BOULDER, Colo., July 21 — Generative adversarial networks, or GANs, introduced by Ian Goodfellow in 2014, and the more recent diffusion models like Stable Diffusion, released in 2022, can create photorealistic images. But when these tools are used to “enhance” wildlife photos, they may combine features from multiple species. That’s because the models lack biological constraints.
The result can be a kind of digital chimera. An AI editor might, for example, superimpose the plumage of a western tanager onto the body of a warbler.
It might even add an extra leg. These fabricated images can pass an initial visual inspection by someone who isn’t an expert. The problem is made worse by AI upscalers that introduce hallucinated details — details that never existed in the original photograph.
How AI Creates Wildlife Chimeras
In 2023, researchers at the University of Colorado Boulder demonstrated just how easily this can happen. They trained diffusion models on images from iNaturalist, a popular citizen-science platform where people upload photos of plants and animals. The models could then generate plausible but completely fake species photos.
These synthetic images look convincingly real to the untrained eye. iNaturalist relies on a computer vision model to suggest identifications for uploaded photos. The early findings suggest that if a fake image gets uploaded, it can confuse that model.
That could lead to a misidentification, and the error could spread through the entire system as others rely on the platform’s suggestions. The encouraging part is that researchers are now aware of these vulnerabilities. But the phenomenon highlights a core weakness in artificial intelligence: when AI systems are trained on data that itself was generated by other AI systems, the results can suffer from what researchers call model collapse.
The Risk of Model Collapse
Model collapse is a term that describes how data diversity shrinks when AI trains on AI-generated content. Instead of capturing the full range of variation in the real world, the system latches onto a narrower set of features. Over time, the output becomes less diverse and more artificial.
That reduction in data diversity is exactly where the trouble starts for platforms like iNaturalist, which depend on genuine, varied observations to train their identification models. None of this means that AI is not useful for wildlife photography or conservation.
But it does mean that users and platform operators need to be careful. The combination of GANs and diffusion models — both already widely available — can produce images that look authentic but are biologically impossible. The problem compounds when those images are fed back into machine learning systems as if they were real data.
For now, the best guard against these chimeras is human expertise. Researchers recommend that anyone who uses AI tools to enhance wildlife images check the results against known biological features.
Platforms that rely on computer vision for identification should also build in safeguards to detect AI-generated content. The encouraging part is that the scientific community is actively studying these weaknesses, and the same tools that can produce fakes may also be used to detect them. If you use a citizen-science app like iNaturalist, it is worth staying aware of these developments.
Talk to your local naturalist group or a trusted expert if a photo looks off. The more we understand about how these models work — and where they can go wrong — the better prepared we will be to keep our natural history records accurate.
