A rare bird turned up in central Brazil, or so it seemed. A photo on the wildlife platform iNaturalist showed a red-winged blackbird. That North American species had never turned up in that part of Brazil. It would have been a notable first.
It was not real. The bird in the original photo was an epaulet oriole, a common local species. The photographer had asked an AI tool to make the image “look better.” The software helpfully added features from a different bird.
The problem with ‘look better’
That case sits at the heart of a warning from researchers, published as a commentary in Nature Ecology & Evolution. They say generative AI is starting to pollute citizen-science records. Those crowd-sourced sightings tell scientists where species live and how they move.
There are two ways it happens. The rarer one is outright fakes, images generated from scratch and passed off as real. The more common one is subtler. A birder asks AI to remove a branch or sharpen a blurry shot. The model then rebuilds the bird, erasing the field marks that identify it.
The researchers say they have found several hundred suspect images already, spread across the Macaulay Library, iNaturalist and Brazil’s WikiAves. The true number is unknown, because many slip past unnoticed.
Why the records matter
These platforms are not just hobbyist galleries. iNaturalist alone holds more than 610 million images. Scientists mine them to track how wildlife responds to a warming climate.
“Regular people are posting information that a scientist could probably never get at scale,” said Tony Iwane of iNaturalist, a co-author on the paper. He called the network “almost like a sensor of what is happening on Earth in real time.” His catch: the information needs to be accurate.
Feed it bad data and the inferences go wrong. A single AI-invented sighting can suggest a species has shifted its range when it has not. There is a second cost, too. Manipulated images used to train AI identification tools can quietly degrade them.
Slop meets the wild
The outright hoaxes are usually easy to spot. “Nobody is falling for a toucan sighting in Siberia,” Dr Alexander Lees told The Guardian. The Manchester Metropolitan University ecologist led the paper, and says the quiet edits are the real danger.
Lees put it bluntly: a huge share of the wildlife photos he now sees on Facebook are simply AI-generated. It is the same AI slop swamping the rest of the web. Here, though, it corrupts a scientific record, not a feed. The same tools that recreate a goal that was never filmed can conjure a bird that was never there.
Fighting back
The platforms are starting to respond. iNaturalist now lets users flag images two ways. A fully-AI-generated flag hides the image. An over-manipulated flag drops it to “casual” grade, so it never reaches research databases.
So far only about 1,400 of its 610 million images carry an AI flag. Roughly 600 are marked as fully generated. That is either reassuring or a sign of how much is slipping through, depending on how you read it. Detection is hard, as Meta has found with its own AI image detector.
The researchers want stronger tools: image-authentication checks and metadata verification. Above all they want education, so birders learn that “improving” a photo can break the science. The wider lesson is familiar from every corner of the internet AI has touched. Once the fakes are good enough, trust becomes the scarce resource.
