If you display 10 random animals (or shapes), and ask a user to pick the right one, a dumb spam script would have a 10% success rate.
If you display one image of an animal and give 10 possible answers, a dumb spam script would have a 10% success rate.
What if I combine three pictures in each challenge - e.g. "cat house triangle"?
Plus you constantly have to update your image library, which a huge pain.
Also, recognizing 10,000 images will take me around one day and less than $1000 with Amazon turk, thus giving me a perfect 100% success rate. After that you would have to completely renew your image database.
YES, it would be a pain to update the library, which is why I'm commending this particular concept for solving that problem ...
If you present 10 images (still a stretch), bots will have 10% success rate just answering randomly.
EDIT: Wait, from what I see you mean that the user will have to write "cat" or "dog" or whatever? That's better, yes. Communication, however, is hard, which is why me the GP didn't understand what you meant.