Outside of iNaturalist, I regularly use AI image editors such as Gemini to clean up microscopic pictures from debris and sometimes improve focus/lighting. Just a year ago, the quality of such edits was not acceptable at all with hallucinations being common. Nowadays, the performance of not just editing but photorealistic generation as a whole has reached a new level entirely. Take this messy shot of a bryophilic mold as an example:
I assume that AI-generated content is generally not allowed on iNat but do wonder if editing cases without fraudulent intention pose an exception to this rule, and it they do, whether images are excepted to be marked as AI-edited at all times.
Thanks, I was not aware of the article and it answers my original question.
It would still be interesting to hear about other’s experiences with AI editing of microscopic images as my example does seem to violate multiple rules. (upscaling, removing the host leaf of the fungus)
Ebird recently made a good effort towards being clear about this topic, I recommend reading it. It can generally be followed for iNat as it applies to data quality in general: https://support.ebird.org/en/support/solutions/articles/48001064357-photo-upload-guidelines
Here is Cornell’s “AI enhancement policy” from Levi’s link:
In gauging whether or not an AI edit is suitable, I would always ask the question: Has important identification information or context from the original image been lost or altered? In the example you demonstrate, are the apparent cilia important? Are there other organisms in the image which provide context for the primary organism? Etc. (I don’t need answers to these hypothetical questions; these should be what you ask of an AI edit.)
Along those same lines, if you’re not adding a scale bar to the image, the AI “enhancements” can make it difficult to judge scale, which is often important for IDs of microorganisms. If one has looked at enough microscope images, one can get a feel for what images look like at different magnifications, based on the apparent depth of field in the image. Most of the microscope images on iNat have no scale bars, but it’s usually apparent to an identifier roughly how big the subject is. I wouldn’t have any idea what size I’m looking at in that AI version of the image, as there’s no context and everything has been modified to be more in-focus. It’s like a pseudo-focus stack almost. It looks cool, but the original is more likely to be identifiable.
Personally, I don’t use AI for much of anything and that includes my iNat photos, though I do occasionally edit my photos in other subtle ways. That being said, when I do so, I like to include both the pretty version with the slight editing and also the unedited version in my observation. Then, if I accidentally obscured something important in my edits, it doesn’t get lost!
