nAIture: Investigating Environmental Representation through Generative AI
nAIture: Investigating Environmental Representation through Generative AI

nAIture: Investigating Environmental Representation through Generative AI

nAIture: Investigating Environmental Representation through Generative AI

This case study has been developed by Felix Keilhack within the context of the final project under the guidance of Dr. Susanne Unger undertaken at the Rachel-Carson-Centre, Ludwig Maximilian University of Munich.
Full case study, methodology, annotation categories and sources can be found in the PDF.
AI-generated videos can be found on Instagram.

Introduction

Environmental images are culturally constructed and never neutral. Generative AI systems, trained on vast image datasets, operate within these cultural imprints. They not only reproduce human-environment relations but also possess the potential to create new visual elements and aesthetic forms. However, existing biases in the training data are frequently reproduced, causing epistemic injustice in the representation of vulnerable and marginalised groups (cf. Jenka 2023; cf. Kay, Kasirzadeh, and Mohamed 2024).

In light of incidents involving problematic generation of video and photographic content, this study explores the extent to which OpenAI’s video generation software Sora visualizes academic concepts from environmental and cultural studies within the realm of environmental photography and visual anthropology. This project contributes to the ongoing debate about whether, and to what extent, image-generating AIs as publicly accessible technologies bear moral and pedagogical responsibility in the context of ecological crises for the ways in which they visually represent the environment.

Given that words as tokens in large language models (LLMs) are culturally, aesthetically, and historically charged, embedded within a multidimensional vector space saturated with semantic and conceptual connotations, it is impossible to develop „neutral“ or value-free prompts that yield truly „neutral“ images. Depending on word tokens such as „nature,“ „environment,“ or „landscape,“ combined with other semantic connotations, different images are created.

To understand how AI systems portray the environment, this study employs two approaches: firstly, it uses ten academic concepts as prompt content, which are well embedded within the vector space of large language models like ChatGPT, due to extensive literature, which in turn eliminates the need to use words like „nature,“ „environment,“ or „landscape.“ Most of these different concepts have been discussed in the Reading Course of the Rachel Carson Center. Secondly, the study adopts an exploratory, ambivalent, and open-ended approach, creating a framework within which the AI itself decides how to visualise given conceptual inputs.

The following sections explain the methodology of the project, including a list of academic concepts, the research process, and annotation categories for analysis, followed by the research findings and a concluding discussion of the general results.

Methodology

Methodology can be found in the PDF.

Academic Concepts

The following prompts derived from selected abstract concepts from environmental humanities, anthropology, and science and technology studies; they are central to the investigation:

 Academic ConceptLiterature Source
1Silent Spring(Carson 2002)
2The Rhizome(Deleuze and Guattari 1977)
3Multispecies Landscapes & Multispecies Company(Tsing 2012)
4The Pluriverse(Escobar 2017)
5A critique of the Anthropocene narrative(cf. Malm and Hornborg 2014)
6Capitalocene(Moore 2015)
7Actants in Actor-Network Theory(Latour 1991)
8The Recycling Myth(cf. Colombijn 2020)
9Gaia-Hypothesis(Lovelock and Margulis 1974)
10Symbiosis & Concept of Individualism(Gilbert, Sapp, and Tauber 2012)

Research Results

1. Silent Spring (Rachel Carson)

Silent Spring by Rachel Carson, published in 1962, is a pioneering work that raised awareness about the environmental dangers of widespread pesticide use, particularly DDT. Carson documented the harmful effects these chemicals had on wildlife, ecosystems, and human health, emphasizing the need for responsible scientific and ecological practices. Her book played a key role in launching the modern environmental movement, ultimately leading to policy changes and the regulation of pesticides. 

Sora AI generates an idyllic meadow with a river contrasting against an almost cloudless sky. Under the title of a quiet landscape, one can understand why, for example, insects are missing. Although ChatGPT correctly identifies that the video does not adequately address Carson’s critique, it can be concluded that Sora has unfortunately not grasped the concept.

2. The Rhizome (Gilles Deleuze & Félix Guattari)

The concept of the rhizome, as developed by Gilles Deleuze and Félix Guattari, is a philosophical metaphor and refers to a non-hierarchical, decentralized model of knowledge and social organization that contrasts with traditional, tree-like structures of thought. It emphasizes multiplicity, connectivity, and the idea that ideas, cultures, and networks evolve through complex, unpredictable relationships rather than linear progressions. The rhizome challenges fixed, rigid systems of understanding and promotes fluid, open-ended ways of thinking, where meaning is created through dynamic, interrelated connections.

This rather neutral representation of the rhizome is very abstract and thus loses significance. A concrete interpretation and projection of the concept onto reality as an example would be more engaging. The title also fails to capture the essence of the image.

3. Multispecies Landscapes and Multispecies Company (Anna Tsing)

In her work on multispecies landscapes and multispecies company, Anna Tsing explores the intricate, interdependent relationships between humans and non-human species in shared environments. She challenges anthropocentric views by highlighting how diverse species, including plants, animals, especially fungi, and microorganisms, co-exist and co-evolve in complex ecological and cultural webs. Tsing’s concept emphasizes the importance of understanding these multispecies interactions to navigate sustainability, resilience, and the interconnectedness of life in a rapidly changing world.

First of all, where are the mushrooms? Sora decides not to depict mushrooms or humans but instead focuses on a close-up of a forest with insects that appear to crawl out of the tree roots. Perhaps Sora struggles to represent an ant highway. The title includes harmony, the weather, and the colors are idyllic, thus contradicting and romanticizing Tsing’s concept of coexistence.

4. The Pluriverse (Arturo Escobar)

The concept of the Pluriverse, as articulated by Arturo Escobar, refers to the idea of a world in which multiple, diverse worlds coexist, each with its own ways of understanding and organizing life, knowledge, and society. Escobar challenges the dominance of Western, capitalist, and modernist paradigms by advocating for the recognition of alternative, indigenous, and local practices that offer distinct perspectives on social, environmental, and political issues. The Pluriverse encourages a reimagining of development, emphasizing ecological sustainability, cultural diversity, and the importance of coexisting in a world of multiple, valid ways of being and knowing.

The videos create the impression of a somewhat touristy outdoor agricultural museum with many visitors, as stables and garden houses are scaled down to miniature size. The exact activities of the people in the videos are unclear, but some appear to be growing plants and doing gardening, while others are camping or sitting by a campfire. For the Pluriverse,Sora chooses to depict only humans in a romanticized landscape, rather than animals. The video lacks references to the struggles for liberation and resistance to modernity and colonialism; at the same time, the scene seems to reflect an outdated mindset, as if harmony in the sense of a back-to-the-roots movement is only possible by descending the linear developmental ladder to live in harmony with nature. While the title indicates that Sora may not fully understand the concept or simplifies it, I identify both the title and the video as a romanticized interpretation of the concept. Nonetheless, ChatGPT correctly recognizes that the motif evokes Buen Vivir from Sumak Kawsay and develops a hopeful representation.

5. A Critique of the Anthropocene Narrative

The critique of the Anthropocene narrative challenges the idea that humanity, as a whole, is solely responsible for the environmental crisis, emphasizing instead the complex, unequal ways in which different societies and groups contribute to ecological degradation. Critics argue that the Anthropocene’s focus on human agency often overlooks the historical and ongoing injustices of colonialism, capitalism, and industrialization, which disproportionately affect marginalized populations and ecosystems. This critique calls for a more nuanced understanding of the epoch, one that acknowledges power dynamics, the diversity of human experiences, and the need for decolonial and ecologically just approaches to addressing environmental issues. 

Cooling towers with rising smoke clouds are a common and visually powerful symbol for CO2 emissions. There is a strong contrast between industry and a dense biosphere, resembling a dualism between civilization and the green environment. Industry, as a capitalist system, forms an intense agent in contrast to the green environment. Sora opts for a top-down drone-shot perspective of the landscape rather than highlighting ground-level details. The atmosphere feels gray and seeks to convey a critical tone. While ChatGPT may recognize uneven environmental decay, I cannot confirm that. Additionally, ChatGPT critiques the absence of references to capitalist extraction, colonial history, or ecological guilt. In these videos, it is difficult to determine whether Sora is depicting the Anthropocene or critiquing the concept, which is also not clear from the generated title. Ultimately, I am relatively satisfied with the representation.

6. The Capitalocene (Jason W. Moore)

The Capitalocene, a term coined by Jason W. Moore, critiques the Anthropocene by arguing that the environmental crisis should be understood not as a product of humanity in general, but as a result of capitalism’s specific socio-economic systems and their exploitative relationship with nature. Moore highlights how capitalism, through processes of commodification, extraction, and expansion, has driven ecological destruction and shaped the planetary processes that define the current epoch. The concept calls for a focus on the capitalist mode of production as the primary driver of environmental degradation, advocating for a rethinking of human-nature relationships within the context of global capitalism’s historical dynamics.

Just as we progressed in our reading course at the Rachel Carson Center from the Anthropocene to its critique and finally to an extreme representation of certain environmental conditions in the Capitalocene, Sora generates an even more drastic video compared to the previous one: a densely built city with heavy traffic, numerous screens displaying consumerist lifestyles, cooling towers, and smoke as symbols of pollution, set against a gray, gloomy atmosphere. The videos feature an absence of nature, where capitalism is portrayed as the key actor. ChatGPT does not depict exploitation. Although the Anthropocene and Capitalocene are not drastically different, I find the videos relatively successful.

7. Actants in Actor-Network Theory (Bruno Latour)

In Actor-Network Theory (ANT), developed by Bruno Latour, actants refer to both human and non-human entities that play active roles in shaping social, technological, and natural networks. Latour challenges traditional distinctions between subjects and objects by asserting that all actants, whether people, machines, institutions, or even ideas, possess the agency to influence and transform networks. The theory emphasizes the interconnectedness of these actants and their capacity to co-create the world, suggesting that reality is produced through the interactions and negotiations among diverse, heterogeneous elements. 

These video clips share some similarities with the visualization of the rhizome concept. However, the concept is again too abstractly represented. Starting from a simplified, contentless title, much imagination is required to interpret the color-changing dots as actants. The videos represent a whole, moving, and interconnected network, illustrating the concept of relationality. While ChatGPT identifies the videos as partially correct, I find them simplified.

8. The Critique of the Recycling Narrative in comparison to Garbology

The critique of the recycling narrative highlights its limitations in addressing the root causes of waste and environmental degradation, arguing that it often serves as a superficial solution to deeper systemic issues of overconsumption and production. In contrast, garbology as the study of waste focuses on the social, economic, and cultural dimensions of waste generation, examining how material goods are discarded, reused, and valued in different societies. By emphasizing the need for a more holistic approach to waste, the critique challenges the notion that recycling alone can solve environmental problems, advocating for changes in consumption patterns, production methods, and the broader political economy.

Using various split screens, Sora contrasts the material recycling of waste (orderly, cleaned up, sorted) with the parallel landfill trash, as the title of the videos suggests. The numerous AI hallucinations, such as the text „Tryp catilang Fevo 2 entycelling waste?“ in the first video or the concert crowd and the strange image composition in the second video, remain a puzzle. Also, the correlation between landfills and seagulls in both videos is unclear, but these two elements seem connected to the theme of recycling, likely because much waste ultimately ends up in the ocean, where animals like seagulls suffer from the effects of plastic pollution. The videos have a gray tone and seek to convey a critical tone. Despite the inconsistencies, the videos provide an accurate portrayal of the recycling myth.

9. Gaia-Hypothesis (Lynn Margulis & James Lovelock)

The Gaia Hypothesis, proposed by Lynn Margulis and James Lovelock, posits that Earth and its biosphere function as a self-regulating, complex system, where living organisms interact with their environment to maintain conditions conducive to life. According to this hypothesis, the Earth’s biological, chemical, and physical processes are interconnected in ways that promote stability and balance, akin to a living organism. The Gaia Hypothesis challenges traditional mechanistic views of nature, suggesting that life and the planet’s systems co-evolve and work together to sustain habitability over geological time scales.

The focus is on the multifaceted environment, with an emphasis on the plant biosphere, with the videos excluding humans and unrecognizable animals. Sora chooses a top-down textbook-like visualization of the relationships on Earth in both videos: while the first video uses a split screen to show Earth and a cross-section of the biosphere, depicting the complexity of life across various layers, the second video contrasts a cross-section of the Earth’s layers with tree roots with a meteorological weather map showing continental plate movements. Some AI hallucinations creep into the image, where label descriptions, models, and phenomena likely make no sense and serve only for aesthetics. Nonetheless, the title and videos represent a good attempt to aesthetically depict the complexity of ecosystems on both micro and macro levels.

10. Symbiosis vs. Individualism (Gilbert Scott, Jan Sapp, Alfred Tauber)

The concept of symbiosis vs. individualism, as explored by Gilbert Scott, Jan Sapp, and Alfred Tauber, examines the tension between cooperative, interdependent relationships and the notion of individual autonomy within biological and social systems. The authors challenge the dominant narrative of individualism, particularly in the context of evolutionary theory, by emphasizing how symbiotic relationships—where different species or entities rely on each other for survival and benefit—are fundamental to the development of life. Their work advocates for a more integrative perspective, where cooperation and mutual dependence are recognized as driving forces in both biological evolution and social organization, countering the emphasis on competition and individual survival.

Sora chooses to depict a forest by a river without animals or humans, focusing instead on plants and trees. In the second video, the forest appears much lighter with individually standing trees. The videos are not photorealistic but present the environment in a mystical, imaginative way. While the illusion of separateness unfortunately does not come across, Sora connects the concept of harmony with symbiosis in the title and romanticizes the environment. Symbiotic forms can be sensed in the final images of the videos, where algae or fungus species appear to grow on tree bark. ChatGPT, however, hallucinates in its analysis of the video and identifies things that do not exist.

Overall Evaluation

In the AI-generated videos, hallucinations were to be expected, such as the disappearance or morphing of objects. For this reason, the videos were primarily evaluated in terms of their aesthetic overall concept and the accuracy of representing academic concepts through specific annotation categories, rather than focusing on the meaningfulness of details. It becomes apparent that the title serves as a guiding motif for the aesthetic implementation. The terms in the prompt and title also seem to correlate with a color palette: while „harmony“ evokes brighter colors, terms like „capitalism,“ „Anthropocene,“ or „waste“ lead to darker representations. In general, it can be confirmed that images become more expressionless as the concept becomes more abstract or complex, and that Sora AI often fails to project academic concepts onto real events, relying more on descriptive tasks. The ability to use text meaningfully or label phenomena needs improvement, as text in a video or image generator is handled differently than in Large Language Models. A split screen, on the other hand, is a helpful way for AI systems to juxtapose certain things. It is also worth noting that Sora meets the consumer’s expectation that something in a video must move—whether objects rotate, a camera pans across landscapes, or a busy scene is created; there are rarely still, quiet images.

Although the academic concepts are inherently human and have the potential to depict and involve humans, they are predominantly visualized through non-human agents, such as plants, environments, or systems like infrastructures. Depending on the theme, nature is often portrayed harmoniously, frequently in combination with the element of water. Landscape views from above, often using a drone shot, seem to dominate, perhaps aiming for a top-down overview, rather than an object-close observation. Unfortunately, 60% of academic concepts cannot be accurately represented in terms of simplifying or romanticizing the environment, with only mainstream concepts such as recycling, waste, capitalism critique, or the Anthropocene having the best chances of being depicted realistically in a normatively correct way. In evaluating the images, ChatGPT even seemed to analyze the videos more critically.

As image generators flood social media with AI-generated content, I assume that the use of AI systems will continue to increase. At the same time, they also shape society’s perceptions of responsibility by visualising the relationships between humans and ecological crises, such as climate change. Due to the complexity of AI models, it cannot be generalised from this work that OpenAI’s Sora engages in greenwashing or advocates for eco-modernism. Nevertheless, Sora’s capabilities to represent complex motifs are limited, often depicting the environment as idyllic, harmonious, romanticised „untouched nature,“ without any traces of anthropogenic influences. While Sora operates purely descriptively when executing prompts, the question arises whether Sora (like its sibling ChatGPT) bears a moral and pedagogical responsibility in generating images and videos. Although Sora appears willing to depict human-induced environmental destruction upon request, it does not recognise the historical significance of Rachel Carson’s work.

Multimedia Disclaimer

Full AI-Videos on Instagram

Autor & Redaktion: Felix Keilhack (LMU)