the length of the enquiries and answers. For one person this may not seem much but when it’s multiplied by millions of other users, the environmental impact of AI becomes huge. In general, it is challenging to find exact information on the environmental impacts of AI since companies are not giving this information easily to the public.
When we asked ChatGPT about its climate impacts and carbon footprint (prompt: what is your environmental impact and carbon footprint? Give specific numbers), it answered:
“🔋Energy Consumption:
• GPT-3: One query consumes approximately 0.3 Wh of electricity.
• GPT-4: One query is estimated to consume 0.6–1.5 Wh of electricity (2–5 times more than GPT-3).
• GPT-3 training: Used about 1,287 MWh of electricity, resulting in 502 tons of CO₂ emissions (equivalent to a 2-million-kilometer car journey).
• GPT-4 training: Energy consumption has not been published, but it is likely significantly higher than GPT-3.
☁️ Carbon Footprint (CO₂ Emissions):
• GPT-3: 0.15–0.3 g CO₂ per query (depending on the energy source).
• GPT-4: Estimated 0.5–1.5 g CO₂ per query.
• Comparison:
o Google Search: ~0.2 g CO₂ per query
o 10-minute YouTube video (1080p): ~30 g CO₂
o One email (without attachment): ~4 g CO₂
o 1-hour Google Meet video call (HD): ~1 kg CO₂
💧 Water Consumption:
• Data center cooling: GPT-4 usage consumes approximately 500 ml of water per 5–50 queries.
• GPT-3 training: Consumed about 700,000 liters of water.
• GPT-4 training: Water consumption has not been published, but it is likely higher than GPT-3.”
In the answer ChatGPT tells us that its carbon footprint is significantly smaller than in the calculations done by D-mat. However, the background assumptions and other factors (e.g. whether the training is included in the carbon footprint or just the individual enquiries or simply OpenAI’s desire to show the numbers lower to increase the use of ChatGPT) can explain these differences.
ChatGPT and other AI models can, at their best, make work much more efficient. For example, they save time in text generation, translation and coding, which can, in addition to saving time, also improve productivity and decision-making. However, efficiency also comes at a significant environmental cost.
D-mat
References:
Li P. et al (2023). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models (pdf)
Wong, V. (2024). Gen AI’s Environmental Ledger: A Closer Look at the Carbon Footprint of ChatGPT. Pictochart. https://piktochart.com/blog/carbon-footprint-of-chatgpt/
Picture: ChatGPT