- ChatGPT said its carbon footprint depended on the energy use of its computers and servers.
- The climate impact of AI is a newer area of research, but the effect is likely small.
Millions of people, like me, are now using ChatGPT, the artificial-intelligence tool that can write everything from layoff emails to dating profiles to detailed instructions on how to make cocaine.
I haven't asked it to create any of those things, though I did wonder whether ChatGPT could spit out an answer on its environmental toll.
Concerns about ChatGPT's carbon footprint, which is likely small at the moment, could rise as more people turn to it for day-to-day search. The chatbot had some 590 million visits in January alone, according to UBS analysts who described it as the fastest-growing consumer app in history.
The models underlying ChatGPT that train it how to understand and generate human language require heavy-duty processing power. And the data centers hosting its cloud networks must be kept cool. The energy demands for AI are likely to go up as more people ask it questions like, "How to negotiate a raise?" For now, a lot of mystery surrounds the carbon footprint of AI, including ChatGPT.
The murkiness isn't surprising. It's often difficult to total up a single human's carbon tab. I eat dairy and travel on airplanes occasionally. AI does neither of those — nor does it harbor the green guilt I sometimes feel for my actions. I asked the chatbot what its carbon footprint was, and its answer felt, well, human: It depends.
"As an AI language model developed by OpenAI, I do not have a physical body or personal actions that generate carbon emissions," ChatGPT wrote. "My carbon footprint would be limited to the energy consumption of the computers and servers used to run me and process requests. The exact carbon footprint would depend on factors such as the energy sources used to power the computers and the efficiency of the data centers."
ChatGPT added that OpenAI didn't disclose how many computers and servers were used to run and process requests. It's "likely a large-scale infrastructure consisting of multiple data centers spread across different regions," it wrote. The numbers might change based on demand.
And because ChatGPT's most recent data is from 2021, it's unlikely any answer it could give would reflect its recent surge in popularity.
Sasha Luccioni, a researcher at the AI startup HuggingFace, wasn't surprised that ChatGPT was vague — the environmental impact of AI is still an emerging area of research.
Last year, she coauthored a study that looked at the life cycle of emissions from HuggingFace's own language model, Bloom. The study, which wasn't peer-reviewed, measured emissions starting from the manufacturing of computer hardware to the energy used to train the model. Those emissions totaled nearly 51 metric tons of carbon dioxide — or about 60 one-way flights between London and New York, according to MIT Technology Review.
HuggingFace also tracked electricity use as Bloom fielded hundreds of thousands of queries from users during an 18-day period, and emissions totaled far less than 1 metric ton of carbon dioxide.
The emissions associated with Bloom were low compared with language models of similar size.
A 2019 University of Massachusetts Amherst study of popular language models found in the case of one model, for example, that its training could emit the equivalent of 284,000 metric tons of carbon dioxide, or five times the average gas-powered car over its lifetime.
If you take a step back, the entire global tech sector accounts for somewhere between 2% and 4% of greenhouse-gas emissions, and AI is only a slice of that field. The aviation industry similarly accounts for about 3% of global emissions.
That doesn't mean it should be ignored, Luccioni said.
"Most AI programs are on the low end of the emissions spectrum," she said. "A select few models seem to be disproportionately polluting, though, so it's important to understand where the sources of carbon are. Then you can at least start setting up guidelines."
ChatGPT also told me that OpenAI's goal was to handle an increasing number of requests in an "efficient" and "cost-effective" way, while reducing its environmental impact. Those strategies include optimizing energy use in its data centers, looking to source renewable energy for its operations, and purchasing carbon offsets.
OpenAI didn't return a request for comment on its environmental goals.
The entire exchange left me feeling like I had interviewed a politician because there were few direct answers and I still had a lot of questions, including what the carbon footprint of my query was.