It seems like every other news story is about a breakthrough in the field of artificial intelligence. From writing code to cleaning up manuscripts, large language models—a type of artificial intelligence with the power to analyze and create language—have become everyday tools in people’s professional and personal lives.
Another type of artificial intelligence (AI), machine learning, has been used in ecological research for years. Machine learning trains computer systems to pick up on patterns in data, like identifying wildlife in camera trap imagery, and gets better over time. The popular citizen science app iNaturalist relies on machine learning to identify photos of organisms, from kingdom all the way to the species level, while the Cornell Lab of Ornithology’s eBird app uses similar technology to identify birds by their call.
Since released to the public in 2022 with the launch of OpenAI’s ChatGPT, language models (LLMs) have proven to be exponentially more powerful. Some academics think that the advent of this technology will transform ecological research and conservation science forever.
The Wildlife Society spoke with Ahimsa Campos-Arceiz and Christos Mammides, two authors of a new study, to dig into the potential applications of the technology in ecological research and conservation science.
Originally from Spain, Campos-Arceiz has been working in Asia for nearly 25 years and focuses his research on megafauna conservation and human-wildlife conflict. Christos Mammides, originally from Cyprus, studies protected areas and uses acoustic monitoring to understand more about the region’s bird biodiversity. They both work at the Xishuangbanna Tropical Botanical Garden of the Chinese Academy of Sciences in southern China near the borders of Laos and Myanmar. Their responses are edited for brevity and clarity.
What is an LLM and how long have they been around?
CM: LLMs are deep learning models that have been trained on large amounts of data so they can understand and generate human language. LLMs, as we know them, were released to the public in 2022 and then took off.
When most people think of LLMs, they usually think of the chat-based interfaces, like ChatGPT and Claude. These interfaces are built on top of the deep learning models that I just described and are fine-tuned to make their responses more useful and ensure that they comply with certain rules and guardrails.
How might LLMs be used across the field wildlife biology?
CM: Several of the pilot studies we analyzed connected LLMs to devices like camera traps and acoustic recorders. One exciting example used LLMs to study animal behavior, which is often too complicated to do with a simple machine learning model. What is really powerful about this approach is that you can oftentimes do this without having to retrain the model for a specific species. In general, I feel that the most impactful and innovative uses of LLMs will likely come from integrating them into wider workflows, rather than relying only on chat-based interfaces.
ACA: There are many examples where LLMs accomplish more than what human capacity can handle. I think that because of the use of LLMs in conservation science, there is going to be a huge increase in the amount of information we have access to and can process. We are going to be able to monitor natural processes and natural phenomena at a much finer resolution and in much larger volumes. I expect in the next five to 10 years, we will have a much better understanding of ecological processes.
What do you see as the major barriers to using LLMs in conservation science?
ACA: LLMs can hallucinate, which is what we call it when they use or create non-real data that looks very convincing. There are also biases based on the training databases, which means the models won’t perform well outside their area of expertise.
CM: The issue of the hallucinations is quite interesting to me. If you read arguments around LLMs, this is one that comes up often. But if you ask a colleague a question and they don’t get it right, it doesn’t mean you’re not going to ask them again. The errors are part of the process, and you just need to have the right workflows in place to make sure that those errors are not propagated.
If I ask a question about statistics to either a person or an LLM, I should double-check their answer. Ironically, one solution to the problem might be passing an LLM’s output to another LLM to help spot possible errors. If you have thousands of lines of code, for example, you can’t realistically manually check all of them.
What do you think are some of the most legitimate concerns about using LLMs?
CM: I’ve been following the debate about LLMs very closely to expose myself to people’s arguments against them. Recently, Nature published a series of interviews with researchers who don’t use LLMs, including some ecologists, and they made many valid arguments.
One thing I see as problematic is the potential for students to treat their work more passively, like reading much less than we did. You need to read extensively to develop critical thinking and have the right knowledge on a topic. We also need to develop some guidelines and university courses around the ethical use of LLMs.
ACA: I’m concerned about skill erosion. LLMs make things very easy in the short term, but as an ecologist, you still need to understand things from a very wide perspective. I also think that people might end up with common opinions on everything because LLMs will create an average value of almost anything.
Besides the practical concerns of a deskilled workforce, there are also ethical and environmental concerns of AI. How do you address those?
ACA: In terms of energy and water use, I think between the development of nuclear energy and more efficient models, the energetic cost of running the LLMs will decrease. But it’s definitely a factor that has consequences.
CM: There are also issues with data sovereignty and copyright. There needs to be more transparency in terms of how the models are trained and also mechanisms in place for giving the proper credit where it belongs.
What advice do you have for wildlife biologists today as they face the AI revolution?
ACA: There is a lot of fear and emotional reaction to AI. It’s important to be scared to some extent, because some amount of fear can bring caution. But I don’t want young people to be scared away from them entirely. I feel that all these concerns are normal when the tools change, especially when the change happens so quickly and the tools are so powerful.
LLMs, just like coding, are tools to get the answers we want—like knowing what happens when an elephant defecates here or picks up a branch there. We want to be ecologists, not coders.
There are risks associated with LLMs and with technology, in general, but the opportunities are huge. I’m very excited to see what projects my students are developing now. When I was in their career stage, we were much less ambitious about what we were able to address and the potential conservation benefits in our work.
Article by Olivia Milloway
