A deluge of AI-enhanced bird photographs flooding popular forums and platforms is casting a shadow over painstaking scientific research, experts warn. The trend, fuelled by enthusiasts eager to boost their photo portfolios, risks creating spurious records that could taint crucial data used to map species habitats and monitor responses to climate change.
Citizens science projects, including those on iNaturalist and the Macaulay Library, are essential for gathering vast amounts of information on wildlife distribution. For many birdwatchers, capturing a species outside its typical range is a notable achievement – as seen in June's high-profile sighting of a Western Reef Heron in North Wales. However, the proliferation of AI tools has created an environment where users can rapidly produce high-quality fake images or subtly manipulate existing ones, introducing inaccuracies that can have far-reaching consequences.
Dr Alexander Lees, an ecologist at Manchester Metropolitan University and co-author of a Nature commentary, highlighted the scale of the issue. He noted that many wildlife photos shared on social media now appear to be AI-generated, making it increasingly difficult to use them for understanding long-term species distribution patterns. While outright hoaxes are rare, the danger lies in AI algorithms introducing features from different bird species during enhancement, resulting in credible but false records.
Dr Lees pointed to a reported sighting of a red-winged blackbird in central Brazil – a species typically found in North America – as an example. The bird was later identified as a common epaulet oriole after the photographer requested AI to "enhance" their image, resulting in features from the red-winged blackbird being added and creating the erroneous record.
Organisations managing citizen science platforms are still assessing the full scope of the issue. iNaturalist's director of community support, Tony Iwane, confirmed that while only a small fraction of the platform's 610 million images have been flagged for AI use, the actual number of undetected alterations remains unknown.