I think AI images should have their original non-altered version posted with them, without which AI images shouldn’t be present at all. AI happily invents things structurally even when just colourising. I imagine the use case for AI is enhancing colours in some way to make clearer, though I think that should always just be done with a standard level change in a graphics program.
Saving to JPEGs which probably most of us do has already altered the photo. It can create strange patterns and artifacts, and I think we’ve all seen observations like that. Whoever looks at the image has to use their judgement. I manually run most of my photos through software but limit alterations to crop, brightness, and contrast. I fill in areas of blank background when an image needs to be rotated to avoid distraction. Since dates and locations can also be altered, bogus observations can’t be guarded against. Hopefully people will have uploaded enough similar observations that any attempt to deceive will be easy enough to spot. Whether the deception is done manually or automatically really doesn’t matter.
iNaturalist I gather saves at a very high quality level, something quite a bit over 90%, I imagine artifacts might occur in very specific cases such as microscopic stripes or when you are straining at microscopic organelles, but generally I’ve never spotted it in plant photos.
For people’s home use I would recommend either just keeping the original or if sizing down (e.g. 75%) saving as JXL format (e.g. qual 85 compression 6; XNView will do it).
TECHNICAL: When keeping the original, the size can actually be made much smaller by running through jpegoptim.exe (cameras often write bloated jpegs) and if desired transcode losslessly and reversibly to JXL using cxl.exe.
Sharing this article here. iNat is mentioned and has a quote from tiwane too. Includes a link to the article in Nature.
came specifically here to post this article and see it already up (the discussion as well)
INaturalist recently had a job posting for Senior Manager of Community Moderation & Curation. I hope this new hire will help create policies about how curators and moderators handle AI generated images, identifications, observations.
Here’s the job listing.
https://app.beapplied.com/apply/wsr3uznk0l
Yep, all digital photos contain some level of AI image ‘editing’ whether it’s a default smartphone processing path woven into the camera system, or a high end pro camera model.
Technically, virtually all camera sensors capture tones and colours invisible to human eyes and squeeze them into the human range to optimize color realism (or a subjective ‘perceptual’ realism), tonal, and even geometric detail.
Most of the sensors now have integrated, optimized, software inside the sensor chips themselves. That’s why I give a gentle eyeroll when I come across those ‘purists’ who insist that they never use any AI processing, only SFTC (straight from the camera) files.
Myself, I prefer to manage the interpretation in post RAW processing, focusing on getting as much ID friendly stuff I can, without worrying too much about final finish and some light AI edge or detail noise, if it doesn’t ‘create’ stuff that just wasn’t there.
As for smartphone shots, I think that detail generation and color sat are probably the two most difficult traps for most people to control. I generally shut off as much of that stuff on a phone’s camera as possible, and use an app like ProShot to get even more shutoff type management.
If anyone is looking for a great photo tuning app that does NOT use AI content generation, I highly recommend Radiant Photo. Avoiding AI content was the whole focus of this app’s design. It’s become my favorite final step for iNat image work. Doesn’t add pixels but does fantastic shadow, color and detail tuning.
It’s available as a more limited phone app or a desktop version. I bought the phone app after trying the trial early last year. It’s about a fifth of the cost of the full-featured desktop version and it’s fast, intuitive and easy to use.
My digital camera is not high-end. It is expensive (for me), but it is a glorified point-and-shoot. It is not programmed with any generative AI in it.
It is specifically generative AI that “purists” object to. Because generative is fundamentally incapable of not making stuff up. A photo that has been edited with generative AI cannot be reliable evidence.
There are all sorts of advanced applications that are not generative AI, but which may be called “AI” (a term which in my opinion has become meaningless in any case). Those are fine, or at least arguable. Let’s not obscure the issue by lumping it all in together.
Generative vs interpretive AI. When that point and shoot stores to anything other than RAW, it has to interpret the scene and ‘generate’ based on built-in models, the data that you save to JPG.
When I first moved to a RAW workflow I took some of my very old RAW files (I didn’t save a lot in that format back then) from 10 year old cameras and was amazed at what I got from an uptodate RAW processor (DxO) that supported those old camera RAW formats. This convinced me to switch to a RAW editing workflow where I get to call (no pun intended) the shots of that interpretation that I previously trusted the cameras too. Biggest advantage by far was the ability to mine the data that saving the file to JPG dumnps to rescue good high ISO shots with the amazing noise reduction. As they say in their ads, its like gaining two full stops.
By the way, my primary bird and travel (combined with a Raynox in the back pocket) is also a used bridge point and shoot superzoom (Nikon P950) with a modest 16MP sensor. It allows RAW storage and when I shoot (usually in much higher ISO than most would dare), I have no idea of what the actual detail will be because it shows you the camera’s ‘interpretation’ on the camera screen, when you shoot RAW. But I’ve learned to not pay too much attention there and trust the DxO software to pull out the details and give me smooth shadow/highlight detail and colour.
Here’s an example if you’re interested: https://www.inaturalist.org/photos/513149977
It is specifically generative AI that “purists” object to. Because generative is fundamentally incapable of not making stuff up. A photo that has been edited with generative AI cannot be reliable evidence.
There are all sorts of advanced applications that are not generative AI, but which may be called “AI” (a term which in my opinion has become meaningless in any case). Those are fine, or at least arguable. Let’s not obscure the issue by lumping it all in together.
Part of the issue is that the lines between these things are becoming increasingly more blurred. All modern smartphones and also all modern cameras have pretty advanced algorithms built in to create a “good looking” image file from the data the sensors are reporting*. Whatever we want to call them, it’s software that makes decisions that influences the pixels in the final image. Even if it’s not “generative” it has a sufficiently strong impact on hue, saturation and value of every pixel that the result may significantly deviate from what we perceive with the naked eye. While it’s mostly fine, the best practice advice should probably be: “Be aware that it’s near impossible to create images that are completely free of ‘editing’ choices by software/AI, but try to keep it to the minimum that’s workable for you.”
*Sensors themselves are also not neutral elements in this.
I disagree with this statement. For one thing, it’s really hard to actually evaluate since no definition of “AI” is provided. But some/many of the aspects of onboard image processing done by many cameras, especially older ones, wouldn’t meet most people’s definitions of “AI”. They are often more algorithmic than AI.
Reasonable people can disagree whether something should be called AI or not. It likely is besides the point though. We probably all agree that between light hitting camera sensors and an image being uploaded to inat a lot of digital processing happens and there are various degrees to how truthfully the result will be a representation of the organism that was encountered. The question really is what degree of “editorial freedom” can be tolerated for the purposes of inat, no matter what we call the software doing it. Edit: With the understanding that for technological reasons it’s in practice impossible to have “editorial freedom”=0
With the above in mind, please be mindful that for many observers, the only economically feasible tools will be the phones in hand, and provide best practices, not barriers.
Many years ago I was introduced to an instance where a subject had been scanned multiple times by a scanner and the resultant images averaged, for the purpose of noise reduction. Cameras could adopt similar strategies without the use of generative AI. For example it could take multiple images, line them up somehow, and then average them. Google Pixel phones can, I’m told, achieve remarkably good night sky photographs in this way.
Algorithmic forms of computational AI are intended to capture what’s seen by the camera while reducing artefacts from motion blur, limited depth of field, camera shake, lighting conditions, etc. They might not always be successful, but they’re unlikely to produce outright hallucinations.
If we want to get really technical, your eyes aren’t all that reliable and accurate, either, and your brain is doing quite a lot of work between what your eyes perceive and what you are sure that you have seen.
Trying to get the final image to match what you saw in the field is a matter of trying to get technology to do something that produces a result that looks according to you what it looked it according to you. The technology, whatever it does, is sandwiched by your brain doing its best to tell you that what you perceive is what is actually there, and buffered by your memory.
…but I don’t see (ha!) where it gets us to think this way. We have no access to objective reality except as mediated by our brains interpreting our senses.
My point is that maybe it isn’t as helpful as all that to get that technical. We all have some intuitive understanding of what it means to say “the image should match what you observed in the field”. We equally have some intuitive understanding of what it means to say “the computer processing should not add details that were not there”.
Excellent article in the Guardian. “Scientists are appealing to birders to limit their use of AI when editing images over fears that it could undermine the credibility of popular citizen science platforms such as iNaturalist and Macaulay Library, which are routinely used in scientific research to monitor species’ habitat range.”
Many of us have seen these images on social media, I’ve often seen them innocently shared as part of a post focused on bird conservation. Having curated Surrey’s iNat program since 2019, I’ve only seen a few observations that were clearly shots of computer screens or TVs, but deep fakes are becoming increasingly sophisticated and harder to detect.
Fortunately the majority of dedicated community scientists out there haven’t bowed to temptation. Especially since we have plenty of cool, unique occurrences to share without having to resort to manufacturing them!
Check out the article here: https://www.theguardian.com/environment/2026/jul/20/ai-slop-manipulated-fake-images-birds-citizen-science-aoe
*This relates to the post in General from 2022 Coming soon: Deep fake nature photography? Things seem to have evolved exponentially since then!
I moved the above post to this thread which already has posts and discussion of this article to keep the conversation together.