Not much 

AI is significantly less pollutive compared to humans: https://www.nature.com/articles/s41598-024-54271-x

Published in Nature, which is peer reviewed and highly prestigious: https://en.m.wikipedia.org/wiki/Nature_%28journal

AI systems emit between 130 and 1500 times less CO2e per page of text compared to human writers, while AI illustration systems emit between 310 and 2900 times less CO2e per image than humans.

Training GPT-4 (the largest LLM ever made at 1.75 trillion parameters) requires approximately 1,750 MWh of energy, an equivalent to the annual consumption of approximately 160 average American homes: https://www.baeldung.com/cs/chatgpt-large-language-models-power-consumption

The average power bill in the US is about $1644 a year, so the total cost of the energy needed is about $263k. Not much for a full-sized company worth billions of dollars like OpenAI.

For reference, a single large power plant can generate about 2,000 megawatts, meaning it would only take 52.5 minutes worth of electricity from ONE power plant to train GPT 4: https://www.explainthatstuff.com/powerplants.html

The US uses about 2,300,000x that every year (4000 TeraWatts). That’s like spending an extra 0.038 SECONDS worth of energy, or about 1.15 frames in a 30 FPS video, for the country each day for ONLY ONE YEAR in exchange for creating a service used by hundreds of millions of people each month: https://www.statista.com/statistics/201794/us-electricity-consumption-since-1975/

ALL data centers in total (not just for AI) consumed about 149 GWh (17 GW * 365 days * 24 hours) in 2022 and is expected to grow to 306.6 GWh (35 GW * 365 days * 24 hours) by 2030: https://archive.ph/QL9LB

The US consumes 4,000,000 GWh each year: https://www.statista.com/statistics/201794/us-electricity-consumption-since-1975/

Image generators only use about 2.9 W of electricity per image, or 0.2 grams of CO2 per image: https://arxiv.org/pdf/2311.16863

For reference, a good gaming computer can use over 862 Watts per hour with a headroom of 688 Watts. Therefore, each image is about 12 seconds of gaming: https://www.pcgamer.com/how-much-power-does-my-pc-use/

One AI image generated creates the same amount of carbon emissions as about 7.7 tweets (at 0.026 grams of CO2 each, totaling 0.2 grams for both). There are 316 billion tweets each year and 486 million active users, an average of 650 tweets per account each year: https://envirotecmagazine.com/2022/12/08/tracking-the-ecological-cost-of-a-tweet/

With my hardware, the video card spikes to ~200W for about 7.5 seconds per image at my current settings. I can generate around 500 images/hour, so it costs 0.4 Watts each, which amounts to a couple cents of electricity or about 1.67 seconds of gaming with a high end computer.

https://www.nature.com/articles/d41586-024-00478-x

“ChatGPT is already consuming the energy of 33,000 homes” for 13.6 BILLION annual visits plus API usage (source: https://www.visualcapitalist.com/ranked-the-most-popular-ai-tools/). that's 442,000 visits per household, not even including API usage.

Models have also become more efficient and large scale projects like ChatGPT will be cheaper (For example, gpt 4o mini and LLAMA 3.1 70b are already better than gpt 4 and are only a fraction of its 1.75 trillion parameter size).

Everything consumes power and resources, including superfluous things like video games and social media. Why is AI not allowed to when other, less useful things can? 

In 2022, Twitter created 8,200 tons in CO2e emissions, the equivalent of 4,685 flights between Paris and New York. https://envirotecmagazine.com/2022/12/08/tracking-the-ecological-cost-of-a-tweet/

Meanwhile, GPT-3 (which has 175 billion parameters, almost 22x the size of significantly better models like LLAMA 3.1 8b) only took about 8 cars worth of emissions (502 tons of CO2e) to train from start to finish: https://truthout.org/articles/report-on-chatgpt-models-emissions-offers-rare-glimpse-of-ais-climate-impacts/ 

By the way, using it after it finished training costs HALF as much as it took to train it: https://assets.jpmprivatebank.com/content/dam/jpm-pb-aem/global/en/documents/eotm/a-severe-case-of-covidia-prognosis-for-an-ai-driven-us-equity-market.pdf

(Page 10)