The catch tends to be if you're able to detect when you're in a more complicated situation and bail out accordingly. One very useful thing is auto-labeling, which turns out, getting it wrong is more problematic than one would expect.
Take for example the 'Gorilla' issue that Google encountered.
(This one isn't as big of a problem, as the frustration was probably more due to the compulsory usage of the new bookmarks, but it still shows some hazards of auto-categorization).
With social things, it can sometimes be very hard to solve a problem that works for 95% of the population, but doesn't piss off the remaining 5% (often with good justification). Unfortunately, detecting when you're in that situation in the first place is sometimes equivalent to the problem you're trying to solve.
I don't understand why a big company like Google can put some feature online if it's not thoroughly tested or thought. It is because it is a web product and it wants the users to be the testers?
Simple procedure: take a large sample (1M pictures or so), tag them with the system and then show each picture with the label to a person (using Mechanical Turk or similar) and ask "is this label offensive?"
Take for example the 'Gorilla' issue that Google encountered.
http://nyti.ms/1JlTBab [www.nytimes.com]
Then there was the issue with the bookmarks and suicide categories
http://bit.ly/1JFkmCr [productforums.google.com]
(This one isn't as big of a problem, as the frustration was probably more due to the compulsory usage of the new bookmarks, but it still shows some hazards of auto-categorization).
With social things, it can sometimes be very hard to solve a problem that works for 95% of the population, but doesn't piss off the remaining 5% (often with good justification). Unfortunately, detecting when you're in that situation in the first place is sometimes equivalent to the problem you're trying to solve.