Google recently introduced their new version of Gemini Flash: 3.5 (with the Pro version following next week). Whenever a new version drops, I always check how well the AI performs at depicting species when only provided with their scientific names. The leap from 3.1 (the previous version) to now is absolutely massive. Yes, there are still species—especially rarer ones—where you can tell that something is off. But even for insects, it’s becoming increasingly difficult to distinguish them from real photographs. I am very curious to see what iNat can do to prevent the flood of AI images we are already witnessing on social media. Here are some examples:
Of course, these images can be further refined by feeding the AI additional settings. So please don’t confuse this with ‘you can tell by the lighting that this could never be real.’ This right here is the 1st result of entering nothing but the scientific name.
I guess there is something called the Coalition for Content Provenance and Authenticity, also known as C2PA. The idea is to incorporate information about provenance (that this is actually the organism that was observed) into the metadata. https://en.wikipedia.org/wiki/Content_Authenticity_Initiative
I guess C2PA has some issues, but I expect that image verification standards (whether on the blockchain, or through other methods) will become more common as ai becomes more embedded into every aspect of modern life.
Perhaps iNaturalist will require these types of verifications at some point.
Here’s the (Completely fictional) 'Black-spotted Red-breasted Quail, Cyrtonyx palustris’ that Microsoft’s Copilot made.
I do have to say, some of those insect pictures look extremely realistic. I would not have been able to tell the difference. The Vertebrates are somewhat easier to tell though.
At the same time an orange blurry UFO on a branch can be identified with confidence as an Eurasian red squirrel. If we want to cheat there’s no need that many details.
Curious to see how data veracity will be treated not only by citizen science but by any open data set in general.
The vertebrates do look suspiciously posed but I wouldn’t instantly think them to be fake if I saw them on iNat. I’ll have to start being less credulous.
Thing is, we’re rapidly approaching a time where we’re not going to be able to tell.
I reckon Staff is soon going to have to decide to either implement technological procedures/requirements to keep AI images off the site, or to just let anything ride.
My hope is that we see some technological procedures/requirements implemented, as there’s not much point to a lot of this endeavor if it’s impossible to tell reality from fiction.
It often goes too far in the other direction nowadays, where any real person who is attractive, or who is doing anything mildly impressive, is instantly accused of being photoshopped or ai.
“Guys, I’m real!”
I’m sure the same thing will be applied to any cool find on iNaturalist: “This must be ai!”
I agree that AI image generation is only going to improve and we should be alert to it. But aside from the “how”, I wonder about the “why”. I mean, why would someone upload AI-generated images to Inat? What is the incentive? For individuals, this platform is a way of recording their finds in nature. From what I’ve seen over here in the Hong Kong plant corner since I’ve been ID-ing, users fall into 3 categories: a) those exploring and recording plants they see, b) those treating inat as an AI plant ID app; c) students doing assignments/learning activities. Only in case c) there might be an incentive to upload AI generated images (deadlines, too much homework). Maybe there are different motivations in other corners of Inat. But I wonder if there would really be enough users uploading fake images for it to seriously impact data quality?
I have a feeling this will become a really big issue with students sooner than later. There’s already countless observations posted from student projects (where their instructors force them to contribute to iNat for points) which have stolen images or terribly incorrect metadata, blind agreements, etc. It’s exhausting to go through them, and since the students don’t actually care about what they’re submitting, they just want their credit, they rarely fix things. If it gets to a point where students can just upload AI images to get their points, they will. In my opinion instructors should not even be allowed to use iNat as schoolwork for students in the first place but that’s a different issue. Of course I think it’s great for them to learn about iNat for those who do care, but most of them simply do not, so they should not be forced to use it.
Yes, maybe this happens when nature-based education is poorly delivered - learning activities focusing just on knowledge rather than developing the affective dimension (connection to nature) - i.e. “actually caring” as you say. I’m getting off the original topic so I won’t elaborate, but I agree!
What the technology is capable of doing is one thing, but the risk of the platform actually being flooded by it is another. Who would deliberately try to disrupt iNaturalist at scale? If it’s just a handful of users, they can be flagged and dealt with, much like people who currently reuse online photos or take pictures of screens.
Personally, I see the value of this post as highlighting what GenAI is now capable of, rather than presenting anything close to a top-10 risk for the platform.
The bottom one came out much worse and looks nothing like the species i put in the prompt. But thats easy for someone who’s spent two years learning this group to say. How many average people will it fool?
Why is a why needed? I have subscribed to and go through all the dinosaur observations that are put on the site (a somewhat frustrating experience), and I constantly come across random, inexplicable observations. Just two days ago, someone uploaded a photo of a squirrel cuckoo IDed as an Archaeopteryx, and a while back their was a string of earthmoving equipment IDed as various dinosaurs and contemporary animals.
So I’m going to add three more to your list: d) trying to be funny by putting an AI obs on and assuming everyone will catch it (and then they don’t); e) they are just downright mean and take pleasure in confusing the poor sciency people (“iNat AI trolls”, if that makes sense); f) they do not know how to actually use iNat and dump there AI images here.
Then you’ll still get edge cases, ones that are just plain inexplicable (perhaps an observer with a single observation of an AI turtle).
I think one of the most sure-fire ways to tell a real observation is going to be multiple angles of the same organism. Ai doesn’t have the capability to have continuity like that across multiple images. With that said, this greatly saddens me. Over the past few years, photography has become one of my greatest passions and favorite ways of engaging with and sharing my love for the natural world. I hate to see it cheapened like this. The thing about photography that I love is that it is a way of capturing in a single frame the beauty of an entire creature, a whole life. Photos are impactful because they represent an image of something that’s meaningful. AI images are the very opposite. They are a cheap and hollow imitation of these beautiful creatures that compel us to photograph, document, study and preserve them in the first place. Even more, I hate the sick irony of these seemingly pretty nature photos being generated by a machine that is draining our natural resources and destroying our environment. It’s just deeply, deeply wrong, and it makes me very upset.
I’m generally much more concerned about AI “enhancements” of photos than of fully AI-generated images (although those will of course continue to get better and are a small problem, currently). There’s less of an incentive to post a fully AI-generated image to iNat than there is to “clean up” and “improve” a real image that will bring it into the realm of untrustworthy evidence. I think it’s already pretty accepted among wildlife photographers to use generative AI to remove leaves blocking the subject, for example, or to make an image less blurry, or to remove other distractions. The Pixel 10 also has 100x “Pro Res Zoom”, which can create artifacts, and I imagine more phones will be incorporating functionality like this.
