Back when the App Store was a goldmine, I contemplated creating an app that would help people find public restrooms. I knew that mothers with babies were very sensitive about cleanliness and good changing tables etc, so this would be a "yelp for bathrooms." I mocked up the UI etc, and gave it a good thought and then did some schema work etc. Then I realized that all the users would be uploading the worst pictures imaginable; and we all know how bad bathrooms can be. There would be no way to moderate that type of thing, and I flushed the idea...
It's a tricky idea but honestly one that sounds worth pursuing. I certainly wouldn't mind knowing where all the nice restrooms are hiding within a random city or suburban block (for example that the one on level 22 of a particular skyscraper is public-access, or that the one on the lower level of a particular upmarket underground carpark is appreciably hard to reach, or that nobody realizes the one behind the service elevator lobby (that you only know about if you go to level 2 in this particular mall) is actually not in a restricted area).
I do sadly agree that the app's eww-score would rapidly converge to represent the worst-case-scenario baseline. This is the trickiest part; squinting and looking from the right angle I could see Apple pulling such an app from the app store, and possibly Google as well. (The apps would definitely just be WebView containers with little investment.) But I see that as a possibility, not a certainty; they would provide a useful service as well.
There are a lot of interesting ways to potentially solve the moderation problem. Fundamentally the most complex yet most viable approach would naturally be to train an ML model to distinguish along a spectrum from "WHAT" from "wait is that nearby? that looks better than my current favorite." This would need training data and lots of manual classification until it reached a point of viability, so you would very probably need a small army of assistants helping to both classify image inputs and override the model's output. The good news is that with the service just starting out "army" could probably mean "1 or 2 people"; the caveat-emptor news is that you'd need the ML model ready go from day 0 in order to not wind up behind the curve. You might also be able to achieve sustainability by requiring users to perform 5 or 10 moderation tasks, say per fortnight, to gain/retain access to the service. (You'd send the same tasks to multiple users and aggregate the results, not just to weed out junk inputs, but because people will score things differently.)
While controversial some algorithmic magic would probably be necessary to keep the service viable. For example if all you have are terrible photos of one particular location it may be prudent to hide all of them just in case people are only posting when it's bad, and show more prompts to take photos "especially if it looks good". Figuring out how to find the signal in the noise would rapidly become the long-term headache: for example, that a particular location is only cleaned by the thorough janitor every Tuesday, but because the building entrance that would lead people to that restroom closes early that day very few people use that location, and the only uploaded photo ever submitted on a Tuesday was from a passer-by who happened to be in the building staying late. 98% of that context would never be recoverable. The service would need to still provide actionable insights.
TL;DR I genuinely think that moderation is a solvable problem and that this is a cool idea - it's getting knowledge about it into people's heads that's the real challenge IMHO.
As an aside I think it would also be both useful and interesting to publish data dumps of all of the uploaded images (possibly with added watermarks) and attached classification data to give the ML community an extra dataset to play with - this would be an unusual but effective way to raise awareness about the service. Maybe even release the model, and possibly even open-source the website and app. This would be a sure-fire (but high-stakes) way to attract suggestions for improvements to the ML model, receive drive-by fixes for obscure bugs people hit in the field, etc.
The restrooms for the employees at a company says a lot about the company.