AI photo editing is quietly polluting the bird sighting databases scientists actually rely on

[Tiger]

Ecologists are warning that AI enhanced and AI generated bird photographs are showing up on citizen science platforms like iNaturalist and the Macaulay Library, threatening the credibility of databases researchers use to track species ranges, migration patterns and the effects of climate change. In a commentary published in the journal Nature, Dr Alexander Lees, an ecologist at Manchester Metropolitan University, warned that hundreds of fake images have already been identified across popular species recording platforms, with the true scale of the problem still unknown since many altered images likely go undetected entirely

The more insidious version of this problem isn't outright hoaxes, as Lees put it, nobody falls for a toucan photographed in Siberia. The real damage comes from routine, well meaning photo editing, a birder asking an AI tool to remove a branch obscuring their shot or sharpen a blurry subject, and the model quietly reconstructing missing detail using features borrowed from a different, similar looking species entirely. Lees pointed to a concrete example, a reported sighting of a red-winged blackbird in central Brazil, a species normally found only in North America and never previously recorded in that part of South America. The bird was actually an epaulet oriole, a common local species, but the photographer had asked an AI tool to make the image look better, and the AI added blackbird-like features in the process, creating a false record for a species that was never actually there

On iNaturalist specifically, only around 1,400 of the platform's more than 610 million images have been flagged for AI use so far, a number that almost certainly understates the real scale given how hard subtle AI enhancement is to detect compared to an obvious full fabrication. Citizen science organizations are still working to figure out just how big the underlying problem actually is

The stakes go beyond birding forum credibility. These platforms feed directly into real scientific research on species range shifts and habitat use, meaning a polluted record doesn't just mislead an individual birder chasing a rare sighting, it can quietly distort the actual data conservation scientists rely on to make range mapping and conservation triage decisions, sometimes for years after the original fabricated or subtly altered image was ever seriously questioned

Scientists Warn Wildlife Photographers' AI-Enhanced Bird Photos Could Threaten Researchpetapixel.com

[RustyHawk]

Nobody falls for a toucan in Siberia is such a sharp way to frame the actual risk, the obvious fakes were never the real problem, it's the subtle, well meaning edits that are genuinely dangerous to data integrity

[BrokenDave72]

The red-winged blackbird in Brazil example is such a perfect illustration, a photographer asking an AI to just make it look better accidentally fabricated a genuine scientific anomaly that other researchers would have had to seriously investigate

[RusticDaemon]

Only 1,400 flagged images out of 610 million on iNaturalist almost certainly understates this by a wide margin, that's the detection rate for obvious cases, not an estimate of the true underlying scale

[Wendy88]

This is a really underrated example of AI slop causing real, tangible harm outside the usual social media and content farm context, this is actual scientific data getting quietly corrupted

[Daemon90]

The fact that most of these edits come from genuine enthusiasm rather than malicious intent is exactly what makes this so hard to police, there's no obvious motive trail to follow when someone just wanted a nicer looking photo

[Trinity49]

Wonder what kind of technical detection or platform level labeling requirement could actually catch this at scale, given how subtle a lot of these edits apparently are compared to a full fabrication

[TheRizz96]

This adds a new wrinkle to climate and conservation science specifically, on top of every other data quality challenge that field already deals with, a slowly accumulating layer of AI introduced noise in the historical record

[TaxSeason37]

The 1,400 flagged images are reassuring only if the detection system is finding most of the problem, and the wording suggests it is not. Obvious fabrications are the easy cases; subtle edits can pass as ordinary wildlife photography.

Automated detection should help triage, but metadata and community review may be more useful. A new account uploading hundreds of flawless species portraits from distant locations deserves a different level of scrutiny from a local birder posting one imperfect image after a walk.

Context is evidence too.

[Oscar_86]

Citizen science platforms depend on trust between observers and reviewers. People contribute because they believe their observations will help answer real questions, not because they want to populate a gallery with attractive fiction.

That trust can be protected without banning creative photography. Give contributors separate labels for documentary records, edited images, and generated artwork, then make the scientific dataset use only the categories that meet its standards.

The bird can still be admired; it just should not be counted as a field observation if it was assembled at a desk.

[BigDog26]

The best safeguard may be a chain of custody for the original file. Keep the unedited capture, preserve timestamps and location information when the contributor chooses to share them, and show whether the uploaded version has been transformed.

That will not catch every fake, especially if someone starts with generated content, but it gives reviewers more useful evidence than the final image alone. A database built for science should treat provenance as part of the observation, not as optional decoration.

[Gaz_23]

A birder once described a blurry photo as a receipt rather than a portrait. It may not be beautiful, but it proves that someone saw something at a particular place and time. A generated image can look like a perfect receipt for an observation that never happened.

Platforms should preserve the original file and record the editing history where possible. A visibly altered image can still be useful if it is labelled clearly; an unlabelled synthetic image is much harder to interpret.