Chat GPT and carbon footprints: how much CO2 does your query produce?

“Each [ChatGPT] query emits around 4.32 grams of CO2“, according to a “detailed calculation” from Smartly.AI, an AI agent for customer service with a blog discussing the the quantity of CO2 produced per usage of ChatGPT. The breakdown of this environmental impact can be boiled down to the two phases required in the developing of generative AIs.

The training phase of large AI models is very resource intensive. This comes at no surprise, with LLMs and generative AI’s requiring huge amounts of electricity to run, and massive amounts of water to keep systems cool. To train these models, there needs to be the processing of vast datasets over several weeks or months, leading to this large consumption of water and electricity.

Then, there’s the inference phase. This is the process of when we actually use ChatGPT and ask it our queries. This usage of ChatGPT requires “continuous use of computational resources”, and industry estimates suggest that “each generative AI query uses four to five times more energy than a standard search engine query”.

Smartly.AI’s blog then provides some further numbers to put the environmental impact of ChatGPT even more into perspective: “15 ChatGPT queries equate to the CO2 emissions of watching one hour of video streaming; 139 queries are roughly equivalent to the emissions from one load of laundry washed; 92,593 queries would match the carbon footprint of a round-trip flight from San Fransisco to Seattle”.

Now we look at the wider picture of how ChatGPT usage is affecting our planet. According to the Master of Code blog, ChatGPT handles “10 million queries daily”. This means that with some simple calculations, the total carbon footprint of the usage of ChatGPT in one single day is 43,200,000 grams Of CO2. That‘s 43,200 kilo grams of C02. Or 43.2 tonnes of CO2.

If a flight from London to New York has the total carbon emissions of 0.59 tonnes per passenger, that means the total daily carbon emissions of the usage of ChatGPT is the equivalent per passenger of 73.2 flights between London to New York daily.

So here lies the question: how can we justify the tremendous widespread use of generative AIs when they are producing such a large environmental footprints?

It is undeniable that this use of AI has brought about significant changes to work places, the education sector, and other industries. An MIT study demonstrates how generative AIs can boost the productivity of highly skilled workers, arguing that it can “improve a worker’s performance by nearly 40%” in comparison to workers who do not use AI.

Furthermore, a study done by Scientific Reports suggested that the carbon emissions of writing and illustrating are less for AI than they are for humans. Their findings revealed that “AI systems emit between 130 and 1500 times less CO2 per page of text generated compared to human writers”, with similar numbers found between AI illustration systems and human illustrators. Therefore, while AI is not yet able to substitute for all human tasks in these sectors, the finding concludes by saying that “the use of AI holds the potential carry out several major activities at a much lower emission levels than can humans”.

After all this, we have reached a dilemma of sorts. Usage of AI has the ability to transform industries, increase productivity, and perhaps in certain sectors have a reduced carbon footprint than their human counterparts. However, instead of targeting all those who are using generative AIs, should we just be limiting usage for when people use ChatGPT or other LLMs for ‘menial queries’, as these are contriubting to the carbon footpring

Nonetheless, the environmental impact of AI training and inference is staggering. And to improve and progress with AI, there is only going to need to be more testing and training. In order to ensure companies give customers the most updated generative AI systems, they will have to continue with this testing - as discussed in one of my previous articles, ensuring that companies are giving customers the most accurate AI systems they can is something they should be obligated to do.

Therefore, in this debate between generative AI’s carbon footprint and the need for developed AI, there is a calling for sustainable AI development. These should be strategies such as implementing more efficient AI modelling architecture, and ensuring data centres are powered by renewable energy sources.

To end on a positive note, some of the smartest minds in the world right now at companies such as Google, Microsoft and OpenAI are now tackling questions of how AI can assist in energy production, green power and renewables. These are very promising in general, and particularly in the sustianble development of AI too.

Articles referenced:

Smartly.AI: What is the CO2 emission per ChatGPT query?

Master of Code: 50+ Eye-Opening ChatGPT Statistics: Tracing the Roots of Generative AI to Its Global Dominance

Travel Navigator: Carbon emissions from London to New York

MIT Sloan: How generative AI can boost highly skilled workers’ productivity

Scientific Reports: The carbon emissions of writing and illustrating are lower for AI than for humans