Hospital. Ambulance. Camouflage. Angel. Baseball bat. Nail. What kind of images come to mind when you see these words? How would you doodle them if you only had 20 seconds?
Google’s Quickdraw is an online web-based game that you can have fun with while you teach Google’s doodle recognition AI (yes, you’re helping train Google’s AI for no pay, but you already do that each time you use Google, so what’s new?)
Also not new? The way this neural network reproduces dominant culture. It does not harm anyone in the sense mentioned by Safiya Noble (stereotyping objectification of women of color) or Cathy O’Neill (increasing discrimination against African Americans in courts)… But it’s an almost transparent look at how the neural network learns from data sets and is therefore fascinating to watch… A bit like my undergraduate thesis, but more obvious to non-techie people because it’s visual.
Here is how it works. The game asks you to doodle a word it shows you in 20 seconds. Like pictionary. The Google AI is supposedly blind to that, and tries to guess what you’re drawing as you draw it, and if it guesses correctly, you move on to the next word. If not, you get the full 20 secs to keep trying. After 6 of these, you get to see your results. If you got it wrong, you can see what Google’s AI thought your drawing looks like. And what others had drawn that it considered correct.
Here are the main issues I have with this game (they’re not really issues I guess, except maybe the first two)
- The name Quick Draw. I originally didn’t think much of it, remembering a cartoon I once saw of a Western (mouse? Horse?) character called Quickdraw McGraw. But as soon as I mentioned this to hubby he asked me, “You know what it means, right?” and I suddenly realized it meant someone (cowboy I guess) who draws his gun quickly. Ugh. In a time of gun violence in the US causing so much tragic death in schools, and the ugly history of the West and what they did to native Americans.. They couldn’t find a more neutral name?
- Linguistic incompetence. I was pleasantly surprised to find the game available in many languages including French and Arabic. Not so pleasantly realizing you have to read the word Arabic in Latin alphabet to open the Arabic version and at way bottom of the screen not top. Not the coolest, and an easy fix imho. But that’s not a showstopper. However, the weird thing is that the Arabic version, even if you get one right (as in, Google’s AI recognized it as similar to what it learned) , when you click later to see details, tells you Google’s AI didn’t understand your drawing.
- Baseball. I cannot believe the French and Arabic versions ask for baseball as one of the key terms. Do they not know that very few people in the Arab world (much of the world?) play baseball? Rolling eyes
- Church and ambulance. A key aspect of what identifies a church or ambulance is a cross sign. You know in Israel they use a David’s star and in Muslim countries we use a red crescent, right? But the majority of data this neural network is seeing comes from Christian-culture countries. I’m gonna keep playing and see if the Arabic version expects a crescent. It should…. If it learns each language on its own… Or if it starts to connect across cultures. I think it connects across cultures because…
- Angel. I consistently got this one right. Even in the Arabic version (even though it says I’m wrong in the image I inserted earlier, because see #2). The reeally weird thing is that in Muslim culture (which is dominant among Arabic speakers), angels don’t look like this at all. I don’t know what they look like because we’re not supposed to depict them. So this is either a sign that the AI learns across languages… Or that Arabs doodle angels that look like Westerner angels because they don’t have their own imagery for angels. Interesting?
- Camouflage. This one was damn annoying. First of all, because connotation of camouflage for me is just how butterflies or chameleons blend with their background so their predators or enemies can’t see them. Apparently, the connotation for Americans is army uniforms intended for the same purpose. And to get it right, for the AI to understand me, I have to think like an American. Grr. Also, when I mentioned the term camouflage to an American yday she started saying it had an uncomfortable negative connotation for her and so obviously I understood she made an immediate connection to military. Not at all what came to my mind. A reminder that fluency in a language and familiarity with a culture are never identical to living that culture and being native to the language.
- Nail. I don’t know why they chose that word because it has multiple meanings. Nail as in fingernail or toenail, but also as in nail/hammer. Apparently most people thought fingernail coz the AI couldn’t at all understand the other kind of nail. I think eventually it will get smarter but that requires users to not learn to follow the AI, you know? For example, if people keep doing both types of nail, the AI should eventually learn both are OK, in the same way it understood a face view and a side view of a kangaroo (this is really cool btw).
- Envelope in Arabic. This one was just weird. The word it used to say envelope in Arabic made no sense to me. And I am familiar with 3 different dialects of Arabic. I don’t understand where the word they used come from. I’ve already forgotten it, actually!
- Stop sign. Oh no. The Arabic stop sign expects u to write Stop in English not Arabic. It clearly isn’t differentiating between languages. It’s possible it’s not programmed to find that difference. I wonder if it might learn that on its own or if it has no idea what language input the user sees… Depends what variables it’s allowed to consider while it learns (and that’s how machine learning differs from human learning).
So that’s my reaction to it.
It is a fun game though. I let my kid play as a fun way to practice reading and drawing and we discuss the cultural connotations, too. It’s fun for me to see how other people draw things…
And I discovered this game by a fun coincidence that my students discovered it while doing some ds106 assignments for extra credit! Lucky me 😉