AI Church Toolkit

The AI Church Toolkit looks at the cost of using AI, particularly environmental impact

>> Mercedes: Welcome to episode four of the AI Church Toolkit. Today we're taking a closer look at the cost of using AI, particularly the environmental footprint. While businesses everywhere are embracing generative AI to cut costs and improve services. As ah, faith leaders, we're called to consider all the costs, not just those that affect our bottom.

>> Peter: Unpack some of the claims and criticisms about AI's environmental impact. We'll also discuss a recent article that challenges some of these concerns that Mercedes brought to my attention, which we'll put in the show notes.

>> Mercedes: And as always, we'll bring it back to our faith and the baptismal covenant.

There are plenty of claims about artificial intelligence's environmental impact

>> Peter: So AI's environmental impact is a hot topic with plenty of claims out there about its carbon footprint. I hear about this all the time and it's probably the main complaint I hear from people here in my circles in California, in the Bay Area in particular, about using generative AI. Mercedes, you have this article that really lays out some of the arguments against AI from an environmental perspective. Can you lay them out for us here?

>> Mercedes: Sure. Uh, these are probably all sound bites that our listeners have already heard. Uh, there's four main ones. One each interaction with ChatGPT emits 10 times as much carbon as a Google search. Every time you prompt chat GPT, it supposedly uses 500ml of water. Like pouring a whole water bottle down the drain, ChatGPT, uh, uses as much energy as 20,000 households every day. And training an AI model has the same environmental impact as 200 plane flights from New York to San Francisco. These are some of the big talking points when it comes to AI's environmental impact. And in that recent article we mentioned, uh, by Andley Mason, he does a great job of breaking down these claims and providing some helpful context to, to consider. So if you want to dive deeper into this topic with some helpful visuals, I definitely recommend checking out the article.

>> Peter: Yeah, so these are the things that I've heard over and over as well. These four points you laid out. Uh, but they're not all true, right?

>> Mercedes: Not quite. Three out of four are actually true. Guess which one doesn't hold up.

>> Peter: Uh, is it the water one?

>> Mercedes: Uh, it's the water one.

>> Peter: I knew it. I remember seeing some tweet a year ago and reading the linked article and thinking that the numbers just didn't add up.

>> Mercedes: Exactly. That original report said that 500ml of water is used for every 20 to 50 chatgpt prompts. But somehow through the media and sharing that got twisted into people thinking it used that much water every single time. You, uh, prompted ChatGPT.

>> Peter: That's wild. And it makes such a huge difference whether that's the case for every prompt you input or whether it's for every 20 to 50 prompts. So what I'm hearing is, yes, water does get used, but it's more like 1/20 to 1/50 the amount that this sound bite suggests.

>> Mercedes: Right. And in the article, Maisley also points out that focusing only on the water use of ChatGPT leaves out the reality that all data centers, including those for email, cloud storage and stream streaming services, consume water for cooling as well. The focus should be on improving the efficiency of the industry and data centers in general, not just singling out AI's use.

>> Peter: Okay, yeah, that makes a lot of sense. So what about the other ones?

>> Mercedes: Okay, so estimating the exact numbers is tricky since we're dealing with proprietary information, but generally speaking, this claim seems Mostly true. ChatGPT likely does use about 10 times as much energy as a Google search, at least historically. But that's only part of the picture.

>> Peter: But aren't these models becoming more and more efficient all the time? I just recently saw a report that using ChatGPT 4.0, the current standard model, uses about one tenth the energy that its predecessors did, which means it's about the same as a Google search now.

>> Mercedes: Yes, that's a great point. Uh, things are evolving quickly and these models are getting more efficient all the time. For A while, the 10 times claim was a fair estimate, but as the technology improves, they may be getting much closer to parity with Google searches.

A single Google search uses around 0.3 watt hours of electricity

>> Peter: Shall we talk numbers for a moment so that people have something concrete to follow, like how much electricity are we actually talking about? And how does that compare to something like driving a car?

>> Mercedes: Sure. So, uh, let's break it down. We've put together, uh, an example using watt hours, which is a standard way to measure electricity use. A single Google search is estimated to use around 0.3 watt hours. Even older versions of ChatGPT, which might have used 10 times as much, were only around 3 watt hours per interaction.

>> Peter: But how much is a watt hour? And how concerned should I be about using 3 watt hours of electricity? Can we put that in relation to something like driving a car?

>> Mercedes: So here's a bit of context. One gallon of gasoline contains about 33.7 kilowatt hours of energy. Now, keep in mind, in doing the math, a kilowatt hour is 1,000 times more than a hour. It's not a, uh, perfect apples to apples comparison because of the different ways energy is used. But for simplicity's sake, that can give us a frame of reference.

>> Peter: Okay, so I know I use about 2 gallons of gas for a round trip commute to my church on Sundays, and the other four days a week, I'm usually saving that much by either commuting by public transit during the week or staying at home. So, uh, you're saying that by saving two gallons a day, four days a week, that's eight gallons a week. You know, multiply that by 33.7, I'm saving 269.6 kilowatt hours of energy each week.

>> Mercedes: Yes, the math checks out.

>> Peter: So that means that I would have to reduce the number of times I interact with ChatGPT. Let's, you know, run these numbers. Uh, divide that by 0.3 watt hours or 3 watt hours. Using the older numbers of the less efficient models, it would have to be 89,867 fewer interactions with ChatGPT just in order to save as much energy as I'm saving every week by driving less.

>> Mercedes: Yeah, when you look at it that way, it really kind of puts things in perspective and it doesn't seem like as much.

>> Peter: Yeah, so here's the thing. Even if we just take the outdated higher number of electricity usage for ChatGPT, it would take me almost 90,000 interactions to decrease what I use by the amount that I use in a week of reducing how much I, uh, drive. So for all the times that I've heard people share this sound bite, it seems like our efforts at reducing our environmental impact could be better directed elsewhere in other directions than focusing on how we're using AI.

>> Mercedes: Yeah, that also, uh, raises a good question. If ChatGPT does use 10 times as much energy as a Google search, when was the last time you actually worried about the impact of your Google search? What people don't say on top of that is that services like Netflix use more energy streaming for an hour than hundreds of AI prompts. But nobody is saying that you should stop binge watching your favorite shows. I think this kind of highlights how these concerns can sometimes feel a bit selective or disproportionate in a sound bite.

>> Peter: Right. And then it's also helpful to know that these numbers are going down as the technology becomes more efficient. And it seems to be that the energy use might actually be on par with a Google search nowadays.

The idea that ChatGPT uses 20,000 US households per day is misleading

>> Mercedes: Okay, so let's move on to the next claim. The idea that ChatGPT uses as much energy as 20,000 US households per day.

>> Peter: Is it true?

>> Mercedes: It is true, but it's also misleading. We need to make sure we're comparing apples to apples when we hear that ChatGPT as a whole uses as much energy as 20,000 US households per day. That is an aggregate measure, not an individual one. That's important because it's easy to hear a big, scary number like that and feel like it means we as individuals should opt out of using generative AI altogether.

>> Peter: I know that when I hear these big, scary aggregate numbers, it's easy to believe that it means that I, as an individual, should opt out of participating in something. But that doesn't seem like a helpful way of understanding how my individual use of generative AI could impact the world. That's. That's just guilt by association.

>> Mercedes: Yeah. Instead, we should ask, how does our individual use of generative AI compare to other actions that I can take on an individual level? Or on the flip side, how does the aggregate use of generative AI compare to other aggregate numbers when it comes to the environmental impact? Once we're asking those questions, then we can know that we're comparing apples to apples and kick out the oranges.

Netflix reportedly uses twice as much energy as ChatGPT

>> Peter: Okay, so we've already talked a bit about our individual use of ChatGPT and compared it to Google searches. That's the individual side. What about the aggregate numbers?

>> Mercedes: Okay, well, let's put that one in perspective with another aggregate number from Maisley's article, depending on whether you're looking at just the data center usage or considering the entire energy cost of video streaming. Netflix uses the equivalent energy of either 40,000 or. Or as much as 800,000 households.

>> Peter: So that's either twice as much or 40 times as much as ChatGPT in total. And I hear that. I gotta wonder, like, why don't we hear people worrying about the environmental impact of streaming video on Netflix?

>> Mercedes: Uh, I know. Video streaming has become so normalized in our daily lives that we rarely question its environmental impact and we have fun with it. And while video streaming does account for about 1.5% of all global electricity, it's not the primary concern when it comes to climate change. Yeah, I think it's a good reminder that new technologies often face more scrutiny, while established habits that we really like fly under the radar.

>> Peter: It seems like this sound bite is being used to scare people into not using something that could actually be quite useful and have an environmental impact that is far less than many other activities that we engage in on a daily basis. It's helpful to know that ChatGPT uses at most half as much as energy as Netflix does, and Perhaps as little as 1 40th.

Training an AI model has the same environmental impact as 200 plane flights

Okay, so what about the last soundbite?

>> Mercedes: Here it is again. Training an AI model has the same environmental impact as 200 plane flights from New York to San Francisco.

>> Peter: Okay, so plane flights have a huge carbon footprint. That sounds really bad, but how does my use of ChatGPT impact training costs? Doesn't training all happen before I use it? And really the numbers we should be looking at have to do with the inference rather than training.

>> Mercedes: Okay, so there are some important terms to understand right there. What do you mean by training and inference, Peter?

>> Peter: Right, so I can explain. Uh, training is pretty straightforward. This is the process of training an LLM to be generally useful. This is a one time process that occurs every time a new LLM is created. Once the LLM has been trained and begins to be used, however, it shifts over to inference, the computing cost of responding to its users prompts. Training is a one time thing. Inference, however, is ongoing.

>> Mercedes: Right, so we're back to apples and oranges in the comparison. The training process is one time an aggregate cost, whereas individual plane flights are repeated over and over. What really matters to us as everyday users is the environmental impact of inference, not training.

>> Peter: Yeah, so I only have control over the environmental impact of inference, not training, because when I use these tools, it's inference that happens. The training has already happened long ago at best. I guess it seems like one could argue that by boycotting ChatGPT, you might disincentivize the training runs of future models. But our own individual purchasing power is pretty small compared to the big numbers we're talking about. And because we're again talking about aggregate numbers, what seem like shockingly scary numbers, like the amount of energy it takes for 200 plane flights across the US is not actually as big as we might think.

>> Mercedes: And so, you know, as with all emerging technologies, these numbers and comparisons are going to continue to change over time. And as generative AI becomes more efficient, efficient, so might our understanding of its environmental impact. We will want to keep revisiting these questions with fresh eyes and considering other emerging, uh, measurements about the impact of AI. But instead of reacting instantly to sound bites, we also have an opportunity to be critical thinkers and thus engage with AI thoughtfully and as we've mentioned before, consider with discernment how to respond and use AI in our daily use.

>> Peter: Well, thank you Mercedes, for bringing this article to us. Uh, and for synthesizing these, these different viewpoints about the environmental impact of AI. I think that's, it's really helpful to have these more Helpful, useful comparisons to understand what it is we're actually talking about here. Yeah, so what about our baptismal promises? Uh, what comes to mind for you as we think about how this relates to the vows we make?

>> Mercedes: You know, so we often think about, uh, creation, care, and that being especially under persevering and resisting evil, um, and repenting. So, you know, we do have a responsibility to be mindful about AI's energy use and also to educate ourself about its true impact. But I do wonder about how we also might look at the other side and avoid misplaced guilt or even shaming over using what could also be helpful technology. What about you, Peter?

>> Peter: I think that's a helpful check to what we've just been discussing, because it's not about just like, have at and go wild using generative AI all the time. We want to understand the true cost of these tools so that we can discern how to approach them faithfully with, um, you know, environmental stewardship in mind. And that, you know, the. The other promise that comes to mind is to seek and serve Christ in all persons, loving your neighbor as yourself. Obviously there is. The environmental costs have an impact on other people, uh, other inhabitants of this planet of ours, other neighbors. We all know. And so there are many different ways we can do this. But one of the things that we haven't actually mentioned so explicitly is that there are many different models out there. ChatGPT4O is probably the most commonly used one currently. But if you have a ChatGPT subscription, you can choose different models to use. And, uh, a lot of them are more powerful and therefore more energy intensive. So it's helpful to stay informed on which models are the most energy efficient, which are more powerful, what they're good for, and use the ones that are best suited for your tasks. So environmental stewardship and caring for others means understanding the cost of our actions and discerning the right course of action accordingly. You know, uh, going back to the example of driving a car, we don't completely not drive cars at all. There is a cost there, and we discern when it's appropriate to do so.

Peter: AI's environmental impact may not be as extreme as some claims

>> Mercedes: All right, well, Peter, let's then summarize with our final takeaway. AI's environmental impact may not be as extreme as some claims, but we do recognize that we should still use it responsibly as we act as faithful stewards of the environment. Okay, so that finishes us up for today. So if you found this information helpful, uh, we'd love it if you would subscribe to the aichurch Toolkit and leave a, uh, review.

>> Peter: And as always, we've got some listener homework for you. So do some research about the environmental impact of various models, as well as learning about what each of the various models are best suited for. If you have a subscription to ChatGPT or some similar product, experiment with its different models, see how some models are more powerful than others, and learn to use the most environmentally conscious one that is capable of completing the task you're asking it to do.

>> Mercedes: Well, thank you for listening today. Don't forget to check out our show notes for the various articles we've, um, talked about.

>> Peter: And thanks for joining us for episode four of the AI Church Toolkit.

>> Mercedes: Remember, AI is tool, but our mission remains rooted in faith and community. See you next time.