Opportunities

Exploring the opportunities for environmental and ecological research, many are related to enhancing our ability to communicate (Figure 1). First, ecologists use images and videos to communicate with other scientists. For example, they develop conceptual models that come to life in the form of visual representations. These depictions are very important, because they help other researchers understand complex relationships. Perhaps in the near future, the ecologist's workflow could be streamlined by creating such visuals using text prompts, saving time (and money) for other work. This way, ecologists could communicate their thoughts and ideas about ecological topics more effectively. Certainly, all of biology deals with complex topics, but in ecology visuals could for example be created for future scenarios of climate change, to depict concepts in theory-heavy fields such as population or community ecology, or to help explain experimental designs using tailor-made diagrams.

Second, ecologists also use images for communicating with the public, for example, policy makers or journalists, and in environmental education (Schäfer, 2023). Environmental scientists could likely use AI-generated images (or videos) to help them better explain the often complex environmental issues (such as pollution) and intricate ecological relationships, such as in biodiversity research. This could be done more cheaply, and more effectively, thus potentially reaching more people with visually appealing content. One powerful application could be the creation of visual what-if scenarios, illustrating risks for the environment (Luccioni et al., 2021). Imagine creating an image or video of your city in the year 2050 given the presence of certain global change factors, including climate change, invasive species, pollution, land use change and other drivers. Perhaps reporting on ecological research findings overall could benefit for exactly the same reasons, including reporting done by journalists or institutional press offices. Given the command line-based image creation process, images and movies can be easily adapted to different target groups with different backgrounds or different scopes. However, we should also be aware of and prevent potential negative impacts on scientific communication professionals (illustrators, artists, storytellers); in the near future, generative AI is unlikely to surpass their specific skill and professional input, which is essential for excellent communication.

Direct benefits for the environment could occur by replacing energy-intensive operations with AI-generated video. For example, traditional filmmaking has a high carbon footprint and could be partially supplemented or replaced with AI-generated content (careful carbon budgets would need to verify this assertion). This could make certain industries, such as entertainment, more environmentally sustainable.

Could generative AI for images also be useful in the scientific process of ecology? There are several opportunities that can be explored (Table 1): we identify image data augmentation and gap filling (Kebaili et al., 2023; Shivashankar & Miller, 2023), biodiversity monitoring and enhanced citizen science as potential areas where AI can be employed to enhance research outcomes.

Risks and dangers

There are direct environmental risks that are similar to those of text-based AI, in that AI-generated image and video will consume computing-related resources, leading to effects on energy and resource consumption and carbon emissions in addition to those of text-based models (Bender et al., 2021; Bird et al., 2023; Rillig et al., 2023), since they also make use of large language models as part of their model architecture. Energy needs arise during model training, during inference (that is, queries made to the model), and also for server infrastructure, and corresponding carbon emissions depend on the carbon intensity of the energy source. One estimate per query is 0.5 g CO₂e for a large language model (Vanderbauwhede, 2023). Luccioni et al. (2023) estimate the carbon cost of BLOOM (a 176-billion parameter language model) to be 50.5 t CO₂e over its life cycle.

However, we contend that indirect effects on the environment and the science of ecology are likely going to be of greatest concern (Figure 1). First and foremost, we see risks associated with spreading misinformation via deep fakes (Zhang et al., 2023). Imagine central figures, environmental activists, representatives of environmental NGOs, politicians or ecology experts, for example, being depicted in compromising settings that cast doubt on their integrity, defaming them; thus undermining the validity of the message they send. In addition, prominent communicators of science, or simply well-recognized people in public life, could be instrumentalized in spreading fake facts about the environment, and it will be very difficult to get these under control once they have been spread via social media. For example, imagine a politician, well-respected TV personalities and conservationists, or the president of an ecological society in a deepfaked video declaring that all problems of environmental pollution and climate change are well under control or that there is no biodiversity crisis to worry about. The damage to the credibility of the science of ecology caused by bad actors (governments, organizations, companies, individuals) generating such materials would be immense.

A threat to the integrity of ecology from within looms large as well. Image or video generation opens a door to rapidly and easily fabricating data; this is an already existing problem in science (Bik et al., 2016) but could be exacerbated immensely by the availability of advanced text-to-image tools (Gu et al., 2022). Perhaps images can also be manipulated to a point that it will be extremely difficult (or impossible) to detect. This will affect all areas of environmental science in which visual evidence plays a role: think of gel images, microscopic evidence, and really anything that can be used as supporting visual evidence in a scientific study on an environmental or ecological topic. Such deepfaked visual scientific data produced by bad actors would not only be a nuisance but might also undermine some of the trust in environmental science and ecology overall.

Risk mitigation

How could we effectively counteract some of these effects, for example protect reporting on ecological topics from fabricated, false information? How could we safeguard the integrity of our science? We will likely have an arms race between image generating AI tools and those that can detect their products; but improved detection will only solve part of the problem, because once images are posted on social media, effects will be very difficult to control. Banning these AI-tools will not be a realistic option as they are already widely adopted, and it is unclear how effective regulations would be when certain governments step in to regulate these products. We here propose local solutions, at the lab or institutional level, that can be rapidly implemented; however, we stress that in the long-term regulatory action on both the national and international level is needed.

To counter deepfaked visual information, we envision transitioning to radically open science as one potential option to safeguard ecological research and reporting. In the long term, multi-platform science communication in ecology has the potential to build public engagement and trust (Pavelle & Wilkinson, 2020) and might be less prone to fabricated visuals. Since such social media open science activities go beyond the open science policies already implemented in funding agencies (e.g. the European Research Council), these would need to be included in future funding schemes, providing adequate support to ecologists. Transparent reporting of the responsible use of AI-generated content in scientific work including detailed (and ideally machine-actionable) provenance information must be part of such an open science approach (Wahle et al., 2023). This would also safeguard gap-filling techniques (Table 1) from misuse via fabricated data.

Responsible use of generative AI also includes taking steps to protect others' careers with connection to environmental science, such as those of science communication experts. This includes carefully weighing the options when working on communication-related tasks, for example by asking how visualization could profit from professional input.

Another important step that can be taken by institutions or individual research groups is the development of policies on the fair use of generative AI (see here (Rillig, 2023) for an example). Such policies provide a starting point for larger-scale agreements in environmental science and they offer an opportunity to train scientists, early career researchers and more experienced colleagues, in these issues and involve them in the discussion during this critical time of transformation.