The Curated Wild
Welcome to the brave new world of artificial intelligence for conservation
By David Biello
Artwork by Christopher David White
First, the flying drones scout the landscape recording the swell of hills, the temperature and humidity of soils, the location of streams and rivers via radar and GPS data. This information feeds back into computers, which use
machine learning—recording the information in photograph after photograph after photograph—to determine the best locations for planting a species, whether a mangrove in south Asia or a pine in western North America. Then the drones deploy to fire seed pods into the ground, planting more than 100,000 in a day.
Meanwhile, a swimming drone patrols the Great Barrier Reef, navigating by sonar and camera while scanning the area for destructive crown-of-thorns starfish. Using artificially intelligent software, the COTSbot spots the purple, thorny arms of the animal (even when wrapped around coral or partially hidden under it) and a pneumatic arm deploys to inject poison. The lethal drone can eliminate as many as 200 starfish in eight hours, helping preserve the reef from another echinoderm that can spawn millions of young.
And today, photo or audio files pour into a central location where computers that are set up as neural networks—a collection of programs training each other on a specific task, a crude imitation of the human brain—scan the collected works. The programs identify the whistle-sizzle of birds colliding with power lines, in hopes of developing better methods for avoiding such deadly strikes. Or they confirm sightings of an endangered plant or animal, such as a whale shark.
“We built an intelligent agent that replaces me,” explains Jason Holmberg, executive director and director of engineering for Wild Me, one of a slew of outfits employing artificial intelligence (perhaps better named ‟machine learningˮ) in pursuit of conservation goals. Holmberg spent two years data-mining YouTube vacation videos for encounters with whale sharks. He and his team then used that data to train the artificial intelligence program Wildbook to recognize whale shark sightings in order to enable scientists to gain a better understanding of populations, behavior, and other important information. “The users of our [machine learning] are overwhelmed with visual data,” he says. Already, neural networks allow a computer program such as Wildbook to be trained on one set of animal-picture data to learn the characteristics of another species and its ecosystem, rather than having to build a program anew for each and every task.
Welcome to the brave new world of AI for conservation, which offers a new possibility for this increasingly unnatural world: automating wildness. We’re used to thinking about AI in manufacturing and self-driving cars, but the technology is finding new uses in agriculture, health care—and now environmental conservation.
Can humans design technologies to curate ecosystems? And would these ecosystems be more or less wild than human-curated ones?
This automated future raises some very old questions, with some decidedly new twists—starting with, What is wild, anyway? ‟We tend to think of a species as ‘wild’ if it does not exhibit evidence that it is controlled or shaped by humans,” argues Laura J. Martin, an environmental historian at Williams College. But there are few, if any, organisms or ecosystems which fit that definition in the world today, whether the human-shaping is direct or indirect. In fact, the Anthropocene is a world of novel ecosystems, a mix of animals, plants, and other organisms living under rapidly changing environmental conditions. These novel assemblages of plants and animals sometimes thrive so much that it may be impossible for them to tolerate a return to some original condition. And even if the original condition could be restored, how could it endure the shift in climate that is upon us? The Arctic of today is different from the Arctic of yesteryear or the Arctic of the next century. The only constant is change.
Consider the Amazon rainforest, sometimes known as the lungs of the world. Will the Amazon as we know it today persist, or will it be cleared for farms and ranches? Will it be cleared anew, only to grow back—as the eastern forests of the U.S. have? In fact, the regrowth of the Amazon seems to have happened before—just a few hundred years ago, when Europeans reached the so-called New World and killed off their fellow humans largely through diseases such as smallpox. The overgrown urban outposts of the Amazon suggest that the rainforest has not been left to its own will for millennia.
In 1851, writer and naturalist Henry David Thoreau penned the now famous assertion, “In wildness is the preservation of the world.” That line has become an inspiration to the modern environmental movement. But Thoreau himself struggled to define the wild, ultimately coming to the conclusion that wildness is more of an attitude or idea than a reality. A century after Thoreau’s death, the poet Wendell Berry added an inventive corollary to the famous quote, one that seems befitting of the Anthropocene: “In human culture is the preservation of wildness.”
The wild, in other words, requires human imagination and human choice. But it also requires human forbearance and—maybe—technology unleashed by humans. Imagine, for example, a robot capable of learning from its environment through images and other data. Imagine people giving this robot the task of a conservationist, to preserve and protect a particular ecosystem or ecology. Imagine this robot freed to pursue its task as it deems best, removing signs of human intrusion and impact.
For example, this “Wilderdrone” might be asked to control the excess nitrogen flowing off farm fields and into a “wild” area. It might choose to create buffer strips of particular plants to consume the nitrogen and shield the core from too much fertilizer. Or it might reshape the land itself, either to better retain nitrogen or to promote ever faster flow through the protected area. It would learn from each action and implement better strategies based on the results, constantly improving its ability to remove, reduce, or replace human influences. Would the resulting creation—free from human interference and control, even human conception—be wild? “Can humans design technologies to curate ecosystems?” Martin asks. “And would these ecosystems be more, or less, wild than human-curated ones?”
After all, human conservationists already do all the above, but with the stigma of human interference in what ultimately is only a philosophical conception of the wild—whether that be a national park removed from historical context or a novel ecosystem growing on the margins of human activity.
The wild, in other words, requires human imagination and human choice. But it also requires human forbearance and (maybe) technology unleashed by humans.
There is a computer game called Universal Paperclips. The game asks players to adopt the role of an artificial intelligence programmed to make paper clips. At first, players simply make paper clips, sell them, invest in paper clip–making machines, and make ever more paper clips. Over time, however, the game incrementally and insidiously reveals that the AI’s single paper clip–making goal brings to an end the reign of humans, as it stops at nothing in its quest to make everything into a paper clip. The game—and the thought experiment behind it—illustrates the unintended consequences of too much autonomy for a capably intelligent but narrowly focused agent.
Similarly, an artificially intelligent and autonomous “Wilderdrone” tasked with conservation—or even with simply managing human interference—might rapidly come to the conclusion that eliminating humans is the key step to preserve an endangered species or stop light pollution.
Even short of that twin AI-environmental apocalypse, AI for conservation could privilege an endangered animal over the people with whom the animal co-exists, a repeat of the many people-versus-parks conflicts from around the globe that have turned settled peoples into conservation refugees or misidentified a local community as a group of poachers. And it is not clear whose vision of the world has been, or will be, programmed into any such software nor who gets to choose what the right state for a given ecosystem should be. “Ultimately, humans need to make decisions among each other about what world to live in,” Martin notes.
If there’s one thing wildness is not, it’s static. And yet a kind of artificial stasis—an ecosystem artificially protected from the vagaries of nature itself—is what AI can promise, and that may prove an all too seductive offer for nostalgic humans.
If we use AI to try to create and enforce a kind of artificial stasis—inhuman gardeners on the most massive scale and ecosystems insulated from the vagaries of nature—won’t we be disappointed by the results?
Alternatively, AI could become a way to further blend the wild and the tame. If any of the world’s woodlands, fields, and wetlands end up conserved, the first thing required will be keeping a better eye on them. And satellite monitoring or drones paired with artificially intelligent computers suggest a cheap, easy, and potentially accurate way to do so. What a drone and a computer program can do in minutes takes hours, days, or even months for a person or team exposed to the elements and the attentions of biting insects, among other hindrances, as they attempt to survey a plot or plant a new forest. Once ground-truthed, artificial intelligence promises an end to this kind of work. Drones can now sense air pollution, monitor animals, and stop poaching, among other conservation pursuits.
And AI could go even further. It “has the potential to profoundly shape other species,” Martin says. AI “is changing how we collect data on ecosystems, and therefore how we view and understand ecosystems.” AI, however, may remain forever constrained by its human trainers.
The strategies of conservation are, in the end, simple: management through technology and surveillance. “Imagine a world where a cloud of hundreds of thousands of photos from tens of thousands of researchers and citizen scientists could accurately predict population sizes and even individual animal life histories quickly and with high accuracy, iterating population estimates weekly and allowing wildlife-conservation authorities to constantly monitor and protect populations,” Holmberg says. “Now imagine those photos largely collected by remote drones, reversing human encroachment and instead promoting larger wildlife refuges in which wildlife populations exist in a very natural state while still being carefully monitored and protected.”
In such a near-future world, will we reserve half for nature and simply do nothing to the land and sea? Or will we simply outsource to our technology decisions between competing visions for how a place should be used, adding a veneer that may obscure the very human power struggles and politics? If so, the Anthropocene might rapidly become the Robotocene. This is a common story when it comes to people outsourcing decisions to AI: just think of how we wrestle with whether to program an autonomous vehicle to run over a grandmother to save a baby, or not. These will be hard decisions and must be made with wide participation and democratically. The decisions must also be subject to change in response to the needs of humans or animals, circumstances or goals, clear mistakes or clear victories.
“Conservation is a form of care,” says Martin. “Can we design AI that care? That love? That cherish other species?” For the moment, a love of the wild is found only in human minds.
David Biello has been covering energy and the environment for more than a decade and is the science curator for TED as well as a contributing editor for Scientific American. In this article, he draws from his latest book, The Unnatural World: The Race to Remake Civilization in Earth’s Newest Age.
What to Read Next
Using agricultural waste as fertilizer led to healthier soils, less invasives, and more tree canopy cover
Elephants inhabit complicated and changing landscapes—and they have a knack for covering themselves with mud. Machine learning solves the problem.