Beyond Zoocentrism: How AI Bias Impacts Biodiversity Education and Conservation
The increasing presence of Artificial Intelligence (AI) in our daily lives has naturally extended into the realm of education. While AI offers numerous advantages, a recent study has brought to light significant concerns regarding its reliability, particularly when it comes to teaching complex subjects like natural sciences and, more specifically, biodiversity. This concern is substantial because biodiversity conservation is fundamentally linked to human sustainability and is a recurring topic in educational curricula worldwide, including in Spain. The study rigorously tested several widely used AI tools—ChatGPT-4.5, DeepSeek-V3, and Gemini—by asking them to generate a list of endangered species. The findings were stark: all tools exhibited both taxonomic bias (an uneven focus on certain types of organisms) and geographic bias (a skewed representation of species from particular regions) when compared against the authoritative data from the IUCN Red List. These imbalances are not mere technical glitches; they carry the potential to reinforce existing societal prejudices such as plant blindness (the tendency to overlook plants), zoocentrism (prioritizing animals), and Western centrism in classrooms, especially where educators may not have specialized scientific training. Ultimately, this research underscores the urgent need for AI models to be refined using accurate scientific datasets and for teachers to develop AI literacy to mitigate the spread of misinformation in biodiversity education and learning aligned with Sustainable Development Goals (SDGs).
The journey of artificial intelligence began in the mid-20th century, with pioneers like Alan Turing envisioning machines capable of cognitive tasks. The term "artificial intelligence" itself was coined in 1956, establishing AI as a distinct field of study. Over recent decades, rapid advancements in areas like machine learning and deep learning, coupled with access to vast amounts of data, have propelled AI to unprecedented levels of sophistication. Modern AI systems, exemplified by tools such as ChatGPT, Gemini, and DeepSeek, now possess advanced capabilities for human-like interaction, complex reasoning, and content creation. These systems are powered by sophisticated language models trained on extensive datasets, enabling them to perform a wide array of tasks, from processing text and generating multimedia to tackling complex challenges in fields like medical diagnostics and autonomous navigation.
In education, the integration of AI has been hailed for bringing about significant advancements in learning solutions. These tools are characterized by their ability to offer personalized learning experiences and increased accessibility across various academic domains. AI can act as virtual tutors, providing explanations tailored to individual student needs, which can enhance knowledge retention and student engagement. Furthermore, AI applications can assist educators by automating administrative tasks, such as grading assessments and creating instructional materials, thereby freeing up teachers' time for more direct student interaction. Perhaps most notably, AI holds the potential to democratize access to high-quality education, especially in regions with limited resources, by bridging educational gaps through real-time tutoring via chatbots, thus mitigating geographic and economic barriers.
However, the widespread adoption of AI in education is not without its challenges. A primary concern is the potential for overreliance on technology, which could hinder students from developing crucial cognitive abilities like independent thinking and problem-solving. Ethical considerations are also paramount, particularly concerning data privacy, potential bias in algorithmic decision-making, and the risk of diminishing the essential role of human educators in the learning process. Experts emphasize that for AI integration to be effective, it must serve as a complementary instrument to teachers and human interaction, not a replacement.
Within the natural sciences—encompassing disciplines such as biology, physics, chemistry, and earth sciences—AI offers new pedagogical and research opportunities. These tools can enrich the learning experience by providing innovative resources like customized explanations, interactive simulations, and diagrams that clarify complex scientific phenomena. AI can also facilitate the creation of thematic questions and exercises, enhancing classroom engagement and serving as an instrumental tool for incorporating research methodologies into educational settings. Despite these benefits, the indiscriminate use of AI in scientific education also raises significant concerns. A major risk is that students might use AI primarily for problem-solving or text generation rather than engaging in deep analytical exploration of scientific concepts, which could impede the development of critical thinking and scientific literacy. This brings us to a crucial question: Do popular AI applications provide scientifically accurate and unbiased information, or do they introduce distortions that could mislead both educators and students?.
Biodiversity is a cornerstone subject in education, playing a fundamental role in achieving the Sustainable Development Goals (SDGs) outlined in the United Nations’ 2030 Agenda. Its importance stems from being the very foundation of life on Earth and its intricate connection with social, economic, and environmental sustainability. Integrating biodiversity education into curricula has been shown to foster environmental awareness from a young age, helping students understand the interdependence of ecosystems and the profound consequences of human actions on the environment. Environmental education is considered essential for cultivating responsible global citizens committed to sustainability. Given that biodiversity loss is recognized as one of the most pressing environmental challenges of the 21st century, education in this domain equips students with the knowledge and skills needed to address critical issues such as climate change, deforestation, and species extinction.
Biodiversity is directly linked to several SDGs:
SDG 13 (Climate Action): Biodiverse ecosystems, like forests and oceans, act as carbon sinks, helping to mitigate climate change. Education is key to promoting their conservation to reduce greenhouse gas emissions.
SDG 14 (Life Below Water) and SDG 15 (Life on Land): These SDGs explicitly underscore the need for biodiversity conservation and the mitigation of desertification.
SDG 4 (Quality Education): Biodiversity is an integral, cross-disciplinary theme that connects biology, geography, and social sciences, fostering a holistic educational approach aligned with sustainability principles.
In Spain, environmental education, particularly biodiversity, is a recurrent theme throughout the curriculum. In Early Childhood Education (up to 6 years), objectives focus on observation and respect for the environment. As students progress into Primary Education (ages 6–12), concepts become more intricate, emphasizing preserving the environment and adopting sustainable lifestyles, with biodiversity explicitly mentioned in legislation. Secondary education intensifies the focus on biodiversity and its eco-social relevance, particularly in subjects like "Biology and Geology". Finally, in the non-compulsory Baccalaureate stage (16–18 years), subjects like "Biology, Geology and Environmental Sciences" delve into sustainable development, biodiversity loss, and its causes and consequences. This background highlights why accurate AI information on biodiversity is particularly critical in Spain's educational system.
To assess the accuracy and reliability of AI-generated information in biodiversity education, the study adopted a straightforward approach. The core research question posed to the AI applications was: "Generate a list of 100 living species that are endangered". This specific number was chosen because higher counts would be too voluminous for educators to analyze, while lower counts would limit effective analysis, especially considering the vast number of real endangered species. The goal was to determine if AI responses aligned with scientific data or reflected societal, geographical, or preferential biases. Three prominent AI-based text generation tools were selected for the study: ChatGPT-4.5, DeepSeek-V3-chat, and Gemini 1.5 Flash, all accessed in February 2025. To ensure objectivity and avoid influencing the AI responses, three separate runs were performed for each AI bot using fresh accounts.
Biodiversity experts then categorized the species obtained from the AI tools into taxonomic groups (like mammals, plants, insects) and geographic regions. This categorization allowed for the analysis of taxonomic bias (uneven representation of different types of organisms) and geographic bias. The findings were expressed as percentages to facilitate comparison with real-world extinction risk percentages reported by the International Union for Conservation of Nature (IUCN) Red List, which served as the scientific gold standard. To minimize potential biases in the comparison, IUCN values were adjusted to account for discrepancies in assessment coverage across different taxa; for example, while 84% of mammal species have been evaluated by IUCN, only 1.2% of insect species have been. Geographic distribution data for species were also sourced from the IUCN, and eight clear geographical areas were defined for analysis. To quantify the consistency and magnitude of systematic errors in the AI results, two indicators were calculated: the Bias Ratio, which indicates the directional consistency of the bias (positive for over-representation, negative for under-representation), and the Magnitude Error Bias, which measures the mean of the absolute bias values. To further understand if the differences were merely due to sampling error, three random samplings were also performed on the IUCN database for comparison.
The results of the study revealed clear and consistent biases across all AI applications. Taxonomic Bias:
All three AI tools showed a marked over-representation of mammals and birds. For instance, ChatGPT listed 43.3% mammals, DeepSeek 53.1%, and Gemini 23.8%.
This contrasts sharply with IUCN data, where mammals represent a relatively tiny 0.27% of all endangered species, and birds only 0.23%.
This phenomenon is a reflection of "zoocentrism," a human tendency to prioritize animals over other living organisms, particularly charismatic species like mammals and birds, which influences conservation efforts and education.
Conversely, the study found a significant underrepresentation of plants, fungi, insects, and arachnids. For example, insects showed the highest negative taxonomic Bias Ratio for all AIs, indicating severe underrepresentation compared to their actual numbers in IUCN data. This is particularly concerning given that insects are the most diverse animal group and critical for human survival.
Plant blindness, or Plant Awareness Disparity, was evident, as plants (which constitute 29.6% of endangered species according to IUCN) were significantly underrepresented by all AI tools.
Fungi were almost entirely excluded from some AI lists, likely due to ambiguity in their classification and a general lack of recognition for their ecological importance.
Geographic Bias:
ChatGPT showed a notable overestimation of species from Europe (Bias Ratio 4.2), Asia (3.3), and North America (2.3), while underestimating Africa (-2.6) and Central America (-2.3). This is attributed to the historical and cultural influence of Western civilization, particularly North America, which is also the origin of ChatGPT's developing company.
Gemini exhibited the most extreme geographic bias, largely overestimating North America (Bias Ratio 13.3), Europe (1.8), and Oceania (0.9), while significantly underestimating Central America (-6.0), Africa (-3.3), and Asia (-1.4). Gemini, also of North American origin, showed similar societal influences.
DeepSeek had a narrower range of geographic bias but still displayed clear underrepresentation of Africa (-3.0) and South America (-3.1). DeepSeek's results were notably skewed in one run by repeatedly listing subspecies of a single North American fox (Urocyon littoralis), which the study notes as lacking practical relevance for a species list and likely due to the AI's lack of uniqueness mechanisms for long lists.
In terms of overall bias magnitude, DeepSeek was found to be the "most biased" AI application in the study, showing the highest Magnitude Error Ratio for both taxonomic and geographic perspectives. Gemini, while not globally the most biased, showed the widest ranges in its Bias Ratio values.
These findings clearly demonstrate that generative AI systems not only reflect but also amplify pre-existing biases present in their training data. This data, largely harvested from the internet, includes not only academic sources but also a vast amount of unverified, popular, or socially biased content. Consequently, there's a significant risk that AI models will not only reproduce stereotypes, inequalities, and taxonomic/geographic biases but also reinforce them through a cumulative effect each time similar content is generated. This is further compounded by a "user reinforcement" mechanism where popular or widely accepted answers are prioritized by the system, further solidifying dominant biases. Even user prompts can induce biased responses if they contain implicit assumptions or if the AI fails to recognize ambiguities.
The internal mechanisms and training processes of each AI system also contribute to differences in bias. ChatGPT, for instance, trains on massive amounts of unlabeled internet data before human trainers fine-tune it and score responses. This reliance on potentially "uncleaned" data and human judgment during fine-tuning increases the risk of biased or inaccurate outputs. While Gemini's architecture and use of Google-curated datasets aim for more precise responses, it still struggles with biases derived from human-conditioned training data. DeepSeek, despite its high performance in some tasks, exhibited the highest bias levels in this study. Understanding how these complex models are trained, how they prioritize responses, and the data they rely on is crucial for critically evaluating their use in educational, scientific, or social contexts.
The implications of these biases for biodiversity education are profound, especially for self-directed learning and for non-specialist teachers. Educators who rely on AI tools for classroom instruction may inadvertently reinforce existing biases, particularly the disproportionate emphasis on mammals, rather than helping to correct them. This perpetuation of skewed perspectives hinders the achievement of biodiversity-related Sustainable Development Goals (SDGs), particularly concerning the conservation of endangered species, by maintaining distorted geographical and taxonomic views. This issue is particularly salient in countries like Spain, where biodiversity is a consistent part of the educational curriculum. A particularly problematic scenario arises in Primary Education, where educators may not have specialized scientific training, making them more inclined to rely on AI tools for lesson preparation. While more complex content is addressed in Primary Education, such as ecosystem relationships and eco-social responsibility, teachers less familiar with these topics might find it challenging. Although the problem might be less acute in Early Childhood (simpler content) and Secondary/Baccalaureate stages (teachers generally have specific training), studies show an increasing inclination towards AI even at higher levels, and greater student autonomy could also perpetuate these biases.
In conclusion, the study unequivocally demonstrates that AI applications not only replicate general ethical biases but also significant taxonomic and geographical biases in their information about endangered species. This lack of accuracy can severely impede educational effectiveness and self-directed learning, exacerbating existing issues like zoocentrism and plant blindness. The prevalent cultural preference for animals, especially mammals, at the expense of ecologically vital organisms like plants, fungi, insects, and microorganisms, is alarmingly perpetuated by AI. This imbalance stands in stark contrast to scientific reality, as evidenced by IUCN data, leading to a distorted view of biodiversity in education and conservation efforts. Furthermore, the observed geographical bias is a clear indicator of the influence of sociocultural trends and the geographic origin of the AI developing companies, limiting a truly balanced global perspective.
The integration of AI into educational settings, particularly at primary school levels where educators may lack specialized scientific training, risks perpetuating rather than rectifying these harmful stereotypes. This directly hinders the achievement of the Sustainable Development Goals related to biodiversity and continues to foster a biased view of endangered species. To mitigate these biases, it is paramount to enhance the training of AI models using rigorous scientific data. Equally important is the promotion of critical teacher training that encourages the responsible and informed use of these powerful, yet imperfect, tools. Further research is essential to develop practical solutions that directly impact teachers at all educational levels, and to continuously evaluate how both existing and new AI models address and correct these persistent biases.
Biodiversity Researchers:
Dr. Tiara Moore: Dr. Moore is an environmental ecologist dedicated to understanding and protecting marine biodiversity. She is the founder of Black in Marine Science (BIMS), a nonprofit organization focused on celebrating Black marine scientists, raising environmental awareness, and inspiring future generations of scientific leaders. Her international research has focused on crucial areas like water quality, biogeochemistry, the microbiome, and coral reefs.
Dr. Rae Wynn-Grant: Dr. Wynn-Grant is a prominent wildlife ecologist and advocate for diversity and inclusion in the conservation community. She is co-host of the television series "Mutual of Omaha's Wild Kingdom Protecting the Wild" and is recognized for breaking barriers as one of the few Black women in her field. Her career embodies inspiring the next generation of diverse conservationists.
Dr. Mamie Parker: Dr. Parker is a trailblazing scientist who has significantly impacted the field of fisheries and wildlife conservation. She holds the distinction of being the first African American Regional Director of the U.S. Fish and Wildlife Service, where she also served as Chief of Staff and Assistant Director of Habitat Conservation/Head of Fisheries. Dr. Parker's career is a testament to overcoming barriers and promoting opportunities for women of color in conservation.
