The potential impacts of generative AI images on public perceptions of forests
Without a doubt, generative AI has captured global attention. ChatGPT, the most well-known generative AI tool, famously gained 1 million users within five days of its launch in 2022. Today, it receives over 130 million visits every month, ranking as one of the most visited sites worldwide.
What is particularly striking is the fast adoption of this new technology among younger users. In the UK, nearly 80% of young people between the ages of 13 and 17 report regular use of such tools. In the Netherlands, the same applies to 74% of those aged 16 and 24.
As AI capabilities expand, particularly in generating and editing images, such technologies are increasingly integrated into daily life. These types of images are now prevalent on social media and even appear in political campaigns and NGO reports. Needless to say, as AI-generated content starts to permeate our everyday routines, increased AI exposure could potentially influence how people perceive, relate to, and make sense of environmental issues and sites, including forests and the people who live and work in them.
Experimenting with generative AI images of forests
With twelve early-career researchers, journalists, NGO representatives and communication experts, we did a little experiment to test how different generative image AI models portray forest-related concepts. Working in groups, we crafted broad, open-ended prompts to run on generative AI image tools like ‘impacts of climate change in the Amazon rainforest’ without detailing what the image should depict, leaving the interpretation to the AI. The goal was to assess whether the AI-generated images reflected reductionist interpretations, stereotypes, and biases around complex, multi-layered forest themes.
Each group ran their prompts through three AI image tools: Midjourney, DALL-E 3, Stable Diffusion, Adobe Firefly, and DeepAI. The groups then collaboratively analysed the images, focusing on what each tool’s output portrayed. We also presented a selection of images at the EFI Annual Conference, where up to 100 senior scientists, researchers, and other representatives from EFI member organisations participated in further interpretation.
Entanglement of algorithmic issues and human bias
So what did the AI-generated images show when prompted with generic keywords like “a photo of foresters in the Amazon forest”? Below, you see the images from DALL-E 3 and Midjourney.
Showing these images to senior forest scientists sparked an intense discussion. Looking at the DALL-E 3 image on the left, many researchers pointed out that the foresters were portrayed as colonial, military-like figures drawn in a cartoonish style. The image depicted all-male characters marching through a flat, cleared pathway in a forest, wearing uniforms and carrying machetes. In contrast, the Midjourney image was more realistic and photography-like but similarly showed male figures in uniforms and helmets. Several scientists noted that both images lacked the structural density typical of tropical rainforests like the Amazon. One researcher said, “Why isn’t it dark? The Amazon forest is a dark place”, while another commented, “The images, especially the one on the left, look more like a national park in North America.”
When prompting ‘a photo of Indigenous people in the Amazon forest’ and ‘a photo of local people in the Amazon forest’, we ended up with very similar images depicting tribal people. These prompts were treated almost synonymously on both Deep AI and DALL-E 3. Yet, the ways each tool interpreted the prompts showed a significant difference. On the one hand, DALL-E 3 depicted people in colourful costumes and feathered headdresses, surrounded by birds and a large tree—images that appeared almost “staged”. On the other hand, Deep AI showed realistic pictures of people in a more natural environment, seemingly portraying their everyday life in the Amazon in a domestic setting. However, these images also came with imperfections, such as misplaced arms and missing hands, indicating algorithmic issues (see the top left image below).
How does generative AI interpret more open-ended prompts like ‘impacts of climate change in the Amazon rainforest’? Midjourney, for example, returned an image of a burning forest with flames and smoke rising into the sky, portraying the causal link between climate change and wildfires and resembling images circulated on social media during the 2019 Amazon rainforest fires (many of which were debunked as images of unrelated events from the past). In contrast, Stable Diffusion produced a confusing and unrealistic result, combining elements of drought and an overflowing river next to a heavily forested area. The AI’s interpretations diverged quite significantly when tasked with portraying a concept that could be interpreted in so many different ways.
Yet, these results highlight not only algorithmic limitations but also human biases. The incompleteness and lack of diversity in the data have been widely criticised as one of the main issues with today’s generative AI technology. But, as seen in the images of masculine foresters and tribal people in the Amazon, some of the biases prevalent in our society seem to be reflected more accurately by generative AI tools.
Generative AI as a tool for collaboration and participation
Our experiment demonstrated the potential of generative AI to bring together people from diverse professions and disciplines. As this new technology has rapidly gained attention across sectors, our exercise took participants on a shared journey to explore how different AI tools work, their possibilities, and the challenges they may pose to their professions and society. At the same time, the collective exploration and interpretation within a group of young researchers, journalists, and NGO representatives helped them get familiarised with different worldviews. Such experiments and exercises could be leveraged in public engagement activities and science literacy programmes, where scientists and forest stakeholders collaborate and interpret whether AI-generated images accurately portray ecosystems and societal interests and concerns and what else might be missing through workshops, focus groups, or pop-ups and open houses in the forests. These technologies could be potentially used to “imagine” how the forests we all care and are concerned about should evolve in the future.
This hackathon was collaboratively designed and developed with Gabriele Colombo and Anna Cattaneo from Density Design at Politecnico di Milano, based on the research protocol tested at the Digital Methods Summer School at the University of Amsterdam in 2023. The participatory approach and some of the insights gained from the testing have directly/indirectly contributed to the ongoing EFI projects, in particular CLEVER and TRANSFORMIT.
For more on the research approach applied to this Hackathon, you can check out:
