This seems like the best solution, to me. Here is a recent example where I used that method [warning: this observation contains roadkill]. The first photo is nothing worse than many people will see on a day-to-day basis while cycling to work or school. The other photos show closer views (and yes, a photo of a paw is much more useful for training the AI than an image with a text warning).
I guess this might not have worked in the salamander case unless the observers were willing to take an extra photo of each occurrence. But “taking an initial photo from a distance in cases where the object might be disturbing to others (e.g. roadkill or entrails)” could be added to the information on how to make observations with the aim of making this a norm on the site.