I have used clean up to blue someone’s face in the picture, for privacy reasons. I have another case where an insect seats in a paper with personal information on it (I had nothing else on me).
Technically speaking, are those observations against iNaturalist new guidelines?
No - you can see the reasoning and background for what should be flagged or marked here: https://www.inaturalist.org/blog/118284
I am not computer expert, but wouldn’t checking the users history be helpful. If they are suddenly putting all perfect photos of species on iNat, when their previous photos were the usual mix of good, bad and ugly, you would know there is a problem. The example photos are of a quality that I could never match myself. Yes, there are some great photographers on iNat, but even these people put up not so good photos because they are interested in downloading all the species they see.
If they always have photos of very difficult to find and photograph species that would also be evidence of a problem to me.
There are likely other behavioural based clues, timing clues, location clues etc. that might help.
Yes. And not just for the reasons you mentioned. If the quality increase coincides with the release of a new AI version, that should be suspect also.
Some of Google’s “generated” images now have synthID watermarks, which can be checked using Google (of course). This is a bit of a PITA for ID’ers and curators, but it does exist.
Sometimes the photo size or dimensions may also be a tip-off.
There is a YouTube channel that frequently discusses ways to tell apart AI things - the photos are harder, the videos are still a bit easier.
I think ai is just scrap everything related to scientific name to create new result which is no difference from real photos. This is why these generated results looks so detailed.
It’s very furious to watch that those intrusive generative ai is scrap & steal everything for crappy one-time pleasure, only to deceive people.
Hard to say whether it impacts data quality but there are a dozen pages of open AI flags and most of them seem legitimately AI. The first one I randomly clicked on looked like an isopod that just had aberrant segmentation that looked suspicious, but others have artificial-looking morphology or lighting or an AI symbol in the corner.
Most instances are going to be of common species where a false positive isn’t a big deal, and identifiers might only flag suspicious images because they take a second look when something is out of range. But as generation gets better, documentation of hoax rarities will be more of an issue…
It’s worth noting that the improvement in quality of AI-generated images is the result of them doing more direct plagiarism of our real ones. These [multiple expletives deleted] steal all our copyrighted material, resell it for billions of dollars, and get away with zero consquences.
Already a few actually, and from diverse categories of observers… I didn’t expect that many in few days.
The cheap ai high quality pictures might result in the end of quantity approach for observations (and observers) over quality. Which is a shame as the citizen science strength was in the quantity side mostly.
Perhaps it would be possible to add an “AI-generated” category to the DQA votes. That way, identifiers could take action themselves. Especially since it can be assumed that AI-generated content also affects comments. There are already 12 pages of reports here: iNaturalist Flags for Artificially Generated Content
Anyone can flag an image as artificially generated, which will hide it.
And there is a DQA metric specifically for images that have been altered. See this blog post.
Oh, thanks. I didn’t know that :)
All good!
That’s more than we want but a) I suspect some of those may be incorrectly flagged and b) the number of photos that have been flagged as copyright infringement dwarfs the number of AI flags for images. It’s still so much easier for someone to just steal an image and post it than it is to generate one.
This is a really engaging discussion, and I will definitely reply to some of the other great points raised here soon. For now, I want to focus on the potential impact these generations could have on observation quality and data integrity:
I completely agree with you on the numbers. 422 pages of copyright infringements versus 12 pages of AI flags over the same period is a stark contrast. However, the AI flag is still very new, and I suspect many users are not even aware that it exists yet. And importantly: at the moment we are mostly seeing and flagging the bad or obvious AI generations.
To test this for myself, I ran a small experiment using two of my own photographs as image-to-image references instead of text-to-image as before for Gemini to see whether the outputs could plausibly pass as real observations:
Gemini created completely new and unique images based on my photos and the prompt consisted of a single sentence: “Generate a new image based on this one.” It is nothing that takes a long time or requires any specific technical knowledge. This takes the issue beyond traditional copyright infringement. These are not 1:1 copies, meaning reverse image search would not detect them. More importantly, the AI did not just recreate the macro image convincingly; it also replicated the blurry, imperfect aesthetic of the spider photograph, including realistic sensor noise and lack of sharpness. In other words, it successfully reproduced “mediocrity” as well as technical quality.
I remember you mentioning earlier that AI “enhancements” (for example generative object removal or de-blurring) may currently be a more practical concern than fully generated images, partly because there is less incentive to fabricate observations from scratch. I understand that argument, and I share the concern about heavily processed observations. However, image-to-image generation increasingly blurs the line between “enhancement” and full fabrication.
It also creates a new incentive structure: users could generate entirely novel yet highly plausible observations with minimal effort simply to inflate statistics or participation metrics. At that point, the issue is no longer just aesthetic manipulation - it becomes the quiet creation of fictional data points.
I attached all images below (my original photographs and the AI generations). Looking at them side by side, can we confidently distinguish the real organisms from the generated ones? And importantly: in this example everyone already knows that two of them are fake. Under normal circumstances, users would not have that warning - and many identifiers confirm familiar species directly through the Identify module without ever opening the full observation to take a closer look (which wouldn’t make much of a difference here anyway).
The larger concern is not necessarily what is easiest to upload, but that many users - including experienced identifiers - may no longer be able to reliably distinguish sophisticated AI generations from genuine photographs, especially when the generated image intentionally mimics realistic imperfections. At the same time, iNaturalist’s identification workflow is built around community agreement, and many users understandably rely on existing IDs rather than independently verifying every observation in detail. Highly plausible AI images could therefore pass through the normal consensus process and achieve “Research Grade”.
As AI tools become more accessible and integrated into everyday workflows, the number of undetected generated observations entering the dataset may increase substantially.
A stolen copyrighted image is still at least a photograph of a real organism, even if uploaded improperly or attributed to the wrong person. A convincing AI-generated image, by contrast, may introduce entirely fictional biological records into the database while remaining effectively invisible to normal review processes.
I do not pretend to have an easy solution here. The technology is advancing extremely quickly, and detection is becoming increasingly difficult. But I do think it is important to acknowledge how difficult this problem may become precisely because the fabricated observations are often so plausible.
Out of curiosity, I am considering setting up a small external test among my own Instagram followers to see how reliably people can distinguish the generated images from the real ones.
Perhaps staff members - or a small group of users in coordination with staff - could eventually run a controlled experiment on the platform itself. By using clearly designated test accounts and uploading a limited number of generated observations under supervision, it might be possible to measure how often such content is actually detected and flagged versus how often it passes unnoticed into “Research Grade”. That could help determine whether this is likely to remain a niche issue or become a more significant long-term concern.
I agree that AI images should not be uploaded to iNaturalist, and checking across multiple photos is a neat idea for detecting them. However, the environmental impact of generative AI isn’t a serious issue, that’s something that has routinely been exaggerated. See https://www2.datainnovation.org/2024-ai-energy-use.pdf and https://www.sciencedaily.com/releases/2025/12/251205054736.htm
I agree with this, it would tell us just how big a problem we have; but it may be hard to implement and coordinate.
In my mind, this is another to add to the list of “possible sources of error in the data that can’t be entirely corrected”. I don’t think there’s any question that every one of us could easily be fooled into thinking an AI image is real if we weren’t primed with “spot the AI image” first. This can be added to the list of “uncatchable” errors like
Of course, sometimes these are catchable if the error is egregious enough (observation is out of range, photo shows wrong season, the friend also posts the photo), as are AI images.
And as for stealing images from elsewhere on the internet, again these can be detected and called out using reverse image searches, but how often do we actually reverse image search something to check its validity? Usually if you’re doing that it’s because something else about the observation tipped you off that something was wrong; I’m sure there are many more “unflagged” stolen images that were used in a way that didn’t raise any red flags to the person reviewing the observation.
So my point is that this isn’t something completely new. The data already has many incorrect points in it, as does any dataset, and anyone using a dataset responsibly is aware of this. This is just another source of errors to add to the list. We should absolutely be flagging these and trying to remove them, but I don’t see it as the biggest issue the platform faces right now.
That spider, by the way, is pretty impressive; I had always though it could only make ‘perfect’ images. The motion blurring is pretty effective.
Yes - it has broken the picture ‘too good to be genuine’ I relied on. Stolen from a professional photographer ? Or faked up ?
