Who said it was supposed to be useful or amazing?

I think GP has raised a fair point. The landing page says: "Can a neural network learn to recognize doodles" but the game doesn't demonstrate that the ANN is able to recognize arbitrary objects.

It merely shows that it's able to pick the correct name from a finite list of possible names that could very well be small (e.g. 10 or 20 object names).

A better way to demonstrate that the ANN is truly able to recognize objects with good accuracy would be to ask a person to draw an unspecified object and guess its name.

If the goal is to acquire training data, they could still ask for a name in case of failure.

BTW: this reminds me of detexify[1], the tool that can recognize latex symbols from handwritten drawings. Same problem.

[1] http://detexify.kirelabs.org/classify.html

Fair. It was posted and voted up so perhaps I inferred that it would be useful or amazing.

As explained at the beginning, the purpose of the exercise is to train it.