I am specifically referring to records where AI processing has significantly altered key diagnostic features of the organism, as well as intentionally deceptive AI-generated images that are difficult to distinguish from genuine photographs. I have a growing concern regarding this issue: how should we address the contamination of natural history databases by AI-generated data? As identifiers, have you encountered observations that were obviously modified or entirely generated by AI?
(moved to General since this is directly about iNaturalist)
Oh I love that we can flag things for AI! I feel like the AI conversation is especially important in nature spaces and im so glad to see it being addressed like this.
I run my photos through Lightroom’s AI denoise because shooting at ISO 25600 has…significant noise. AI manupulation of images that affect identifying marks, etc are not allowed on iNaturalist.
I’ve not seen an AI generated image yet, but will flag it immediately if I do, so happy those flagging tools exist!
I generally do not post process except for cropping and adjusting levels for the odd occasion when I have ISO on 200 when it should be 100 or vice versa. I am too far behind to faff a lot with 500 to 1000 photos per day.
That being said, I’ve a feeling it is going to be unstoppable. Apple and Honor are working on generative AI functions on device that will allow a user to “get a photo that has been processed through a professional photography “brain,” with composition, lighting, and colors that better match her aesthetic preferences”; according to industry insiders this will include generative fill and crop before the image is even rendered. Hence, once the user’s AI settings on the phone has “learnt” what the user is focusing on most of the time, there is no way you can tell if there was post processing as even the user will not know what the camera has “seen” as opposed to what the user has seen. Aftermarket opportunity - make an AI tool that will tell us what is AI generated:)
So far, I haven’t found too many images that are clearly completely AI generated, and most do not seem to be done maliciously. They’re usually the result of someone, with good intent, using AI to “clean up” blurry photos, etc, and not understanding that it actually degrades the photo’s usefulness as evidence.
Most newer smartphones will have an tag in the metadata, which we can hopefully detect upon import and not accept. The problem, IMO, is messaging this to the user, since some things may be opt-out, some opt-in, some are not optional, and the whole thing is very confusing and will vary by device. And metadata, of course, can be erased.
I wouldn’t do that. This is not a photography website, the photos aren’t supposed to look good, but represent reality. I am not sure what “AI” actually means here but I would never let anything that isn’t a simple deterministic linear algebra touch photos that I am using as evidence for something in the real world.
“AI models are trained on millions of image pairs (a noisy image and its clean, original counterpart). Because of this training, the AI actually knows what real-world textures (like bird feathers or tree bark) are supposed to look like, allowing it to intelligently recreate the missing detail.”
I’m not sure if creating something that can’t be seen is such a good idea. Unless the recreated details aren’t creating things that don’t exist in reality. But how would one ever know for sure?
The photos my camera takes at 25600 ISO on RAW are so grainy that they do not resemble reality. My un-denoised RAW photos are also too grainy for the CV, even for photos at lower ISOs taken in broad daylight. IMO the denoise makes them look more like what our eyes actually see. I only set the slider midway because I am wary of the algorithm being overzealous and smoothing out important details.
Sometimes the photos are clearly AI-generated. In other cases, the phone’s image processing can introduce artifacts, even duplicating a body segment (there was an example, but the user later deleted the image). Live stacking can also create structures that were never actually present. These three situations can be surprisingly difficult to tell apart.
Noise in a photo is an electonic artifact related to the sensor’s inability to receive enough light. It could as easily be argued that posting noisy photos hides critical details and creates image data that doesn’t reflect real-world conditions.
It’s really worthwhile to understand what noise is and isn’t and what current noise models do and don’t do before lumping them in with AI slop. There are certainly issues that can be found if someone uses too much noise surpression, but in most cases it provides a photo that more accurately reflects real world conditions than does a noisy photo.
FWIW, AI-assisted denoising is currently accepted on iNat. Personally I use DXO Prime for some images and have found it works quite well without changing the underlying image. But yes, if it’s overaggressive it can create issues. I don’t use DXO’s DeepPrime, for example.
Where we tried to draw the line, at least on iNat, is between using tools that target the fundamental issues inherent in photography, such as noise, exposure issues, lens distortion, etc, and tools that “clean up” or meaningfully alter the actual subject and scene. So if someone uses denoise and it clearly is doing things like creating new feathers, or removing feathers, etc., then one should vote “no” for “Evidence accurately depicts scene”, which will make the observation casual grade.
I start feeling anxiety setting ISO over 1600!
If we’re strongly discouraging use of AI on the site, which we should be, maybe we don’t need an AI Summarise button on the forum?
I also disagree with using AI denoise to recreate ‘what you saw’. If you didn’t capture it in the photo, it isn’t hiding somewhere under the noise to be revealed, you just didn’t capture it.
I use that – it is interesting to see what is pulled out, and what is presented, what is misunderstood, where the ‘consensus lands’. Especially for the threads which have run for years, with hundreds, or even over a thousand comments. It takes a certain sort of dedication for someone newly landed, to avoid ‘I haven’t read all the comments …’ and proceed to repeat what has already been discussed to death many times.
(however for this thread the summary collapses before it finishes the last sentence ;~))
I don’t tend to use AI summaries, but I think there are some big distinctions here:
- the forum is run on software by Discourse.org, not iNaturalist
- also, using an AI model to summarize a text conversation, and presented as such, is quite different from a photo presented as real evidence of an encounter with an organism that people are meant to evaluate scientifically.
The two aren’t really comparable.
I have been concerned about the processing algorithms used by phone cameras, when users may not even realise that they have certain ‘filters’ turned on. A lot of photos that I see of people have obviously had lines removed from faces, as well as acne and sometimes hairs. Does the phone camera recognise when a leaf is meant to be hairy, or when there are subtle splotches of colour on a beetle, or are those features being smoothed out?