In the wetlands of Senegal, researcher Alexandre Delplanque uses a drone to monitor waterbird populations, including species like pelicans, flamingos, and terns. While Delplanque controls the drone, artificial intelligence processes the captured images, enabling a dramatic reduction in the time required to count individual birds—potentially saving thousands of hours in survey analysis.
The urgency of this initiative is underscored by alarming statistics indicating that wildlife populations have decreased by more than 70 percent since 1970. Experts claim the world is facing a biodiversity crisis, which some scientists argue constitutes the sixth mass extinction. Historically, the planet has experienced five mass extinctions, with the last one marking the end of the Cretaceous period, approximately 66 million years ago due to a catastrophic asteroid impact.
Understanding the current status of species is a critical first step in efforts to prevent their extinction. However, this task is often complex and daunting, especially when considering that experts believe fewer than 20 percent of insect species on Earth have been cataloged. In another instance, researchers in Panama utilized AI to analyze just a week’s worth of camera trap footage, leading to the discovery of over 300 previously undocumented species.
Supporters of AI’s integration into scientific studies highlight its capacity to process extensive datasets almost instantaneously, revealing complex patterns in species behaviors and genetic structures that may evade human detection. However, the technology is not without its critics, who raise concerns about its environmental impact and the potential for biased outcomes, criticizing the lack of adequate ethical guidelines.
Currently, much AI research in conservation revolves around the analysis of extensive recordings from remote cameras or aerial flights, although its applications are poised to expand further. Researchers mainly utilize object detection models, which can discern and pinpoint species within images and videos. These models typically employ Convolutional Neural Networks (CNNs) for identifying species or determining their presence.
The decision to employ AI for conservation efforts has led to media attention, particularly in instances where it is proposed to “rescue” endangered species. In South Africa, researchers attracted headlines by investigating whether AI could locate the planet’s “loneliest plant.” This initiative involved deploying drones in the Ngoye Forest to find a female partner for a male cycad housed at Kew Botanical Gardens. The effort aimed to identify a species believed to be extinct in the wild, but possibly hidden under the canopy. Critics contend that such claims may be exaggerated, failing to consider potential repercussions.
Critics caution that extensive enthusiasm surrounding AI applications is often unaccompanied by a corresponding examination of their environmental and social costs. Hamish van der Ven, head of the Business, Sustainability, and Technology Lab at the University of British Columbia, emphasizes the need for a balanced perspective on AI’s benefits and drawbacks.
Training AI models, particularly large language models, can require substantial energy resources. In fact, the electricity consumed during the training phase can exceed a thousand megawatt hours. Additionally, the cooling requirements of data centers, which house the infrastructure supporting AI, pose further challenges as they utilize freshwater supplies for cooling systems.
Predictions indicate that AI technology will demand between 4.2 billion and 6.6 billion cubic meters of water each year by 2027, much of which will evaporate. The environmental effects of this rising demand vary geographically, especially as tech companies continue to expand their data centers internationally. Google’s plans to establish new data centers in Latin America, for example, have sparked protests in regions already grappling with severe drought.
Moreover, data centers contribute to public health concerns due to emissions of pollutants, including fine particulate matter and nitrous oxides. A report predicts that the public health impact of such facilities in the U.S., primarily located in economically disadvantaged areas, could incur costs of around twenty billion dollars by 2030.
“The models we’re running aren’t huge – they’re big for us, but it’s not like Social Network Big Data.”
For many biologists utilizing AI, the environmental impact remains minimal at present. Delplanque, for instance, performs image processing on localized computing systems, with his HerdNet model for counting species developing in approximately twelve hours, a stark contrast to the lengthy training times of larger LLMs.
Delplanque reflects on a common concern among scientists: whether their work inadvertently harms the environment they seek to protect. He reassures, stating that the models they employ, while significant for their studies, do not approach the scale of major data operations found in other applications, like social media platforms.
Tanya Berger-Wolf, a computational ecologist, argues that the current low-energy applications of AI do not fully harness the technology’s capabilities, describing image recognition as an outdated practice. Together with Pollock, she authored a paper addressing the “unrealized potential of AI” in biodiversity research.
Berger-Wolf elaborates on the need to evolve beyond mere acceleration of existing practices toward generating new, testable hypotheses and uncovering hidden ecological patterns. Pollock concurs, emphasizing the responsibility to utilize AI not just for immediate tasks but to address profound ecological questions.
One intriguing avenue generating both excitement and skepticism is the pursuit of AI-driven translation of animal communication. Initiatives like the Earth Species Project aim to employ advanced algorithms to foster communication with non-human species. Another project, Project CETI, investigates patterns in sperm whale communication, theorizing that their clicks can be deciphered similarly to morse code. Some researchers have even used machine learning to discern that elephants identify family members by unique identifiers.
Berger-Wolf stresses the importance of thoughtfully selecting specific applications for AI in conservation, cautioning against their indiscriminate use simply because of technological novelty. She posits that it would be irresponsible to allocate resources to projects when the results could be negligible, highlighting the need for data as a precious resource.
Moreover, the effectiveness of AI models hinges on the quality of the training data, which can introduce biases and lead to misdirected conservation efforts. Among the prevalent issues are spatial biases, whereby certain species are more represented in specific geographical regions, and taxonomic biases, where more charismatic species receive greater funding and attention compared to lesser-known ones. Van der Ven notes that AI has the potential to skew public perception and influence the questions posed in conservation.
Concerns persist that AI could inadvertently contribute to overconsumption and resource extraction, overshadowing its potential conservation benefits. In discussing the technology, van der Ven expressed a wish to temporarily eliminate its existence, citing an imbalance between its advantages in conservation and its usage by major corporations, such as Amazon, for profit-driven purposes.
In 2024, Google unveiled an AI model called SurfPerch, designed to monitor coral reef health by analyzing bioacoustic signals. These sounds can reveal the stability of reefs, with healthy reefs emitting distinct audio characteristics. Yet, shortly after this deployment, Google revealed struggles to meet previously established climate goals, primarily due to the environmental costs associated with its AI advancements.
Berger-Wolf insists that using AI in conservation is not inherently hypocritical, emphasizing the necessity of responsible implementation. Nonetheless, she acknowledges the complexities of regulation across geopolitical borders when it comes to both biodiversity and AI technologies.
