Dear Birders,
Over the past year, several birders in California and further abroad have used artificial intelligence (“AI”) to edit photographs of birds and have uploaded those altered photos to eBird and iNaturalist. Even though most of these instances have probably been innocuous attempts to make a poor photo look a bit better, this practice is highly damaging regardless of a user’s intended purpose. Unfortunately, most birders still do not understand how many of these programs work and why they are so negative. Below I have attempted to explain this issue.
There are several types of artificial intelligence, but the variety used by the most popular platforms (Google Gemini, Claude, etc.) is called generative AI. These programs have immense databases which they use to generate responses to user prompts. To put it simply, if you have a photo of a California Scrub-Jay, and you use a generative artificial intelligence program to edit the photo, the program is going outside of the data contained in the photo, searching its database for what it knows California Scrub-Jays look like, and then uses that external knowledge to “touch up” the photo to meet its expectations. This is not much different than printing a photo out onto paper and asking an artist to touch up the photo by painting onto it. The human artist will use their knowledge of what scrub-jays are supposed to look like to improve the image. So does AI.
To further clarify, I asked Google Gemini to explain how it edits photos. Here is what it told me: “Gemini is a Generator: It doesn’t actually “edit” your photo. Instead, it looks at your image, reads your instructions, and draws a brand-new picture from scratch based on patterns it knows. Because it rebuilds the entire image every time, it struggles to keep the original details—like the exact shape of a face, text, or background textures—identical.”
So, if you upload that Scrub-Jay, the generative AI can use its knowledge of scrub-jays to fill in missing details, and make that blurry photo sharp, or that dark spot visible, etc. But what if you upload a subpar photo of an Empidonax flycatcher? Or a seagull (yeah, I said it)? The AI still goes into its database and uses that external knowledge to touch up the photo… can the blurry Western Flycatcher be mistakenly touched up into a Willow? A Yellow-footed Gull into a Western Gull? Even if you upload a good photo, can the Sooty Fox Sparrow be made to look a bit more like a Slate-colored? Might a hatch-year state of molt return looking a bit more adult like? (If the number of remiges can change on a bird when run through AI, as I’ve seen a few times, it seems anything is possible.) How about a perfectly decent photo of an Epaulet Oriole, which when “touched up” with AI came out looking like the first-for-Brazil record of Red-winged Blackbird? More on that last scenario, and a further explanation of AI’s problems as they pertain to citizen science databases, here:
https://www.nature.com/articles/s41559-026-03141-y
In short, generative AI should absolutely not be used when editing photos. These programs rely on pre-existing knowledge to edit photos, and in doing so, inevitably hallucinate details that were not actually there. At best, uploading these photos pollutes scientific databases with photos not representative of reality. (It is crazy that that is the “at best.”]) And, I hate to say it, but with companies shoving AI down our collective throats, the onus is now on the user to verify that new photo editing apps and tools do not use generative artificial intelligence.
One final sobering note: Most of these instances where eBird reviewers or other careful eyed observers have sniffed out AI-altered media have done so because, to the trained eye, most AI edited photos often have the wrong number of feathers, a mishmash of features from different (sub)species, or otherwise appear “strange” or “off.” But AI is improving rapidly. Just last year, I was unable to generate a photo of a bird in Google Gemini that did not seemingly combine features of multiple species and that was not positioned at an unnatural angle. Last week, I generated a photo of a Mourning Warbler with AI–completely from scratch, no pre-existing photo to serve as a “canvas” to work off of–that was scarily convincing. (I sent the photo to a few friends to ask what they thought– the first to reply wrote, “Ok f*** this.”) Soon, we will not be able to tell what is and is not altered.
Please, build good habits now, and encourage the same from those around you by strongly advocating against the use of generative artificial intelligence to edit photos and by educating other birders of its perils.
Thank you
Ethan Monk