Dear ChatGPT: Are you a climate solution? Or climate problem?
Some people think artificial intelligence is the best thing since sliced bread. Others say it’s the beginning of a science-fiction apocalypse. At COP28 – the U.N. Climate Change Conference – tech companies are saying AI is key to unlocking a more efficient future.
But what if the truth is less sensational than all that?
In this episode, how AI tools are helping and hurting efforts to curb climate change. From satellite-based flood maps to the growing energy cost of programs like ChatGPT, we’ll survey the use of artificial intelligence as a tool for climate action… and for climate distraction.
Featuring David Rolnick and Karen Hao
Our free newsletter is just as fun to read as this podcast is to listen to. Sign-up here.
SUPPORT
Outside/In is made possible with listener support. Click here to become a sustaining member of Outside/In.
Talk to us! Follow Outside/In on Instagram, or discuss episodes in our private listener group on Facebook.
LINKS
David Rolnick is one of the lead authors of this paper, called “Climate Change and AI: Recommendations for government action.”
Check out ChatNetZero, an AI climate chatbot that gives you references when it answers your questions.
A University of Washington researcher estimates the energy usage of ChatGPT (UW News)
After a Greenpeace report outlined how tech giants have worked with the fossil fuel industry, Google said it would no longer make AI tools to “facilitate upstream extraction” for oil and gas firms. (CNBC)
The Climate Summit Embraces A.I., With Reservations (New York Times)
COP28 president says there is ‘no science’ behind demands for phase-out of fossil fuels (The Guardian)
CREDITS
Host: Nate Hegyi
Reported, produced, and mixed by Taylor Quimby
Edited by Rebecca Lavoie, NHPR’s Director of On-Demand Audio
Our staff includes Justine Paradis, and Felix Poon
Music by Blue Dot Sessions
Our theme music is by Breakmaster Cylinder
Outside/In is a production of New Hampshire Public Radio
If you’ve got a question for the Outside/Inbox hotline, give us a call! We’re always looking for rabbit holes to dive down into. Leave us a voicemail at: 1-844-GO-OTTER (844-466-8837). Don’t forget to leave a number so we can call you back.
Audio Transcript
Note: Episodes of Outside/In are made as pieces of audio, and some context and nuance may be lost on the page. Transcripts are generated using a combination of speech recognition software and human transcribers, and may contain errors.
Taylor Quimby: Hey Nate Hegyi.
Nate Hegyi: Hey Taylor Quimby.
Taylor Quimby: So I want to take you back to 1990. It’s the World Cup semi-final. England vs. West Germany… and the score is tied.
Announcer: So it’ll go as a 1-1 draw this semi-final, and someone will lose on the penalty shootout.
Tens of millions of people across the UK are watching live on TV when England Mid-fielder Chris Waddle steps up to take a shot.
If he doesn’t score, West Germany will win the game.
Announcer: Here goes Waddle. Ooooh dear! It’s West Germany!
Taylor Quimby: He hits it over the net.
Nate Hegyi: Oh, Waddle.
Announcer: And England’s dreams of emulating their success of 1966, ends in penalty failure!
Guess what happens next in the UK?
Nate: I’m imagining people printing t-shirts with Chris Waddle’s face on it, and a big slash through it. “No more Waddle.”
Taylor Quimby: That probably happened, but I mean literally moments after this goal, electricity usage in the UK surged by 2,800 Megawatts…. As somewhere around 1.2 million Brits all angrily went into the kitchen and turned on an electric tea kettle.
[tea kettle SFX]
Nate: Leave it to the British to angrily make tea.
[mux]
Taylor Quimby: There’s actually a name for this, it’s called “TV Pickup”. I’ve also heard it referred to as the “Great British Kettle Surge.”
And even though it’s happening less now because so much TV is on-demand, it happened after the World Cup quarter-finals, in 2002.
[SFX of soccer goal]
During the Royal Wedding of Prince Williiam and Kate in 2011.
[SFX of royal horn-blowing]
Actually, it happens after every popular British soap opera..
David Rolnick: In the UK, people knew that when a soap opera ended, everyone would get up and make a cup of tea with an electric kettle…
This is David Rolnick, a computer scientist at McGill University in Canada.
David Rolnick: … and that would lead to a 10% spike in electricity consumption.
Taylor Quimby: Wow, really? That's so specific.
David Rolnick: It's extremely specific and quite… quite a large effect.
Eastenders clip: “I want the truth! Was it Mark or was it Steve?” “Steve?? Stever had nothing to do with it!” “Mark then? No!”
Something you should know about electricity, is that we only want to generate about as much as we actually need at any given time. Right?
If you don’t have enough, you have blackouts and power outages, but if you feed too much power into the grid, will wires sag and overheat, transformers can blow up…
Nate: Wow.
Taylor Quimby: Yeah.
David Rolnick: In order to run a power grid, you need to know how much power people are using and how much power is available.
David Rolnick: Both of those things are hard to predict.
David says there is a whole industry devoted to predicting supply and demand of electricity. It’s like a forecast - and forecasters have to factor in economics, they’ve got to think about business, the weather.
David Rolnick: As the sun went behind a cloud or the wind changed, all of your solar and wind power would change with it.
And you have to factor in human behavior - like the Great British Kettle Surge.
David Rolnick: Previously, people would have schedules of television programs to work out when people were going to be consuming power. They would be sending each other faxes with their guesses as to what would happen with spikes in electricity demand,
Nate Hegyi: I haven’t sent a fax in a very long time.
Taylor Quimby But actually serious business, because better energy forecasts help us cut back on fossil fuels… But also…
David Rolnick: People's lives are at stake if you make the wrong forecast. The electrical grid goes down and people die.
[mux intro]
Nate Hegyi: Sounds like a stressful job.
Taylor Quimby: And a hard one. All of which will either make you relieved or concerned to find out that this is just one of many problems related to energy and climate… that are increasingly being handed over to AI.
David Rolnick: So this is the kind of thing that is now being automated...
Taylor Quimby: Artificial intelligence.
David Rolnick.: …And the UK recently instituted an AI based algorithm for forecasting electricity demand, and that has led to a 40% improvement lower error in demand forecasts.
Nate Hegyi: Oh! AI for the win. I feel like we never hear stories about good AI.
Taylor Quimby: Yeah. Right. [pauses, and then they both start laughing] I mean, you know - keep listening. We’ll see where it goes.
News clip montage:
“Artificial intelligence, the new technology that’s already here. It can solve problems…”
“State school students can no longer use an AI program in the classroom, with fears it promotes cheating.”
“You’ve spoken out saying AI could manipulate or possibly figure out a way to kill humans?”
“You believe they can understand.” “Yes.”
“You believe they are intelligent?” “Yes.”
[mux swell/hit]
Nate Hegyi: Artificial intelligence is all over the news these days.
Taylor Quimby: To some, this explosive growth in technology promises to revolutionize the way we live and work.
Others see it as an existential threat - the literal robot apocalypse.
Nate Hegyi: But what if the threat of AI isn’t quite as sensational as that?
Taylor Quimby: Today on Outside/In, we’re asking a different question: Will AI help us fight climate change? Or will it accelerate the forces already pushing us over the brink?
Nate Hegyi: Producer Taylor Quimby here has been talking to some experts, and he’s got five big takeaways. Stay tuned.
Taylor Quimby: I think it’s fine. You know when you’re writing and you’re like, am I on four or five?
Nate Hegyi: He’s got a few big takeaways. Stay tuned.
[mux fades]
Taylor Quimby: So Nate… one thing that’s been driving me bananas lately is that people seem to be lumping a lot of different technologies under the same AI umbrella. Like we’re talking about it all the time - I don’t know that we know what we’re talking about.
Nate Hegyi: No it’s like, I think it’s like when people were talking about the internet in 1995.
David Rolnick: Yeah, the field of AI is older than ChatGPT. I'm sure everyone will be surprised to hear that
So again, this is David Rolnick - who aside from being a professor, is also co-founder of the non-profit, Climate Change AI.
David Rolnick: For a long time, people have thought of AI as something fancy and futuristic and just means any computer algorithm that makes predictions or labels things, And that can be very low tech.
Nate Hegyi: So we’ve had AI probably for a long time.
Taylor Quimby: Exactly! The chess app on your computer, spell-check programs like Grammarly, those use AI.
AI can sound like this:
Siri: I’m sorry, I didn’t get that.
Watson, on Jeapardy:
Alex Trebek: Watson?
Watson: What is clock.
Alex Trebek: Clock is correct.
Nate Hegyi: My Gmail gets organized into “focused” and “spam”. That’s probably AI, right?
Taylor Quimby: That is exactly AI. So when people are talking about AI these days, they might be talking about this broad category of useful computer technologies.
They might be talking about science fiction movies where computers actually gain consciousness….
Clip from The Matrix: A singular consciousness that gave way to an entire race of machines…
And now, a lot of times, they’re talking about this specific generation of programs like ChatGPT and Midjourney that we call generative AI:
News clip: AI software called Midjourney, it’s one of several image rendering programs used by millions…
David Rolnick: But these are different AI algorithms. Ai refers to. Things as diverse as mammal refers to both elephants and mice
[mux swell]
So here’s takeaway number one: When you’re talking about AI, make sure you know what it is you’re actually talking about. Because these differences really matter.
Nate:OK
[mux swells and fades]
We already talked about energy forecasts - I want to give another example of a sort of lower tech tool that uses AI to help solve climate problems.
David Rolnick: There is a branch of AI called Computer Vision, which focuses on processing images.
So you know when you take a picture, and your phone can tell whether it’s a picture of your Dad or a picture of Gilly?
Nate Hegyi: Yeah, uh-huh.
That’s made possible with computer vision.
So is the technology being used by self-driving cars to tell the difference between a person and a lamppost.
It’s also something that can be used to look through satellite data to create special flood maps.
David Rolnick: Now, technically, what's going on is that you give the computer a loose set of instructions and you give it a lot of data. That data comes with labels where somebody has gone through. An expert has said, this is a flood, this is not a flood, this is a forest. This is not a forest or this is an area of cropland which is healthy. This is an area of cropland that is suffering drought. And the algorithm picks out patterns that can help predict those labels, and then uses those same patterns on new data where there are no labels.
Nate Hegyi: Seems like the perfect technology for insurance companies, to figure out how like, are we going to give this house home insurance if it’s been built in this floodplane.
Taylor Quimby: And that’s in the longterm. On the like, very fast-acting, urgent basis, the UN has a satellite imagery center called Unosat.
David Rolnick: UnoSat has adopted algorithms that take imagery of areas that have been flooded, and find out exactly where disaster response teams should go in.
Nate Hegyi: So this is in real-time?
Taylor Quimby: That’s the very clutch part of AI!
Nate Hegyi: Wow.
David Rolnick: These are already deployed, or they're already being used, and they're already helping people around the Indian Ocean monsoon season.
So there are lots of examples of AI tools like this - very specific algorithms designed to do a very specific job really well, and really fast.
And they aren’t capable of starting a robot apocalypse anymore than your TI-84 calculator is capable of whipping up an egg souffle.
Nate Hegyi: By the way, I would watch a movie if it was TI-84 graphic calculators taking over the world. I’d watch that.
Taylor Quimby: But the point is, Tools like this can and are helping us lower factory emissions, monitor biodiversity and prevent extinctions, it can be used to research and build better batteries. This is BIG.
David Rolnick: People talked about big data for years before they had anything to do with the data. Now we have tools that allow us to use that data. For positive applications.
And that’s takeaway number two: AI is already a powerful tool in the fight against climate change.
But close on its heels is takeaway three… which is that these AI technologies being used to fight climate change, can be used to exacerbate it too.
David Rolnick: For every time you see a tech company promoting some new green use of its algorithms, it is also working on massively profitable algorithms to help accelerate oil and gas. And this is not small. It's estimated that it's going to lead to half $1 trillion in additional profit for the oil and gas industry just by 2025, with as much as a 5% production boost.
Nate: So explain this to me Taylor.
Taylor Quimby: Ok, so all three of the big cloud computing companies - Microsoft, Amazon, and Google all have contracts with major oil companies.
And if I can use specifically tailored AI tools to make emergency flood maps or make a factory more energy efficient… fossil fuel companies can also use their own specifically tailored AI tools to say, increase the efficiency of oil exploration - or create quote unquote smart pipelines that can move oil or gas faster, or with less maintenance or downtime. Does that make sense?
Nate Hegyi: Yeah, it’s not more efficient at like, keeping fossil fuels in the ground, or curbing greenhouse gas emissions, it’s becoming more efficient at getting more of it out and into our atmosphere.
Taylor Quimby: Even if it’s more efficient paper usage, it is saving them dollars and making them more profit. You know? Ok, Ok.
David Rolnick: Another way we see Ai being used that’s making climate change worse, is in online advertising. AI is behind almost all of the advertising that you see. And in some cases, that advertising is designed to increase consumption.
[mux]
Consumption of resources, consumption of energy. And we don't know what effect that is having, but you better bet it is huge.
Taylor Quimby: So what I'm getting to is that like, all of these cool AI tools? They’re making everything - it’s like that song - better, faster, stronger. And you know what I mean, and so yeah it might help in certain ways, but in other ways it’s just the opposite.
Coming up - we’re going to talk about the stuff that’s all over the news: generative AI - how much energy does Chat-GPT actually consume? And how else might it factor into climate change?
But first, we want to know how are you using AI. Are you a researcher who uses AI in your work, or a student using it for school? Are you like my Mom, and you’re using it to write newsletters for work? I did tell about ChatGPT and she got very excited.
Are you an AI optimist, or do you think the end is nigh?
Email us your thoughts at outside in at nhpr dot org, or join our private Facebook group and join the conversation there.
Be back in a minute.
…………….BREAK…………….
Nate Hegyi: This is Outside/In, I’m your host Nate Hegyi - here with producer Taylor Quimby.
Taylor Quimiby: Alright so what should I search here? Write me… five… bad… boomer jokes… about podcasts.
Nate Hegyi: Yeah, I like that.
Taylor Quimby: Well, it give you a little warning first. “Please keep in mind that these jokes are meant to be in good fun, and not to offend.”
Nate Hegyi: Thank you. Thanks Chat GPT.
Taylor Quimby: Okay Nate, why did the boomer refuse to start a podcast.
Nate Hegyi: Why?
Taylor Quimby: Because he thought a pod was something you plant in the backyard. [starts laughing]
Nate Hegyi: That…
Taylor Quimby: That’s mean.
Nate Hegyi: That’s mean!
Taylor Quimby: If you think that’s bad. Wait for number two. How does a boomer listen to podcasts?
Nate Hegyi: How?
Taylor Quimby: He tries to stick his cassette player into the USB player and wonders why it won’t fit.
Nate Hegyi: You know boomers aren’t technologically illiterate Chat GPT.
Taylor Quimby: WHen I typed that in I meant jokes by boomers, not jokes about boomers!
Nate Hegyi: Jeez.
[mux swell]
Taylor Quimby: Since last year, people have been freaking out about Chat-GPT. Right?
Nate Hegyi: Absolutely.
Taylor Quimby: FREAKING OUT.
Nate Hegyi: Myself included, if I’m being honest.
Taylor Quimby: It’s one of a new generation of AI-based chatbots that are built from what techno-nerds call Large Language Models or LLMS.
Nate Hegyi: LLMs.
Taylor Quimby: These are machine learning algorithms that are fed absolutely staggering amounts of data, and what comes out the other end is a chatbot that can sound intelligent, creative, and surprisingly human.
But one thing I’ve been wondering is… how much energy it takes to feed these programs all that data?
And I’m not the only one.
Karen Hao: My name is Karen Hao and I am a contributing writer to The Atlantic, covering artificial intelligence.
Taylor Quimby: when did you first in your reporting start thinking about or hearing about the energy cost of AI?
Karen Hao: Yeah. Um, so I was actually the a, I was the one that broke the first major story on this particular thing. It was a researcher named Emma Strobel who was doing a PhD at University of Massachusetts Amherst, who did a paper that was looking at the energy intensity of what were then considered large language models. Now they're considered like tiny potatoes compared to what are now considered large language models. But that was at the time of GPT two.
So GPT-2, which none of us played with, this was back in 2019, before I’d say the public at large was caring at all about this subject.
Karen Hao: So she had done this research and had just looked at how how much energy does it take to train one of these models? And the finding was for a particular model that used a particular type of training method that Google was known to commonly use, that training a single model would emit as much carbon as potentially the lifetime carbon emissions of five cars.
Taylor Quimby: Five cars…
Nate Hegyi. Meh - not that big.
Taylor Quimby: Right? The problem is, the past few years they’ve just been building these large language models bigger and bigger, feeding them more and more data.
People measure this by “parameters”, which are basically the number of variables fed into these AI algorithms.
Karen Hao: Gpt two was 1.5 billion parameters, which at the time was gargantuan. Gpt four is reportedly estimated to be around around 1.8 trillion parameters.
Nate: That’s an amazing amount of growth.
Taylor: So those five cars… that number is now much, much bigger. The real problem though is that AI companies aren’t being very transparent about how much energy it takes to train these huge AI models…
So we don’t know what the number is now.
Another way people like Karen are trying to figure it out, is by looking at the cost.
Karen Hao: The CEO of Anthropic, one of OpenAI's competitors, said at an event publicly that generally speaking, models are around $100 million train right now. And he sees in the next year it'll be around a billion, and in the next year it'll be around like 10 billion. So like. That's the only kind of general proxy that we have right now for truly understanding how energy intensive these are, because the cost of these models is purely the energy bill. It's the electricity costs for paying for these things to run all of the data centers that are crunching the numbers.
And Nate - all this time, we’ve just been talking about the energy it takes to MAKE the AI. Now, you have to factor in the energy from consumers who are actually using it.
Nate Hegyi: Thousands and thousands of us.
Taylor Quimby: In the case of Chat-GPT - which has been available to the public for about a year - we’re talking somewhere in the neighborhood of 180 million users.
Nate Hegyi: Oh wow. So we’re talking heavy energy use. Do we know how much?
Taylor Quimby: There are all sorts of smart folks trying to do some back of the napkin math on the internet to try and figure it out - and I’ve seen estimates that say Chat-GPT searches may use as much power as 30,000 housholds per day. Another one that said as much as 175,000 people per month.
Ultimately, all of this is very unreliable because OpenAI, and the other companies behind Chat-GPT - just aren’t being transparent and telling researchers what they’d need to be able to do the math.
But again, it’s not so much the number NOW that’s the problem - it’s what will happen if these huge AI tools continue to grow, and get used by more and more people.
But I’ve been desperately trying to get some numbers to help put this all into perspective, and what I can tell you is it’s definitely a lot more than a google search.
And the tech companies are planning on that being the case.
Karen Hao: There’re so many layers of environmental impacts that are sort of hidden, and not really visible to the average consumer that might just be searching things on Chat-GPT.
Karen told me that Microsoft has told investors that they want to spend 50 billion dollars building new data centers next year, to support the future of AI.
Karen Hao: The data center itself, you need the concrete to build the data center, which has a huge embodied carbon in it. And then they're running not just the electricity to power it, but you need electricity and water to cool the things because they overheat and they start melting if you don't cool them.
[00:25:25-00:26:07] And all of these companies have climate commitments, they have sustainability commitments and very ambitious ones where they've set targets to be net carbon like carbon neutral by 2030 or carbon negative by 2030, and water neutral or water positive by 2030. And it just seems to me that we are not seeing enough because there is no transparency and no accountability around what actually the numbers are for this kind of massive infrastructure drive to support generative AI. We don't actually have any way of auditing right now whether they will meet their climate commitments in 2030.
Taylor Quimby: And just to make something clear - the targeted types of AI tools we talked about earlier, like the UN flood maps and whatnot, these DO NOT have this kind of footprint at all. This is much a problem for these huge generative AI models.
Nate Hegyi: It just feels like this is one of those moments where it would be great if as a culture if we could just stop for a second and be like, do we need this? Like we were doing alright before generative AI came along to help us write better emails and search. Right?
Taylor Quimby: And especially because a lot of people… well the people around me, the people using Chat GPT are just playing with it. It’s like a meme generator.
Nate Hegyi: We’re just using it to create weird art.
So Takeaway number four: The bigger our AI models… and the more that consumers use them… the larger the likelihood is that AI will become a serious source of carbon emissions.
[reax and mux]
So there are a lot of movies and books where AI becomes so advanced, it threatens all of humanity.
And I’m glad people are talking about the risk of, you know Terminator level AI disasters.
But, Climate Change AI cofounder David Rolnick thinks we’re missing the bigger point. Or in this case, the smaller point.
He’s worried AI will contribute to an information ecosystem that has already bad for climate action.
Karen Hao: I do want to also note that I am frustrated by the way in which the discourse surrounding risks of AI has shifted to focus on large scale, hypothetical, and even existential risks.
The kinds of risks that are most relevant from AI are big, but they are prosaic. They are surveillance. They are biased. They are misinformation. They are the degradation of democratic structures or use of AI to perpetuate and exacerbate systemic inequities.
Nate: Preaching to the choir. I kept wanting to say Amen. Amen.
Taylor Quimby: Which gets me to my final takeaway. I said that AI is a powerful tool in the fight against climate change. It is.
But we have to be careful, especially in the coming years, that those tools aren’t used as a distraction by people that would prefer us to look the other way when it comes to the not-so-good-stuff.
David Rolnick: it can also be used deliberately to facilitate greenwashing. Look at how climate-conscious we are. We designed a new chat bot or look at how climate-conscious we are. We are doing this thing that isn't actually useful but is accelerated by AI.
Taylor Quimby: Can, can I ask, are there examples of that already? Like have you seen this?
David Rolnick: I do not want to single particular examples out, but yes, of course there are examples. You will probably see many of them at the UN Climate Change Conference.
Nate Hegyi: Point those fingers man! LIke, I want to hear who’s doing this!
Taylor Quimby: Spill the tea!
Nate Hegyi: Spill the tea, exactly.
David Rolnick: It is perhaps worth noting that the head of the UN Climate Change Conference this year is also the head of the State oil Company of the UAE. And so there is definitely some interest in focusing on climate solutions that don't involve actually changing business as usual with respect to the fossil fuel industry.
Just to recap here - the head of the COP28, the United Nations climate conference this year - He is also the head of the state-run oil company for the United Arab Emirates. And he’s in some hot water for saying he’s against phasing out fossil fuels - which seems contradictory. But you know what he’s bullish about?
Artificial intelligence.
David Rolnick: In these cases, new technologies can sound like a great idea, even if they don't actually provide the answers to the hard choices that society needs to make.
Taylor Quimby: It reminds me a little bit about the search for happiness. In that there is so much literature out there about how to be happy. And I’ve looked through it, and I’m always like, “oooooh I know this.” I know what the answers kind of are, it’s just hard to do them. It’s like ‘connect with other human beings’,’get good exercise’, ‘sleep well’.
Nate Hegyi: Exactly. We know what the answers are, we just keep trying.. We keep trying to build away our problems.
Taylor Quimby: We want the easy answer, and something like CHatGPT promises the easy answers of all. I cannot tell you how tempted I am to ask CHatGPT how to solve the climate change problem. But I know what it’s going to tell me.
Nate: …What is it going to tell you?
Taylor Quimby: It’s going to tell us to get off of fossil fuels. Or if it doesn’t that’s an even bigger problem. [imitates villain voice] “Yes, human, I think what you should do is extract more oil… “
Nate Hegyi: “Get off of fossil fuels, except when it comes to creating my server farms.”