Use of AI in Conservation
Source: AI generated images, for example only. Prompts from left to right: red fox in the British countryside, beaver building a dam, pine marten in a tree
Introduction
The Mammal Society recognises that the integration of artificial intelligence (AI) into wildlife conservation efforts presents both significant opportunities and substantial risks. While AI technologies can enhance data collection, analysis, and predictive modelling, their use must be carefully managed to prevent the dissemination of scientific inaccuracies. This statement outlines our position on the responsible use of AI in wildlife conservation, emphasising the critical importance of accuracy in AI-generated images and text.
Opportunities
1. Enhanced Monitoring & Data Collection
AI algorithms can perform data analysis tasks in a fraction of the time they would take for humans, and can utilise processing power and machine learning to increase accuracy and consistency.
AI can automate the identification of species in images, videos and raw sound recordings, increasing the efficiency of monitoring efforts and the potential for immediate responses to issues such as the incursion of invasive species. Examples include the use of AI to review ambient sound recording from acoustic sensors to alert scientists to the presence of certain species, the development of tools that allow surveyors to identify species and details about individual animals from their tracks, and bat call identification software.
AI algorithms are also used to analyse genetic data, sifting through massive amounts of data to identify patterns or mutations.
AI-powered analysis of drone footage can allow surveying of large and inaccessible areas, providing valuable data on wildlife populations and habitat conditions.
2. Predictive Modelling
AI can analyse vast amounts of data at great speed to predict trends in wildlife populations and the impacts of environmental changes, aiding in proactive conservation strategies.
Examples include the use of AI models to analyse various environmental variables (such as prey availability, human disturbance and vegetation cover) to predict suitable habitats for species translocations, and the analysis of genomic data to predict population viability and extinction risk.
Pivotally, in these instances AI is used to dramatically speed up the generation of outcome scenarios, but the parameters of the model, the predicted impacts of different factors, and the scenarios modelled, are provided by ecologists on the basis of research and real-world observation.
Risks
1. Scientific inaccuracies
Despite the incredible potential of AI - specifically convolutional neural networks (CNNs) - to facilitate large scale biodiversity monitoring, their use is not without inherent risks and critical limitations.
The key ubiquitous issue when it comes to monitoring wildlife with AI is the introduction of false positives into data, reporting a species was present in a sample, when the species was not present. False positives stem from classification error and are present in all applications of CNNs, even when every possible precaution is taken. Contemporary modelling methods used for animal monitoring (hierarchical models) explicitly account for false negatives (imperfect detection) but have the implicit assumption that false positives do not occur. Even small numbers of false positives can cause extreme bias in population estimates, which can be hugely detrimental in the context of monitoring rare or threatened species.
There is a suite of statistical models emerging that can account for the issues associated with data from automated classification using machine learning (false-positive models, e.g. coupled classification). Whenever AI is used in wildlife monitoring, users should take all steps to minimize the presence of false-positives through 1) setting high confidence thresholds for ID assignment to samples, 2) maximizing precision, over recall, 3) having CNN-human feedback loops and validation, 4) using appropriate statistical models that explicitly model classification error, and 5) finally, being aware and acknowledging the possible biases that stem from the use of AI for automated classification of monitoring data.
2. Misinformation and misrepresentation
Inaccurate AI-generated images or text can lead to the spread of misinformation, and undermine public understanding and support for conservation efforts.
Images
Even detailed prompts for AI image generators can result in inaccurate representations of animal appearance, behaviour or location, and interactions between species that are rare or impossible in the real world. As AI programmes become more sophisticated, the risk becomes compounded, since such generated images will become harder to distinguish from real photographs/videos.
In some cases the generation of AI images of wildlife may be for purely entertainment or illustrative purposes, and may be captioned to recognise their AI origin. However such images may then become divorced from context and picked up as an image reference intended to convey real scientific fact, or even used by other AI tools as a reference or in an analysis of trends in photos. Scientific inaccuracies can thus lead to others being made, with the false conclusion becoming harder and harder to trace back to the original source of the misinformation.
Misrepresentation of species, habitats, or conservation status can lead to inappropriate local or national actions, policy decisions and could lead to the misallocation of vital resources.
Text
An AI content generator can only provide information as accurate and reliable as that available in the public domain on which it draws. Much of this may be unverified or false, and unusual or disputed data may be given equal weighting as robust and peer-reviewed data.
As with images, once an AI content creator has generated text describing a conservation topic with inaccuracies or unverified observations, these may be picked up as sources for other AI outputs, multiplying and magnifying the inaccurate detail until it acquires false authority through prevalence in the public domain. Myths about species and inappropriate advice or guidance for those seeking to monitor or support wildlife may be perpetuated.
AI should therefore never be used as a first port of call to generate text and formulate conclusions for life science papers – even where the intention is to quickly produce a first draft and provide a structure, before human review and editing. Key arguments or conclusions with no firm basis or a misguided evidence base behind them may be carried through to the final paper, while pertinent findings from very new or old research may be overlooked because they are invisible to the algorithm. AI tools should be used for no more than improving language, grammar and structure of papers, and spellchecking human-authored text.
Recommendations for Responsible AI Use
1. Rigorous Validation and Verification
AI algorithms should be rigorously tested and validated against high-quality, scientifically verified data. It should always be possible to identify all sources of data used to power any algorithm used to guide or inform conservation.
Continuous verification processes should be established and maintained to ensure ongoing accuracy as AI models are updated and refined.
We also note that the use of AI, particularly training a model, is very resource intensive, which has led to a marked increase in Google’s carbon footprint. We therefore suggest that the costs of using AI is considered when deciding on its utility, as this cost is generally hidden from the user.
2. Transparency and Accountability
AI developers and conservationists should maintain transparency about the limitations and potential errors of AI systems.
Clear accountability mechanisms should be in place for addressing inaccuracies and mitigating their impacts.
3. Collaboration with Experts
Collaboration between AI developers and wildlife conservation experts is essential to ensure that AI tools are designed and implemented with a deep understanding of ecological complexities.
Ongoing dialogue between technologists and conservationists can help anticipate and address potential issues before they arise.
4. Public Education and Engagement
Efforts should be made to educate the public about the capabilities and limitations of AI in wildlife conservation, and to highlight where certain practices (such as generating images of wildlife and wildlife interactions) introduce risks to conservation outcomes.
Engaging the public in discussions about AI use can foster a more informed and supportive community.
AI-generated ‘photo realistic’ images of wildlife should not be created or shared by conservation organisations, including the Mammal Society, without clear and explicit context. Our communications and reports are seen as a reliable source of scientific fact by professional and public audiences, and may be used as source material by journalists and AI tools.
References
van Oosterhout, C. AI-informed conservation genomics. Heredity 132, 1–4 (2024). https://doi.org/10.1038/s41437-023-00666-x
Delplanque, A., Théau, J., Foucher, S., Serati, G., Durand, S., & Lejeune, P. (2024). Wildlife detection, counting and survey using satellite imagery: are we there yet? GIScience & Remote Sensing, 61(1). https://doi.org/10.1080/15481603.2024.2348863
Green, K.M., Virdee, M.K., Cubaynes, H.C., Aviles-Rivero, A.I., Fretwell, P.T., Gray, P.C., Johnston, D.W., Schönlieb, C.-B., Torres, L.G. and Jackson, J.A. (2023), Gray whale detection in satellite imagery using deep learning. Remote Sens Ecol Conserv, 9: 829-840. https://doi.org/10.1002/rse2.352
Adanma, Uwaga & Ogunbiyi, Emmanuel. (2024). Artificial intelligence in environmental conservation: evaluating cyber risks and opportunities for sustainable practices. Computer Science & IT Research Journal. 5. 1178-1209. 10.51594/csitrj.v5i5.1156.
Gu J, Wang X, Li C, Zhao J, Fu W, Liang G, Qiu J. AI-enabled image fraud in scientific publications. Patterns (N Y). 2022 Jul 8;3(7):100511. doi: 10.1016/j.patter.2022.100511. PMID: 35845832; PMCID: PMC9278510.