I see what you mean about the settings, not 100% convinced it is AI, but yeah definitely have some doubts on it.

A human made image will have intent in a scene like this. The classrooms could be different, but the similarities will be consistent and the differences will be intentional and have purpose. With generated images, it's a poorly structured facsimile. It's actually easier to pick out in series of images like this, because the "this is what a classroom looks like" features exist in each image, but they're slightly different for no reason - the speaker is lower or shifted to the side, the poster board turns into a cork board, the ceiling changes material at random, etc, etc.

IMO, the more reliable sign in images like this is the poses. Those can be spotted even in singular generations. LIMs can't do groups of people very realistically (yet), so it ends up repeating poses throughout the image. In the first two images, all the characters are positioned pretty much identically. In the next to, you have can see differences, but even those get repeated elsewhere in the same image. Part of this is a model having similar patterns for "this is the facing forward pose" and "this is the facing right pose" that you'll see between generations, combined with a tendency for results to have less diversity within individual iterations. If "laughing and pointing" means one thing with a certain seed, it will typically mean that same thing anytime the algorithm results in "laughing and pointing" throughout the image.

It’s unfortunate how much you have to scrutinize content now to figure out if it’s real or AI slop.

It is, and it's going to get harder. But fwiw, my sole experience with this tech is locally hosted LLMs and LIMs that are necessarily much weaker than the commercial variants (since I don't have the means or desire to wreck communities and nature to erect heinous numbers of data centers), but even with the admittedly more impressive models, these tells still exist.