The Cost of AI

The truth is, we have no idea how the AI and big database programs we use affect the environment. And Big Tech wants to keep it that way.

(Listen to the revised radio version here.)

Finding and identifying cool birds such as this Golden-winged Warbler has never been easier. Which is a Good Thing. Mostly.

[This post was edited on March 24, 2026 to reflect information that Peter Repetti added in an important comment. Thank you, Peter!]

When I was in college, I took a bunch of the “ologies”—ecology, mammalogy, herpetology, entomology and aquatic entomology, as well as two ornithology classes. Botany, too. I also took a few courses taught by my wonderful friend and hero, Bob Hinkle, focused on training us to be good all-around naturalists. Through all of this, I took many exams testing my identification skills. I did well on all of them and aced most.

But that was half a century ago. Since then, I’ve focused almost entirely on birds, and a lot of information about other forms of wildlife—info that I thought was securely locked into my brain—leaked out, never to return. When Deb Burns, my editor at Storey Publishers, asked me to write 100 Plants to Feed the Birds, my instant response was straight out of Star Trek: “Dammit, Jim, I’m a birdwatcher, not a horticulturist.”

If I’d had to rely on my memory to write that book, it would have been a disaster. Fortunately, I have access to the Internet and a whole library of my own reference books and materials, so I could find plenty of information about the native plants birds depend on for shelter, nesting material and substrate, seeds, fruits and, perhaps most importantly, native insects. Researching and writing that book made me appear very well informed, especially after the book was given the 2023 American Horticultural Society Book Award, which made me very proud.

But now when I lead bird walks, people think I can name all the plants we come across. They of course only ask about plants they don’t know, which raises the probability even higher that I won’t know them either. Fortunately, there’s someone in every group who, when any question about plant or animal identification comes up, whips out their cell phone, snaps a picture, and identifies it via iNaturalist, Google Lens, or another app. This whole process arouses enough interest to distract the whole group, ending my public display of ignorance.

Why don’t people ever ask me to identify chickadees?

Sometimes I take nice photos of butterflies, moths, dragonflies, or other insects, various plants, and other non-avian forms of wildlife. When I get home, I can look them up in Sparky Stensaas’s wonderful North Woods Naturalist Series of books if they’re from around here. If not, I use those apps.

Unfortunately, every single use of those apps, in the field or back at home, involves AI. I consider identifying plants and animals a wonderful use, helpful to so many people new to nature as well as those of us who’ve spent a lifetime studying and celebrating it, but in terms of energy and water consumption at huge data centers, it’s no different from other uses of AI, and no easier to find out exactly how much consumption is involved. Ironically, the water usage, carbon footprint, and pollution generated at data centers harm the very wildlife these apps are helping us identify.

It took me a while to get good at using eBird, but now I use it whenever I go birding. I generate my own checklist when I’m leading a bird walk or going out on my own, but when I’m on a guided birding tour, I prefer to let the guide generate and share the list—then I use “My eBird” to remove from my own checklist species I didn’t personally see, add species I may have seen on my own or the guide forgot to enter, and occasionally add media. eBird has an enormous database, but none of it involves AI, and using eBird in the field doesn’t involve data centers at all (as long as we downloaded the right regional “pack” ahead of time) right up until the moment we submit, share, or download a checklist. Then our data is added to all the other data on the eBird servers, i.e. “the cloud.”

Many uses of technology are incredibly valuable, and eBird is a prime example. Sure, it’s fun and useful for birders, but its value extends far, far beyond the hobby of listing, providing a huge, accurate body of information critical for bird conservation. Even more important, because birds are such incredibly useful indicators of environmental health, eBird’s value extends to other forms of wildlife and to whole habitats and ecosystems; the more gigantic eBird’s database grows, the more valuable it becomes. And again, using eBird does not involve AI. Nevertheless, we must assess with clear eyes the environmental costs of any huge database accessed by over a million active eBird users before we can possibly determine ways of minimizing its environmental costs.

Google Lens and iNaturalist, both AI intensive, may be helpful for people trying to identify plants and animals, but I’d argue that when it comes specifically to birds, Merlin is a much better choice. [Here’s where I started editing, because Peter Repetti pointed out that Merlin apparently does NOT use AI] ¹ I am also virtually certain that Merlin’s database is more accurate than Google Lens or iNaturalist. And Merlin generates data used by eBird, so it has positive value beyond accurate identifications.

What Merlin is most used for by most people now is real-time identification of bird sounds. In 2025, 11.6 million active users took advantage of Merlin’s Sound ID feature alone! I’ve caught it in a few mistakes, but if it were being graded, its accuracy would earn it a strong A.

I use Merlin a lot when I’m birding because my ears can no longer hear high frequencies. When I’m hoping to see a particular bird with a high-pitched song, such as a LeConte’s Sparrow or Golden-winged Warbler, Merlin tells me whether one is singing anywhere near.

I could never have seen, much less photographed, this LeConte’s Sparrow without both Merlin and Hear Birds Again.

Merlin can’t tell me where a bird is, so if I want to see or photograph it after Merlin confirms its presence, I put on my special binaural headphones, plug them into my iPhone, and turn on my Hear Birds Again app. (I wrote a lot about getting and using the app and headphones on this blog post.)

Hear Birds Again does not use the cloud or AI—it’s simply programmed to lower the frequency of high-pitched sounds coming in through tiny microphones in the headset, and makes them come through the ear pieces at a frequency I can hear. The mics are close to both ears so I can get the direction of the singing bird just as I can when I’m hearing anything without technological aids.

I don’t like wearing the headphones or using the app the whole time I’m birding—it drains my phone battery and somehow, because those high-frequency sounds are altered, my brain keeps trying to convert them to their real pitches, which tires me out. And if I don’t recognize an altered bird sound I’m hearing with “Hear Birds Again,” I have to turn the app off to use Merlin.

[Again, see footnote or Peter Repetti’s comment below suggesting that Merlin does NOT use AI.]

I can’t begin to guess how many people have told me how delighted they are when Merlin tells them about cool birds in their own backyard, or at their cabin, or when they’re traveling and don’t recognize all the new bird sounds they’re hearing. This is genuinely wonderful, and I sincerely celebrate Merlin’s popularity and how accessible it’s made birding. I also love that people can look up plants and non-avian animals with iNaturalist and Google Lens. But wouldn’t it be better for us and the natural world we love so much if we knew and could minimize these programs’ environmental costs?

A few years ago, I contributed to a Kickstarter campaign to support a project called Terra Listens that was trying to develop a small, affordable listening station for people to set up in our yards that could listen 24-7, identifying all the birds vocalizing within hearing range. Terra devices now also pick up wildlife radio tags. It took a lot of time for the small Terra team to develop and start building the units, but I got mine in 2024 and immediately set it up. Suddenly I could be working hard at my desk, glance over at my phone, and see that Trumpeter Swans were flying overhead at that very moment, or my first Brown Thrasher of the season was singing somewhere near. If I was traveling, all I had to do was turn on the app on my phone and I could see exactly what birds had been in my yard, or even listen to them live!

But this winter it hit me that Terra Listens is analyzing the sounds of cars, kids, dogs, our heat pump, my neighbors’ wind chimes, people talking, and everything else it can detect every moment of every day. I can’t imagine how energy intensive must it be to run Terra 24-7. I couldn’t justify it, so I turned my device off and brought it in for the duration. I may turn it back on during migration, especially if I find out some more specifics about just how valuable detections of wildlife radio tags are, but I just can’t see running it all the time anymore—not as more and more secretive billionaires and corporations are building more and more data centers, using more and more electricity and water and damaging more and more habitat to keep AI projects, good and bad, mushrooming.

AI, like every single thing any human has ever made, has costs and benefits. There are genuine, sometimes massive real-world benefits of some AI processes. But unless we reckon squarely with the environmental costs of this mushrooming technology, we cannot begin to work out ways of minimizing those costs and mitigating the damage. Attention must be paid.

1. I wanted to offer one clarification on Merlin, because I think it actually reflects well on Cornell’s design choices.

Merlin’s Sound ID runs entirely on your phone — no data center connection required during active use. Cornell’s own documentation states that “Sound ID runs on your device, without requiring a network connection,” and their FAQ confirms that recordings are not transmitted automatically. The download you do upfront during initial installation is what makes local inference possible. And again, I don’t think there are data available about what it took to train their model initially, but besides battery life on your phone, whether you run it for 60 seconds or 60 minutes, I don’t think there is a significant difference in water/energy use.