Bearly Recognizable: Facial Recognition and the Wild

Hank the Tank and the Criminal Animal

In the summer of 2021, a black bear was making a nuisance of himself in Tahoe Keys, a gated community in South Lake Tahoe, California. Nicknamed Hank the Tank, the bear was accused of breaking into dozens of homes and was identified as the culprit in 152 reports of human conflicts with bears (Franklin 2022). The California Department of Fish and Wildlife (CDFW) described Hank the Tank as using his “immense size and strength” to break into homes, noting that the bear was “readily identifiable due to its exceptionally large size and dark coat with lighter muzzle” (California Department of Fish and Wildlife 2022b). Judged a pest, he was targeted by the CDFW and local police, who attempted to make the easy access to food offered by Lake Tahoe garbage less enticing with paintballs, bean bags, sirens, and Tasers (Lukpat 2022).

When the deterrents proved to have little effect, the CDFW planned a special trapping effort for Hank the Tank, describing this on a website dedicated to providing information and news about bear management in the region as “official state business to capture a specific and ‘severely food-habituated bear’ defined as a conflict bear under CDFW's 2022 Black Bear Policy in California” (California Department of Fish and Wildlife 2022b). The CDFW noted that it was unlikely that they would be able to find a suitable place to relocate Hank the Tank. As a result, the problematic bear would most likely be euthanized, allowing people in Tahoe Keys and the dense bear population to return to their previous peaceful coexistence. The CDFW reminded locals that Hank the Tank would not have become severely food habituated and at risk of euthanasia if people had exercised more care in storing their food and garbage.

Shortly after these announcements, however, the story around Hank the Tank transformed. DNA testing performed by the CDFW on material left behind by the bear demonstrated that this was not the incredible activity of a single, super-strong bear but was rather the work of at least three different bears. In response, the CDFW transformed their management plan. Hank would be spared: rather than euthanizing him, the CDFW asked the community to redouble their efforts to secure trash and food and to bear-proof homes (California Department of Fish and Wildlife 2022a). The ability to individuate bears and distinguish between one Hank or many Hanks was both challenging and mattered a great deal. Was this a single rogue animal who could be eliminated? Or many animals, indicating that the problem of human–animal conflict would require more intensive intervention while also mitigating the responsibility of any individual animal, thus sparing the life of Hank the Tank? While DNA samples were successfully used to distinguish among bears in the case of Hank the Tank, ecologists and computer scientists based in North America have been pursuing alternative methods of individuating and identifying bears, notably through the use of computer vision. Bears are lively agents whose presence draws human attention, which has led to extensive photographic and video documentation of bear populations. All this (the interest and the photographs) has made possible the development of BearID, an algorithmic system for bear identification: facial recognition for bears.

The use of technology in wildlife and ecosystem management has long been an important concern in many fields, including conservation biology and the environmental humanities. Scholars have examined controversies around its deployment, as actors have variously critiqued the use of new technologies as unforgivably contaminating the wild (imagined as pure and separate from the human) or embraced it as providing scientifically valuable data (Benson 2010). Recent research has highlighted how paradoxical new forms of biopower emerge from the digital surveillance of wildlife, as digital interactions make wildlife “simultaneously more abstract and intimate” (Von Essen et al. 2021, 4). As the technology through which wildlife is understood and managed is increasingly algorithmic in nature, it is essential to bring together research on algorithmic knowledge-making and debates about wilderness and the role of technology. Through detailed attention to the creation of a bear facial recognition system, I examine how machinic processes of identification relate to longstanding debates over what the human relationship with wildlife and wilderness should be. As this paper will show, nonhuman facial recognition both emerges from and reconfigures the relations between people and wild animals. These new forms of individuation and identification produce affective relations between humans and wild animals that unsettle boundaries and categorizations like wild or domestic.

While the case of Hank the Tank illustrates the vital importance of individuation, algorithmic facial recognition promises something more. Known as reidentification technology, these tools do not just individuate bears but identify them as specific individuals who can be tracked over time and space and are attached to names or numbers. This means, rather than detecting animals and categorizing them as members of groups (by age-class, sex, or reproductive status, for example), this technology will identify and track each individual as a specific bear even as they change over time, for example, as an individual matures from juvenile to adult, or undergoes seasonal changes in size and coat. Although techniques like banding or tagging have made reidentification of animals possible, they are controversial because they require handling and contact with wild animals that could be detrimental to their well-being. The use of computer vision promises reidentification without contact or marking of animal bodies.

Through a case study of BearID, this paper contributes to work in science and technology studies (STS) that examines the constitution of algorithms through contingent social, material, and technical processes (Jaton 2017; Seaver 2019). Algorithms are often treated as mysterious, unknowable, and purely technical objects. However, following Nick Seaver's (2017) suggestions, I use interviews, observations, and textual documentation to attend to an algorithmic system as not just simply technical objects, but rather as objects composed of human practices (that is, as culture). ¹ The data I draw on here include interviews I conducted with scientists and computer programmers working on the development of BearID, a blog in which they contemporaneously documented their progress, setbacks, and experiences building the tool, and their media appearances and public statements. In addition, I interviewed scientists and computer programmers working on animal reidentification projects more generally, as well as conducting participant observation at a month-long summer school that trained ecologists to develop computer vision tools. Altogether, this dataset of 21 interviews, texts, and fieldnotes provides insight into the collective human practices, materials, tools, and nonhuman life forms that play roles in the production of algorithmic systems. By examining how bear facial recognition is made, this paper considers the algorithm as just one component of a network of contingent elements that included brown bears in the wild, a graphics processing unit (GPU)-equipped computer named Otis, ² a competition in which people cast votes for the fattest bear, and scientists setting camera traps. This methodology crucially allows me to site the production, training, and use of this algorithm within a larger sociotechnical system, and to highlight how the production of nonhuman animal facial recognition is shaped by human beliefs and responses to charismatic animals—which plays a vital role in fortifying that charisma. Drawing on Jamie Lorimer, I understand charisma to be relational. Rather than an essential quality of an animal, “nonhuman charisma emerges in relation to the parameters of different technologically enabled, but still corporeally constrained, human bodies, inhabiting different cultural contexts” (Lorimer 2007, 916). Charisma or lack thereof shapes human engagement, research, and understandings of the moral status of animals, and thus approaches to management and conservation (Clark 2015). In this case, algorithmic facial identification vitally enhances the charisma of bears by allowing people to identify and relate to them as individuals. I first consider how animals have historically been recognized and identified as individuals, and then turn to a detailed examination of the making of a facial recognition system for bears.

The Face and Individuation

Individual identification is a keystone of human governance, policing, and other bureaucratic practices: see the myriad projects to produce means of individual documentation, including drivers’ licenses, passports, and birth certificates. The specific technical and legal means by which identities have been connected to physical bodies range from the fingerprint to more recent efforts to use DNA, vocalizations, or facial features (Cole 2001; Kang 2022; Lynch et al. 2008; Nieves Delgado 2022; Caplan 2002; Taylor, Gulson, and McDuie-Ra 2023; M’charek, Hagendijk, and de Vries 2013). These efforts share a presumption that bodies can be concretely linked to a stable, unchanging identity (Grünenberg et al. 2022). The face, subject to computational analysis, has been a particularly intense target as a means for reidentification. Companies developing facial recognition powered by machine learning promise that it will provide a technological fix for all kinds of human governance and management needs, from the weighty (controlling the movement of people across borders and the identification of criminals) to the mundane (identifying who exactly is in each image in the ever-growing photo database on your phone) (Gates 2011; Grünenberg et al. 2022). However, the extensive social science research on facial recognition technologies demonstrates how these (and other big data or algorithmic practices) reproduce inequality, discrimination, and racism (O’Neil 2016; Noble 2018; Benjamin 2019). Discriminatory aspects of facial recognition are built into the technology at several levels, from the datasets used to train and test algorithms to the statistical procedures used to categorize people (Nieves Delgado 2022; Buolamwini and Gebru 2018).

The extent of the issues with human facial recognition have led Luke Stark (2019, 50) to call it “the plutonium of artificial intelligence (AI),” a dangerous technology with little potential for good. These problems have not gone unnoticed by big tech companies. Notably, Facebook (now Meta) announced on 2 November 2021 they were ending their facial recognition program (Pesenti 2021): the company observed the use or removal of facial recognition technology involves tradeoffs—the loss of key benefits provided by facial recognition technology versus individual rights to privacy, noting, “We need to weigh the positive use cases for facial recognition against growing societal concerns.”

While the face has been the subject of technological innovation, anthropologists Amade M’charek and Katharina Schramm (2020, 321) point out that “the face as an object of critical inquiry is rather absent” and call for analysis that denaturalizes the face and attends to the conditions of its production, as well as its status as a relational object. Human faces are political because they are deployed to sort people into classificatory categories such as race or gender (M’charek and Schramm 2020; Plemons 2017). If the face is a site for the production of sex/gender and race, it is also one for the production of species, as it becomes the site of species identification and categorization (particularly in applications like computerized facial recognition). In contrast to fears around human facial recognition, animals are often held up as unproblematic targets for facial identification: a positive use case with none of the concerns about privacy, misrecognition, racism, or sexism that have dogged human facial recognition. Computer scientists and ecologists have been expanding the field of facial recognition to encompass nonhuman animal life, deploying AI and machine learning in the service of new forms of wildlife identification and individuation.

Species and Individual Identification

Identification of animals often occurs at the level of the species. One common purpose for species-level identification programs is to train people to recognize species that have been marked as “bad” or invasive (Gallo and Waitt 2011; Moore 2019). For example, with the arrival of the spotted lanternfly on the east coast of the United States, government agencies have launched programs training people to identify them at different life stages, to destroy egg masses and kill individuals. ³ The eradication (or attempted eradication) of species labeled as invasive is only the most dramatic example of how categorizations and forms of species identification within scientific or bureaucratic apparatuses can transform the trajectories of species (Mansfield 2003; Helmreich 2005; Lowe 2006; Swanson 2019). For instance, as Jennifer Telesca (2020) has described, the categorization of bluefin tuna as a commodity or biological asset has had existential consequences for the species because it enables a fisheries regime of managed extinction. These programs manage interactions on a species level, directing attention to the impact of species on ecosystems, or the resilience of a species as a whole.

In other cases, what matters is not identifying the species (because it is well known or obvious to the casual observer, as in the case of bears), but rather the individual animal. The interests in identifying individual nonhuman animals exceed the determination of guilt and criminality detailed in the case of Hank the Tank. Aníbal Arregui (2023, 2) argues for the importance of infraspecies interactions in which animals are approached not as members of a species, but as individuals “dealt with according to their own idiosyncrasies.” In the case of domesticated animals, individual animals are identified in order to mark ownership, establish pedigree and value, manage the spread of disease, and enable the traceability and quality control of animal products, among other reasons (Eradus and Jansen 1999; Greene 2010). However, for most human observers, as the case of Hank illustrates, individual animals are difficult to distinguish from one another. Animal physiologist Philip Dziuk (2003, 319) bemoaned the difficulty of recognizing animals, noting, “Unfortunately, few if any, persons that deal with relatively large numbers of animals are able to distinguish each animal from others by general appearance with any certainty.”

Technology and Identification

To solve the problem of identification farmers have deployed a range of methods, from classic tools like ear notching and tagging, branding, and tattooing, to electronic and digital innovations, like radio frequency identification tags in the 1970s. More recent developments include tags with Quick Response (QR) codes that log the movement of sheep and chicken to document the provenance and treatment of animals, for example, as a way to prove to the customer that the animal was truly free-ranging and so worth playing a premium for (Glen 2019; Wang 2020; Bamforth 2021; Awad 2016). These electronic transponders have allowed for increased automation and monitoring of animals (Eradus and Jansen 1999). Agricultural tracking programs may occur at the level of the farm or at larger scales: for instance, since 2002 the United States Department of Agriculture has been working on a national agricultural animal tracking program to enhance disease traceability (Greene 2010).

A classic example of technology for the documentation and identification of domesticated animals is the General Studbook, a breed registry recording the pedigrees of thoroughbred racehorses in Great Britain and Ireland. First published in 1791, it predates the compulsory recording of human births in England (Cassidy 2002). Breed registries for other livestock breeds followed, identifying each animal by name and lineage. The documentation of pedigree was an effort to introduce control over procreation and future generations; it reflected the belief that the qualities and attributes of animals could be shaped and directed by human management (Ritvo 1987). As Harriet Ritvo (1987, 60) argues, the level of documentation required to track pedigrees was important for recordkeeping, and was also consequential for relations between humans and nonhuman animals, as it “increased the personal dignity and individuality of the animals, making it easier for people to identify with them.” Practices of individual identification of horses represent their complex relations with humans: pedigrees and names both signal horses’ status as the objectified products of human industry and control, while also individuating them in ways that enhance their relations of intersubjectivity with caretakers and owners (Cassidy 2002). This alternation between control and identification is all very well for the racehorse or other domesticated animal, which, while admired for its freedom and independent spirit, is ultimately regarded as a product of human design and intention. However, the appropriateness or desirability of both the control and identification with the human implicated in animal individuation takes on different valences in the case of the wild animal.

Challenges of Bear Identification

While domesticated animals may be properly subjects of control, debates over what wilderness is and should be have situated wild animals differently vis-à-vis identification projects. In wildlife biology and ecology, identifying and tracking individual animals has been an essential tool for answering fundamental questions about animal life and behavior. In the case of bears, scientists track individuals to understand bear movement, population size and density, and social behavior in general. In wildlife management, the issue may be how to identify individuals of a particular species, often because while the species is regarded as valuable and worthy of care, specific individuals are marked as problematic or criminal. People have developed multifarious techniques to deal with the problem of identifying wildlife. Writing about encounters between people and monkeys in the central Himalayas, Radhika Govindrajan (2015, 247) describes specific monkeys being identified as outsiders who are causing problems. One interlocutor says “We know that these monkeys are outsiders in the same way that we know when people are outsiders. Their walk, their habits, the way the look, everything marks them as outsiders.” Identification here is based not on specific markings or appearance, but a gestalt, a look. It is everything and nothing that marks the identity of these monkeys. In another case, the governance of tigers in India is shaped by conservation law that dictates that only tigers who have been positively identified as having attacked a human can be hunted (Margulies 2019; Mathur 2021). Both the management of tigers and bears (as the case of Hank the Tank illustrates) often rely on the identification of individuals to classify them as killable, rather than protected. Nayanika Mathur (2021, 170) describes how identifying big cats involves both bureaucratic rationality and forms of intimate knowledge that exceed bureaucratic rationality: “A tiger is not just any tiger. To identify it means to know who it is—its tastes, gender, personality, territory of operation, idiosyncratic tics, mannerisms, and even looks.” The individual identification of wildlife in these cases requires engagement and exchange between human and nonhuman. Attention must be paid not only to the animal as a species, but also as an individual, requiring knowledge of characteristics that are generic to the group and also specific to the individual.

Visual identification of bears in this manner is challenging, but not impossible. The key diagnostic characteristics with which a person might identify a brown bear include body size, shape, fur color, ears, and face. However, bears’ body types and appearance change radically over the years and even within a single season, as bears emerge emaciated from winter hibernation and grow tremendously fat over the course of the summer. Fur color likewise changes over seasons and life spans. Brown bears generally look quite a lot like one another and also not particularly like themselves, making it hard for people to consistently identify them by appearance. As a result, individual identification also relies on observations of behaviors. Rangers at Katmai National Park in Alaska (home to 2,200 brown bears and famous as a site for bear viewing and research) have developed techniques for identifying bears. They sort bear behaviors into instinctual and learned. Learned behaviors are key to individual bear identification: knowing who a bear is requires a sense of what all bears do, and what makes this particular one distinct. A guidebook to Katmai produced by the U.S. National Park Service (2022, 9) notes: “Bears have an instinct to eat high calorie foods like salmon, but fishing is a learned behavior. Not all bears fish in the same way or in the same places. This offers great insight into their individuality.” Other individualized behaviors include levels of aggression, subordination, and tolerance of people. The discussion of bear identification on the Katmai website states that “Individual bears are difficult to identify, especially the first few times you see them, but with practice anyone can identify the most commonly seen bears along the Brooks River” (National Park Service 2021). Identification of bears requires substantial attention, work, and connection with bears, whether in person or through hours of webcam observation.

However, even with extensive practice and experience, identifying bears by sight remains difficult. In an interview with explore.org, a philanthropic multimedia organization that produces documentaries and hosts livestreams of wilderness sites, Naomi Boak, a park ranger in Katmai National Park, noted that the question rangers get asked more frequently than any other is “who is that bear?” and that her “most frequent answer is ‘I don’t know’” (Explore Bears & Bison 2021). ⁴ Scientists often rely instead on technological means of tracking and identifying individuals. There is a long history of using physical devices to address questions of wildlife identification. The simplest of these devices consist of numbered tags or bands, which have been used systematically since the early 1900s (Lincoln 1921). Subsequent developments include tags that can transmit data remotely, including very-high-frequency radio telemetry and satellite tracking devices like Global Positioning System (GPS) tags (Benson 2010; Cagnacci et al. 2010; Hebblewhite and Haydon 2010). Etienne Benson's work on the development of radiotracking documents debates over its use, highlighting that contestations over the deployment of technology were often, at heart, about what wilderness is and should be. As Benson details, wildlife biologist Adolph Murie's critiques of projects that conspicuously tagged or marked animals like bears are exemplary of the issues at stake in wildlife marking and identification. Murie argued that visible technology negatively affected the aesthetics of the park and visitor's experiences. Marking and identifying animals destroyed their wilderness qualities, which were characterized by their separation from the human. These kinds of arguments were typical of the pushback against radiotelemetry projects, which were often characterized as contaminating parks, in contravention of their intended purposes as bastions of raw, untouched nature (Benson 2010). Despite such critiques, radio and GPS collars affixed to bears play important roles in contemporary research. This does not mean that concerns about their use have been resolved, especially because in addition to their impact on wilderness experience, their deployment is both expensive and stressful for bears. Other means of bear identification which do not require direct contact or interaction with bears like genotyping hair or scat are possible, but they are generally expensive and time-consuming. ⁵

Could a Computer Recognize a Bear?

It is in this context that people began to consider the possibilities of bear facial recognition. In 2017, two computer scientists, Ed Miller and Mary Nguyen, were searching for a project that would allow them to hone their skills and knowledge of a field of computer science known as “deep learning.” Deep learning is a subset of machine learning which makes use of artificial neural networks (algorithms inspired by the structure of the brain). Specifically, they wanted to work on a computer vision project, developing a tool that would use a computer to interpret digital images or videos. Computer vision approaches have been championed as ways to more accurately identify animals without corrupting them through human contact. Deploying camera traps and analyzing the resulting data with computer vision could also be a cheaper and more efficient way to study a larger number of bears.

Visual identification of bears was a practice Miller and Nguyen were already familiar with. They were inspired by their own fondness for watching bears on webstreams from Brooks Falls in Katmai National Park on the Alaskan Peninsula, something they termed a “guilty pleasure” (Miller 2017c). While they enjoyed watching bears, it was a challenge to identify and distinguish between them. Miller and I met virtually to discuss his work on BearID. As I sat in my Pittsburgh office, he appeared on my computer screen, zooming in from a nondescript office space in California, where he lives and works for a major semiconductor and software design company. Describing the origins of BearID, Miller told me, “We would see people talking about which bears were being seen on the camera and talking about how they could identify bears. And at first, we were like, ‘what are you talking about? They all look the same.’ As we were watching the cameras though, we could start to recognize some of those bears as well.” This prompted what Miller referred to as an “aha moment.” If they could learn to recognize bears, presumably machines could be made to do so too. Through the conservation tech network WILDLABS, Miller and Nguyen connected with Melanie Clapham, a behavioral ecologist studying bears in British Columbia, who wanted to identify bears in order to study their social behavior and interactions (Miller 2017d). ⁶ The composition of the BearID team indicates the range of people who may use the tool, amateur bear watchers and professional scientists alike.

I met Clapham on the sunny, southern California campus of Caltech, where a group of ecologists had gathered to attend the Computer Vision for Ecology Summer Workshop, to learn how to apply computer vision tools to their research questions. The school took place over an intensive 3-week period, during which students grappled with the theory and practical details of applying computer vision to ecological systems. At one session, Clapham explained her research to the group. Standing in front of the cohort of students and their instructors, she described her work and research site in the Great Bear Rainforest in Da’naxda’xw/Awaetlala territory in British Columbia, Canada. The Da’naxda’xw First Nation owns and operates an eco-tourism lodge where visitors come to watch bears. Profits from ecotourism get funneled back into research: they provide logistical support and funding for Clapham's research, a long-term study to monitor individual bears, their social communication, and human impacts. She noted the lineage of researchers at the site: in addition to her own work, her PhD supervisor and his PhD supervisor both conducted research there. She had a long-term commitment to the site, which meant that she had deep knowledge about the bears. She observed, “It's nice to be able to work on a species that I’ve spent a lot of time around, so I feel quite a connection to it, not to anthropomorphize things. I spent 10 years at one field site, working on one population of bears.”

She was at the summer school to develop her programming skills and to work on a deep learning machine model that could identify specific bears from camera trap video footage. For her research on bear behavior, it was essential to identify individuals. Training a deep learning model to identify bears could help ecologists and wildlife managers to track individual bears across space and time, and thus do things like produce population estimates (which is not possible with current techniques), and better understand bear behavior and human impacts on bears. More generally, individual identification would provide data to support evidence-based decision-making for wildlife managers.

Midway through the summer school, Clapham and I got lunch together. We sat at a table outside one of the Caltech dining facilities, enjoying the relative calm of summertime at a university campus. Clapham reflected on her motivation for developing BearID and the challenges of bear research, noting “I spent a lot of time watching and trying to recognize between different individuals and understand different individuals and behaviors....It's so difficult to recognize between individuals, as you know. They change a lot within a season and across different years as well.” It was this difficulty that drew her to the project. After watching camera traps every day for six months, she gradually realized she could learn how bears changed and could teach herself to recognize them.

If you had ten pet dogs at home, because you see them every day, you’d be able to tell who they were....Even if they look really similar you still would generally pick up which bear is which. Because we were seeing these bears almost every day, not just on camera, but actually watching them and doing direct observations we came to just be able to recognize them, but it was sometimes difficult to explain to other people how.

Like Miller and Nguyen, Clapham wondered if there was a way to automate this process. Their shared interests in bears and machine learning informed their collaboration on a program that would allow a computer, provided with images of bears, to recognize and identify them: facial identification for bears, which they called BearID.

Computer vision introduces new forms of identification and ways of knowing animals that diverge from the human and nonhuman animal engagement that characterizes traditional bear identification. The tools that humans use to identify bears are not amenable to computer vision: behavior, for example, is too computationally intensive to serve as a useful marker of individuality (Miller 2017b). Likewise, many of the bears’ bodily characteristics, like coat color and body size, are too transitory to form the basis of biometric identification.

The use of machine learning and computer vision to identify and track animals is not unique to BearID. Computer vision programs to identify other wild animals exist: Wild Me, a nonprofit organization focused on combining artificial intelligence, computer vision, and citizen science has developed an open-source software framework that uses machine learning to identify animals. Animals like manta rays, whale sharks, and zebras all have features that are amenable to analysis and comparison using pattern-matching algorithms—in fact, Wild Me's whale shark ID platform was built using a modification of an algorithm originally used to compare the locations of stars (Arzoumanian, Holmberg, and Norman 2005). Polar bears could also be identified with a similar sort of pattern analysis—researchers at the University of Manitoba developed the “Whiskerprint project” which uses whisker spot patterns to identify polar bears (Anderson et al. 2010). Unfortunately for computer vision scientists, brown bears do not have similar visually striking, unique patterns. This meant that the scientists involved in BearID needed to develop alternate methods of recognition.

Miller and Nguyen determined that, as with humans, bear faces would be the best feature upon which to build recognition systems. The geometry of bear faces remains stable over time, making them a good target for identification. Crucially, focusing on the face would also allow them to leverage the intensive investment in tools designed for human faces. Bear and human face geometry are similar enough that tools designed for humans could be adapted for bears. The decision to focus on the face was shaped by both the characteristics of bears and the existence of tools like well-characterized data processing pipelines and computer vision algorithms to detect and identify human faces. Specifically, they identified the method used by the developers of Google's FaceNet facial recognition system as a promising process for bear facial recognition. They noted that “Deep learning applications for human face recognition routinely achieve greater than 99 percent mean classification accuracy. Our goal is to adapt a deep learning human face recognition algorithm for use with bears” (BearID Project 2023). Deep learning is a form of machine learning that uses artificial neural networks to process and analyze data. The pipeline developed by the BearID team to identify bear faces uses a computer algorithm to detect the face within an image and then analyzes those faces, producing a numerical representation of the face known as an embedding or vector. This embedding is then compared to embeddings from previously identified bears in an effort to identify those closest to it. The closest embeddings likely represent images of the same bear (Clapham et al. 2020).

Creating BearID required a patchwork of different tools, as Miller and Nguyen tried out existing object classifiers and detection tools. To detect bears in images, they experimented with pretrained object detection systems like You Only Look Once (YOLO). Based on the Common Objects in Context (COCO) dataset, YOLO can detect the 80 preassigned COCO object classes. Fortunately for those interested in bear detection, bears are one of the ten animals included in COCO (along with birds, cat, dogs, horses, sheep, cows, elephants, zebras, and giraffes) (Lin et al. 2014). Choices of which object classes to define in datasets are products of varied factors, including the particular problem the dataset is intended to solve, what objects are familiar and important to dataset designers, and categories that can be accurately identified by the people tasked with labeling images. ⁷ Other tools conscripted into bear facial recognition work included a photo filter called Dog Hipsterizer, developed to add sunglasses and mustaches to dog faces. The similar face plan of dogs and bears made it reasonably adaptable to the task of locating landmarks (eyes and snouts) on bear faces.

Miller and Nguyen's initial attempts to implement the first part of the pipeline (training a detector to recognize a bear face within an image) were stymied by their lack of a computer with a GPU. Unlike central processing units (CPUs), which process data sequentially, GPUs are capable of doing many calculations in parallel, essential for rendering three-dimensional graphics or for machine learning. Frustrated by the limitations of their GPU-less computer, they were inspired to build their own GPU-equipped deep learning computer which they dubbed “Otis,” after one of the most popular bears on the Brooks Falls webcams. Otis made it possible for them to implement the face recognition pipeline they had proposed (Miller 2017a). Although algorithms are often imagined as immaterial, exclusively conceptual tools, the work of BearID's programmers to implement this pipeline highlights the social and material dimensions of constructing an algorithm, which is shaped by the work of gathering data (including building collaborations with scientists and others on the ground, as well as leveraging the popularity of bears), as much as the physical limitations of computing technology.

Making “Ground-Truth” and Datasets for Bear Faces

In addition to relying on the creative redeployment of tools produced for other purposes, creating BearID relied on the existence of a “ground truth” dataset. In this section, I unpack the history and form of the dataset on which the bear facial recognition algorithm was trained and tested. The design and construction of datasets (collections of labeled images) are essential to making machine learning work. Any machine learning for facial recognition, whether directed at humans or wildlife, requires extensive data. Machine learning for human facial recognition makes use of the existence of enormous datasets of identified human faces. Datasets of this size do not exist for nonhuman animals, so one of the challenges for any wildlife identification project is developing or acquiring sufficient labeled datasets, consisting of images in which the animal in question has already been identified (Clapham et al. 2020).

Datasets are human constructions, ones which enact values and politics. By analyzing the process of dataset creation, such as which data to collect and how to categorize and annotate them, we can surface how they are shaped by the culture in which they are produced (Jaton 2017; Denton et al. 2021). In this case, the contours of human attention to and relations with bears shaped the existence of datasets about bears and the possibilities for facial recognition. It was only possible for BearID to produce a computerized facial recognition system because there was already a lively community of observers dedicated to identifying bears. As mentioned earlier, the bears of Brooks Falls in Katmai National Park and Knight Inlet, British Columbia (where Clapham works) have been subject to decades of intensive observation by scientists. They are not anonymous wildlife; instead, they are well-known as individuals, who have long been photographed and documented. Drawing on this previous work, BearID was able to assemble a dataset of 4,674 labeled images of 132 known individuals collected at Knight Inlet, British Columbia, and Brooks River, Katmai National Park, Alaska. The images from Knight Inlet were collected by Clapham and other naturalists for research purposes, while those from Katmai National Park were taken by National Parks Service staff and independent photographers at Brooks River, for a bear monitoring program and for personal use (Clapham et al. 2020).

Scientists are not the only ones interested in bear images (as the popularity of bear cams indicates), and these images are not created in a vacuum. Bear surveillance lives at the intersection of wildlife and ecological research and management, and animals on social media. Internet presence is not just for humans: nonhuman animals proliferate on social media platforms (Chua 2018; Linné 2016; Turnbull, Searle, and Adams 2020). The bears of Katmai National Park have become social media phenoms with the advent of “Fat Bear Week” in 2014. Fat Bear Week is an annual competition hosted by the multimedia organization explore.org, in collaboration with Katmai National Park, in which bears are matched head-to-head in a single elimination tournament. Viewers are asked to vote for “the bear who you think is the fattest.”^⁠ Obviously, voting is not an effective way to determine which bear is fattest. Instead, people are encouraged to vote for whatever bear they find most deserving of recognition. The rules note that “Fat Bear Week is a subjective competition. Be sure to vote and campaign for your favorite candidate” (explore.org 2022).^⁠

While the Katmai bear cams had already attracted a large, passionate audience since their establishment in 2012, Fat Bear Week drew new viewers to the bears, producing a community of bear viewers who are fascinated with particular individuals. In 2021, 793,463 people voted in the competition, making it the most popular yet (Kraft 2021). The winner was bear 480, also known as Otis (the namesake of the main BearID computer), in his fourth victory in the competition (Diaz 2021). In the context of Fat Bear Week bears become the subject of parasocial relationships, scripted into narratives imagined by the viewers. Bears with long histories or who are perceived to have overcome struggles are celebrated. People actively campaign for others to vote for “their” bear, producing memes, campaign slogans, and long narratives representing individual bears as particularly worthy of respect. Campaigning for Otis on the explore.org message board was intense. One user with the screenname Bearb said: “His will to survive was amazing. He became, to me; a real inspiration. Otis never gave up; he persisted, and that is why he is here today. A lot of bears returned looking thin and scruffy, but Otis worked the hardest to qualify for Fat Bear 2021”^⁠ (explore.org 2022). Commentators lauded Otis as an exemplar of an organism surviving and thriving despite substantial challenges (specifically a late arrival at Brooks Falls, age, reluctance to confront or displace other bears from prime fishing locations, and missing teeth). In viewing video footage and photographs of the bears, viewers became invested in their lives, constructing narratives and posting rationales about why one bear is favored over others.

Although the photographs taken by rangers may be representative of a random selection of bears, they are more likely to be nonrandom, shaped by some of the same elements that draw viewers’ attention to particular bears. Further, there is a clear demand for representation and images of certain bears. The lively audience for Katmai bear images clamors for the appearance of their favorite bears, and an image of Otis, for example, has a built-in audience that the image of a less famous bear will not have, rewarding the ranger (or tourist) who captures a particularly compelling shot of that bear. These preferences and the narratives audiences construct about bears may shape how they are documented and which bears get represented in datasets.

Technologically Constructed Charisma

While implementing BearID and other algorithmic methods for individual identification does not entail physical contact with bears, nor do they visibly mark bear bodies, they are nonetheless consequential for the relationships between humans and bears. In an interview, Miller noted that he hoped BearID would eventually be able to provide automated identification of bears as they appeared on bear cams, stating, “One day I hope to receive a notification that lets me know when Otis is on camera so I don’t miss any moments when the big guy is out there fishing” (Explore Bears & Bison 2021). While identification was previously limited to experts or to the kind of passionate amateurs who devoted hours to wildlife camera watching, algorithmic tools mean that no expertise is necessary to attach a name or ID number to an animal, that is, to know it as a specific individual. Automating identification could grow the audience for things like Fat Bear Week and bear cams, by allowing viewers to more speedily bond with specific bears.

Connecting with animals as individuals transforms human perceptions and attitudes. Writer and artist Jenny Odell describes the transformative consequences of recognizing that the crows who visited her window were specific individuals. She notes that learning who these individuals were was part of her developing a deeper understanding of the landscape around her, one that entailed responsibility and commitment, noting that “eventually, to behold is to become beholden to” (Odell 2019, 146). Seeing an animal as individual is essential to perceiving nonhuman personhood in animals, and to developing a concern for them as more than resources for human use or conception (Milton 2002; Lorimer 2007). In analyzing the significance of individuation and attributions of personhood, Owain Jones (2000, 281) argues that “Any possible switch from relating to nonhuman others as collectives to relating to them instead as individuals has profound implications for how we live on this planet, and may be a significant narrative for the future.” By technologically enhancing the human capacity to recognize and identify bears, BearID is a machine for producing bear charisma and concern for bears. Human interest in animals has important impacts, underpinning conservation strategies and scientific investment.

Caring is consequential, but it is not always desirable or embraced. The relationships between humans and bears that algorithmic identification facilitates are sometimes regarded as more characteristic of those with domesticated animals, and thus inappropriate or detrimental to bears’ wild characteristics. The rangers at Katmai point out that individually identifying a bear (whether it is with a name or a number) may appear to make it “less wild, and more pet-like, than an unknown counterpart” (Katmai National Park and Preserve 2022, 18). They regard names in particular as framing human observation of bears in ways that interfere with a more pure wildlife experience, producing one altered by the meanings we inherently attach to names.

Despite fears of domestication and excessive intimacy, BearID also can be regarded as opening new possibilities for scientific research. Difficulties in reidentification of animals has shaped ethological and ecological research, moving it away from analysis of individual behavior and life course toward elements like song, specific interaction, or populations (Benson 2016). Automated reidentification facilitates research that attends to animals as individuals and their behavior over time. Understanding bears as individuals with life histories can influence behavior interpretation and management decisions, for example by allowing wildlife managers to see when an individual bear has become habituated to humans, or when an encounter with humans is out of character.

Clapham highlighted the complex ramifications of identification, telling me that

an effect of this project could be drawing the focus on individual animals, which from a research perspective sometimes people frown on because we shouldn’t be doing stuff based on an individual level. We should be doing everything based on a population level. But people don’t engage with populations and numbers. People, human beings, engage with individuals and with faces.

Recognizing bears as specific individuals transformed people's relationships in ways she regarded as productive, even if it made bears seem less wild. Clapham described the power of individual storytelling about animals: “This is one individual female. This, you know, this is how old she is, this is her number. This is her name and look at how different she can look over the years, and over different seasons. So, just showing people is opening up their lives in a way that people wouldn’t have access to.” Individuation led to human connection, which ran counter to the notion of bears as radically removed from humans, as in the discourse around wilderness and nature. Identification does not domesticate bears but nevertheless brings them into a new relation with human viewers.

Conclusion

By bringing together algorithmic knowledge-making and the study of nonhuman life, this paper has attended to how AI and machine learning are, of course, social products. Even when applied to the wild, the forms they take are shaped by human interests, resources, and choices. Further, they have transformative potential for human understandings of wildlife. Scholarship in the environmental humanities has asked about the practices of immersion, attentiveness, and inquiry we adopt when engaging with multispecies worlds (Van Dooren, Kirksey, and Münster 2016). Technology for wildlife is one critical node in our engagement and attention to the natural world, shaped by existing interests and relations with animals, as well as producing new ones. AI and machine learning are becoming part of the cultural and technological apparatus mediating people's relations and engagement with bears. Examining this case of facial recognition beyond the human attunes our thinking to the multispecies production of technology and its impact.

Forms of identification produce relationships. As machine learning offers new forms of identification, it also generates new possibilities for human understanding and interaction with bears. It creates possibilities for mass individuation, enabling observers to see bears en masse, and to more easily treat them as individuals. Historically, animals have only rarely been singled out, appearing more often as undifferentiated masses and only at rare moments becoming objects of affective relations with humans (Buller 2013). Charismatic bears, in an age of increased opportunities for observation and digital connection offered by webcams, are an exception to this.

Knowing bears as individuals changes human interactions with them, both in terms of governance and affective relations. Individuating bears has consequences for determining criminality and deciding whether to euthanize or relocate, as the case of Hank the Tank demonstrates. It also shapes affective relations with different bears, as Fat Bear Week makes evident. By individuating bears in a particularly intimate way (through the face), facial recognition technology troubles categories like wild and domestic. This is perhaps even more intimate than naming, which has long been a hallmark of individuated, special animals such as favorite horses, cows, or pets and, as mentioned earlier, a practice treated with some concern and trepidation by scientists for its potential to produce inappropriately close connections between humans and wildlife. Technological individuation through face recognition challenges expectations around nature and the wild, which call for separation and that refuse the kind of domesticated intimacy that comes with naming and individuation.

These processes of individuation mark not only technological change, but also a particularly anthropocenic moment in which the abundance of wildlife is dwindling. This decline is notable in mundane moments through things like the windshield phenomena, in which fewer insects accumulate on a driver's windshield, and in more spectacular ones, like the increasing rarity of encountering large, charismatic wildlife (Vogel 2017; Finn, Grattarola, and Pincheira-Donoso 2023). Individuation of animals tends to occur as they are brought directly under human control through processes of domestication, or as their numbers lessen to a point where individuals can be known. The combination of emerging machinic recognition processes alongside the widespread defaunation of the globe and the dwindling numbers of wild animals means that human engagement with multispecies worlds likely will entail processes of identification and individuation described here.