Biased AI models undercut inclusive biodiversity learning

A new study reveals that even the most sophisticated AI tools demonstrate significant bias when providing information about endangered species, biases that could shape how students perceive biodiversity on a global scale.

Published in Sustainability, the study “AI in Biodiversity Education: The Bias in Endangered Species Information and Its Implications” examines the responses of widely used large language models (LLMs), ChatGPT-4.5, DeepSeek-V3, and Gemini, when asked to list endangered species. The findings point to troubling trends of taxonomic and geographic bias, raising concerns about the pedagogical reliability of AI in environmental and sustainability education.

How do AI models misrepresent global biodiversity?

According to the study, AI models consistently favor a narrow range of species and geographies when asked to produce lists of endangered animals. In most instances, these systems prioritize well-known, charismatic megafauna, such as lions, elephants, and pandas, while ignoring or under-representing species that are equally or more critically endangered, particularly those from under-digitized regions.

This pattern of selection reflects a taxonomic bias, wherein vertebrates, especially mammals, dominate AI responses, while other classes such as amphibians, reptiles, and invertebrates are largely omitted. From a geographic perspective, the models show a tendency to focus on animals native to Western or digitally prominent regions, such as North America and Europe. Meanwhile, biodiversity-rich areas in Southeast Asia, Africa, and South America receive minimal representation, despite harboring some of the world’s most threatened ecosystems.

The study attributes these distortions to training data imbalances within the language models. AI systems ingest vast amounts of text from the internet, but online content is skewed toward animals and habitats that receive more media coverage or conservation marketing. As a result, students and educators relying on these tools may receive a falsely simplified or partial image of what biodiversity loss actually looks like across the planet.

What are the educational implications of AI bias in ecology?

Beyond mere factual oversight, the presence of bias in AI-generated biodiversity content poses substantive risks to science education and sustainability awareness. The researchers warn that the simplification and distortion of species information could significantly affect how students conceptualize conservation priorities, ecological balance, and global interdependence.

First, the emphasis on large, photogenic animals may reinforce a narrow view of biodiversity, one that excludes lesser-known but ecologically vital species. For example, invertebrates, which comprise the majority of animal species on Earth and are critical to ecosystems, receive scant attention in AI-generated lists. This could hinder educational efforts aimed at explaining the complexity and interconnectedness of ecosystems.

Second, the omission of species from developing regions risks perpetuating geographic bias in conservation awareness. If AI models consistently under-represent species from the Global South, learners may develop an implicit bias that endangered species are mostly a Western concern. This undermines efforts to promote international conservation solidarity and equitable ecological engagement.

Third, these issues compound when AI tools are deployed in primary and secondary classrooms, where students may not yet have the critical skills to question or verify AI outputs. Without teacher-led guidance or supplemental materials, learners may internalize inaccurate hierarchies of species importance, weakening the goals of biodiversity education embedded in curricula aligned with the UN’s Sustainable Development Goals (SDGs).

Can AI be corrected to support better conservation learning?

The study does not argue for the rejection of AI tools in education but urges a serious reevaluation of how they are designed, trained, and integrated into learning environments. The authors advocate for a multi-pronged response involving educators, policymakers, AI developers, and biodiversity experts.

One of the primary recommendations is for AI developers to expand their training datasets by incorporating underrepresented scientific databases, ecological registries, and culturally diverse sources. This would reduce the dominance of media-amplified species and help ensure that a broader spectrum of biodiversity is represented in AI outputs.

Educators are also encouraged to apply a critical lens when using AI-generated content in the classroom. Teachers should guide students in examining the limitations of AI, comparing automated lists with authoritative sources such as the IUCN Red List, and promoting active discussions about species that may be missing or misrepresented.

Moreover, the study highlights the importance of collaboration between AI designers and ecological educators to develop educational tools that are not only technologically advanced but pedagogically sound. Rather than replacing human expertise, AI should serve as a supplement, an interactive layer that enriches biodiversity learning when deployed with proper oversight.

The researchers call for stronger ethical guidelines and educational policies that govern the use of AI in teaching natural sciences. These frameworks should aim to promote digital inclusion, epistemic diversity, and environmental justice by ensuring that AI tools reflect the full scope of the planet’s ecological reality.