Anecdotally I’ve been noticing an uptick in people posting photos that they’ve asked AI to enhance or improve. For example, maybe they took a blurry photo of an ant or bird and don’t think it’s good enough to post to iNat, so they’ll ask Gemini or another model to enhance or sharpen the photo before posting it.
Unfortunately, these AI models do more than just add or subtract, or just add a little sharpening, they basically create entirely new images with a lot of changes. I just added a new FAQ that curators (and others) can point people to, if they come across this: https://help.inaturalist.org/support/solutions/articles/151000235408
Let’s keep in mind that the vast majority of people doing this are not acting in a malicious manner, they’re just wanting to improve their photos for iNat, and aren’t aware of the issues involved, so please be polite and explain to them why it’s not a good fit for iNat, and that imperfect photos are OK. Hopefully this FAQ can help with that. Thanks!
The description of the AI generated pigeon example in the FAQ is very charitable: it isn’t remotely like the original photo! Hard to quickly tell that it was AI generated though.
In my experience a lot of these photos just look too perfect. Like the feathers on that pigeon are perfect if you look closely. Another clue is that the image lacks metadata if you go to the photo’s page. Not conclusive, but those are both signals.
And you can download images and check them for SynthIDs, which are watermarks embedded in the images by the models that persist even if the image is resized. https://openai.com/research/verify/ will check for SynthIDs from OpenAI, and here are instructions for checking Google’s SynthID.
Also, I’d like to please keep this topic focused on messaging users, and perhaps how to spot these kinds of photos. I don’t want to turn this discussion into complaints about AI, etc. This is about an issue we’re seeing, and how to handle it.
Different groups will have very different experiences. Wing veins are a good tell if you know your group.
Ultimately though, i think the tech will continue improving with the only realistic option being tech to combat it. Like a built in detector.
I think we’ve already reached a point even something like a chiro can be generated from an AI with enough realism to fool anybody that doesnt know the morphology of them well. There are probably beetles, birds, lizards, etc I can no longer tell. Granted it depends on the quality of the AI generator.
You can’t teach everyone every group. The average person, even most long time iNatters aren’t going to be able to say, "that’s an AI chiro because there are wing veins present that don’t exist, or something like “there’s two M-Cu veins.” No chiro on earth has that.
There is also the issue of phone cameras embedding AI in the photo processing. Most users don’t know that the phone is automatically “processing” the photos at all. For small inverts I’ve had my phone totally change the leg stripes and other markings to “enhance” the photo without my permission, and it’s an opt-out feature rather than an opt-in on most smartphones nowadays. I’m not sure how to address this beyond asking users whether the markings in a suspicious photo match what they saw in real life, and then directing them to a how-to disable this feature?
Adding to this. I truly am not trying to be negative, but i just don’t know how to realistically solve this. I’m pretty sure I can’t spot good AI for taxa I’m not very familiar with.
This is where we are at, fake images that can even have the dorsocentral hairs between the vittae on the thorax of a chiro.
“Generate a dorsal view image of a male Chironomidae” Gemini 3.1 Pro
Really the only main tell is the wing veins are completely wrong for Chironomidae, but you can’t just expect people to have a book full of wing veins for every insect. I don’t know the wing veins of a Sawfly off the top of my head.
“Generate an image of the hypopygium of a Chironomidae” Gemini 3.1 Pro
Is it all wrong, just about, but 95% or more of people won’t know that because they are not familiar with insect genitalia. Even THE scale bar is correct. That is about the size of most chiro genitalia.
We really are getting to the point where we need experts or people with 100s, 1000s, of hours of experience with a group of organisms, to be able to tell whats real, fake, or has been altered.
while a help article is a start, it seems like this sort of guidance should be actively pushed out, especially to folks using the apps to make observations, to ambassador guides, etc., if that’s not already in progress. getting in front of the issue is going to be easier than troubleshooting it on the back end.
Unfortunately, these often get a lot of false positives as they can misidentify regular editing as AI, or mistake very mild AI use to do things like light denoising of the sort that something like the Topaz Denoise does or that some of the inbuilt features of recent Photoshop releases do as full AI manipulation.
There is a really large asymmetry between the current capabilities of AI and the ability to detect it.
@riainnature Maria, this is another important topic. Would you please start a different topic on the forum and explain in more detail how your photo app is automatically using AI enhancement? It will be important to include type of phone/camera in use (Android/iPhone/other), software version, photo app in use, etc. Thanks!
Is only the company that produced the SynthID able to identify them? If so that’s unfortunate and rather confusing to me.
I saw a friend complaining recently on Instagram that a photo from her iPhone that she’d edited in Photoshop had been flagged by Instagram as AI content. I didn’t see the image and have no idea what editing she’d done but I wishfully thought maybe Photoshop (or the iPhone) was doing some AI enhancement at the level this thread is about that would also have a SynthID. Guess not.
That’s my understanding, yes. Although I’m happy to be corrected.
Here’s what OpenAI says:
The tool is designed to detect content generated with ChatGPT, the OpenAI API, or Codex. It currently supports images and audio files. Other content can be uploaded, but OpenAI provenance signals will only be detected if it was generated with our tools.
And Google:
While other companies have started to adopt SynthID watermarks, Gemini can currently only recognize content created by Google AI tools.
Don’t quote me, but I swear I saw a discussion concluding that intentionally using generative AI to alter features isn’t allowed, while basic ai denoisers are generally fine since they just reduce grain. I’ll have to find the comments about this but later.
It puts smartphone computational photography in a weird grey area. Since we often can’t disable the aggressive sharpening and upscaling on our phones in certain situations, it seems we’re treating this unavoidable built-in processing differently than someone deliberately feeding a photo into an external ai tool to ‘fix’ it.
I’m wondering about the use of the word ‘accurate’ as in “Observations should accurately reflect the organism and the scene”. I imagine some might argue that an enhanced image is a more accurate reflection of the organism in that it shows what was not visible or evident at the time of the observation but is clearly understood to be a distinctive feature of that organism.
The follow-on phrasing is more specific in terms of defining what ‘accurate’ refers to: “[observations] should include evidence of the actual organism (or trace of an organism) at the time of the observation, observed by the user who is uploading the observation”.
This narrows things down considerably but probably makes ANY alterations after the fact (at any time after the observation) impermissible.
As some people in this, and similar discussions, have suggested, the accuracy of AI-generated images is rapidly improving (especially since everyone keeps testing and refining the possibilities, even those who are anti-AI) so it seems an appropriate word to use at this stage when many errors and mistakes are being made.
But I’m wondering if emphasis should/can be placed on the scope of the word ‘accuracy’ as it pertains specifically to iNat - ie. an accurate reflection of the relationship between observer, environment and organism in real-time. Not just on what is observed but when and how it was observed. It is changing that relationship that makes it “unacceptable on iNat”.
Yes! It actually came up in an Inat plant walk I did just this weekend - after we all tried taking photos of a plant, one participant commented: how about using AI to make the photos better? (of course, I explained why this is not a good idea). Just to reflect that AI use is getting normalized and it’s happening fast.
I saw an AI (Nano Banana, IIRC) image of cowslip (Primula veris) that was good enough to fool me, and that probably counts as a taxon I’m familiar with.
Probably better, and easier, and cheaper, to rely on volunteers to hunt and deal with possible forgeries? As long as “photographic improvements” do not preclude engaging with nature, I can’t imagine the platform bothering en masse its userbase out of a concern with data quality.
The resulting image is actually just a crude mash-up of a genuine photo of Columba livida set against a similar background. The only elements of the original image that remain are some of the electicity cables! This seems more like an example of copyright infringement, than AI-meddling per se.
From an iNaturalist perspective, examples like this aren’t really in the same category as generated enhancements, since it isn’t necessary to use AI to create them. The tools for manufacturing images like this have been around for decades. AI has just automated the process so that it’s no longer necessary to learn how to use photo-editing software to achieve the same results manually.
I understand that it’s important to have an unambiguous example in the FAQ, but it would be helpful to also provide some examples showing more subtle enhancements of the actual original subject, rather than what is effectively just an arbitrary crop and paste.
Many identifiers utterly loathe AI and actively avoid having anything to do with it, so they will usually be much less familiar with what AI is truly capable of (and also what it’s not). Perhaps we need an anonymized rogues gallery of genuine examples found “in the wild” on iNaturalist so people can start to learn how to identify them. At the end of the day, it’s just a bunch of pixels, so there should be no material difference between, say, counting what appear to be bristles on a fly’s leg and spotting the subtleties of automated photo-enhancements.
I think people who are happy to use AI will just have to get used to being challenged more regularly. Ignorance is no excuse. Having said that, though, iNaturalist isn’t really in a position to complain too much either, since it’s actively involved in promoting tools that use exactly the same AI technology. If it turns out that the CV (along with the collective efforts of the identifiers that train it) can’t reliably identify AI generated fakes, what does that say about some of the fundamental assumptions the site is founded on? Pop will eat itself, maybe…
