Games With A Purpose is a small site with a few simple flash games on it. What's interesting is that the games help scientists with their job.
For example, if you've always wanted to hasten the onset of the robot apocalypse, you can play the games that help AI scientists test their prototype cybernetic overlords. Also, prizes are offered!
The Gender Guesser one made my mum laugh b/c she was guessed to be male at a higher percentage than me. Hehehe.
It's not that I really got a boner over the dear in a field and the mopeding through urban asia (ok yeah I did the second one), but when you phrase the question, "Walk on a beach, or watch some stupid fucking sports game OR HEY look it's wicked cool swirly glowing lights whooooooo", it was a no-brainer.
Woa. 97% chance I'm a female. Perhaps I'm am a woman and just don't know. But seriously it was very easy to know what pictures were male or female oriented based solely on stereotypes.
...
Did it again and now I'm 84% male. Doesn't look very scientific to me.
According to our calculations, there is a 91% chance that you are male. heh, yup this is truly a breakthrough. How did it know? Just cos I'm not into sparkly high heels?
I think it's not so much supposed to be able to predict your gender, it's getting you to help it learn how to predict gender.
I don't quite get the usefulness of being able to link gender to aesthetic preference. Short of being able to design gender-centric websites better. Which isn't really that hard to begin with.
(Edited to add: Has anyone played Squigl yet? Jesus, I could play that all day.)
If the Gender Guesser offered more than two options, I think it'd make a much better experiment - a lot of the choices were very extreme, either masculine or feminine.
I've taken it twice, the first time it said I was 58% female, the second time 58% male. Go figure.
With the gender guesser I don't think it's expecting to get good results straight away. It seems it's more like a classification based piece of machine learning. I've written stuff like that in the past, you get pretty decent results after you've trained it. Problem is, the training can take months.
It'll start out with every image weight exactly at 50/50 for male/female (so the first person to take the test will be told they're 50% male), then re-bias each images score slightly based on whether the user tells it it's correct or not. You'd only shift the bias a small amount each time (to prevent outliers breaking the data set), and assuming the picture pool is big enough it would eventually get a pretty solid 'gender score' for each one.
It's never going to get everything right every time, because classification of anything is an amazingly hard thing for a machine to do (we still don't have a great idea how the brain actually does it, just that it seems to be pretty good at it).
Man, that takes me back. Haven't thought about that kind of AI problem since I was at uni 10 years ago.
How long has the site been up? The gender thing might just need more info to make better guesses... Or we're all freaks... Either or, both seem reasonable.