Quick, Draw! And have a neural net guess what it is

I’ve lost entire afternoon’s to GeoGuessr, hitting the button again and again to teleport to a random place in the world in Google Streetview and then try to work out where I am from the scenery. Now I’ve spent a similarly long afternoon making Google do the guessing. Quick, Draw![ official site] is a sort-of game, sort-of web tool in which you doodle images upon request and a neural network tries to guess what it is you’re drawing. Come, play, abandon productivity.

It’s simple enough on the surface. The game tells you to draw a raccoon, for example, and you’ve got twenty seconds to do so in black line. The webpage calls out its guesses, “I see bird. Cat. I see skateboard.” until either it gets it correct or, in the case of at least one of my drawings, it lets out a resigned, “I don’t know what that is.” You draw five images in total and then it takes you to a results page.

On that results page you’re able to see a little more of the complex workings that obviously underpin something like this. The neural network is designed to learn to recognise the images based on what other users have drawn, and so you can click to see that, heck, your drawing of a raccoon wasn’t that far off other people’s efforts. This machine is just throwing shade your way. It also shows you an overlay of why it thought perhaps your beautiful rending of a rodent was a skateboard, along with comparative images of some of its other guesses.

Then you hit the play button again and go round and round.

It’s impressive sometimes how quickly it guesses the correct answer. For example, I got this far through my beautiful rainbow drawing before it got the answer:

Whereas on other occasions it reaches a point of awkward silence, where you’re finished drawing and it’s made all the guesses it can think of and you just start drawing a second one next to the first in the hope that helps but it doesn’t help and oh no. It’s as close as you can get to playing Win, Lose or Draw without Danny Baker sleeping on your couch for weeks afterwards.


  1. Chalky says:

    I guess it’s easier for it to guess when it’s got a very limited set of things that you’re going to be asked to draw. That said, I was asked to draw a portrait at one point, then asked to draw “The Mona Lisa” later on.

    I’m pretty interested to know how it’s expected to distinguish between those two things. I think google over estimates the artistic prowess of the masses.

    • Premium User Badge

      johannsebastianbach says:

      It struggled with my drawing of the Mona Lisa, too. Even though I drew it with the exact frame the Louvre uses for display, trees in the background and everything :-(

    • ButteringSundays says:

      If it’s a neural net as Google describe then it’s likely more complicated than you’re giving it credit for.

      Ie it’s unlikely just to be pattern matching based on the ‘pool’ of requests – as that would be trivial, and no AI would be required.

      That said this isn’t my bag – so I can’t add any more insight than that…

    • Sic says:

      Does it have a limited set, though?

      I played for quite a while, and never got the same word twice.

  2. geldonyetich says:

    3/5 o n my first attempt, and I’m no digital artist. Not bad, little AI!

    • Premium User Badge

      Malarious says:

      Yeah, it’s quite impressive. I got told to draw an ambulance, so I started drawing a box with wheels — it guesses car, then truck — so I add sirens, and it guesses “police car”, and lastly I added a cross to the side of it, and it immediately guessed ambulance.

      Will be interesting to see, 3 months from now, the reverse: you give the neural net an object and it draws what it thinks it looks like, from the thousands of drawings people have submitted.

  3. TomxJ says:

    “Draw Animal Migration”

  4. Robert The Rebuilder says:

    Maybe I’m missing something, but couldn’t the software essentially do this:

    X = name of random object
    Tell user “Draw X”
    Let user draw for 1 minute
    Tell user “You drew X”

    It would be more impressive if it didn’t tell you what to draw.

    • Chalky says:

      They’re training their new captcha system, so it’s important to tell the user to perform a complex task rather than letting them do what ever they want.

      It still needs a bit of work:

      link to i.imgur.com

      • Kitsunin says:

        On the flip side, it can be pretty amazing.

        • Sarfrin says:

          That looks more like a duck poking Picasso’s Weeping Woman than a garden.

          • Kitsunin says:

            Yeah…my thought process was basically “How do you draw a freaking garden? Stuff growing out of the ground, like an uh, carrot top…? Oh god this looks nothing like a garden, a watering can is obviously garden-y right? Oh nooooo 3 seconds left whyyyyyyy.”

  5. try2bcool69 says:

    You do realize that, by playing this, you’re helping Skynet learn how to eventually destroy humanity?

    Plz stahp.

  6. trappski says:

    When it couldn’t guess line i gave up…

  7. Jerppa says:

    Did anyone else draw a wang just to see what it thinks it is?

  8. Mrice says:

    I played it through 4 times and it couldn’t guess a single one. Guess im shit at drawing :(

    • wu wei says:

      It wasn’t even making suggestions with mine, so I checked the console. If you have any extensions that block Google APIs (Privacy Badger in my case), it won’t work correctly.

  9. Seraphithan says:

    If you have one you should try using an ipad or an other tablet, drawing with a finger works so much better than the mouse.

  10. Hmm-Hmm. says:

    It’s rather impressive what other people can draw in just 20 seconds. Still, I only failed once.

    • Premium User Badge

      ooshp says:

      Yeah I was thinking this should be tits with the SP4 stylus, but haven’t got around to trying it yet.

  11. deadpan says:

    It couldn’t recognize anything I drew, so I guess we’re safe. For now.

  12. kwyjibo says:

    I got 6 out of 6. Book, Camera, Arm, Mermaid, Belt, Bat.

    Bat was interesting because I went with the flying rat and not the cricket. I wonder if it knows both. If I look at the results, most people go with the animal, some go with baseball, and no one likes cricket (because it’s a rubbish excuse of a sport).

    And it’s only going to get better at it. This is Google we’re talking about, every game we play takes us closer to Skynet.

    • Al__S says:

      Skynet is a real Military satellite system. It’s the UK’s system. I feel not enough know and are scared about this.

  13. parsley says:

    Oh, that IS fun. Apparently I can’t draw a pick-up truck.

  14. April March says:

    It is fun, but I wish it also told us what the AI was thinking while we drew. I’m told why its final guess as to my bridge was a rake, but it doesn’t tell me why at a point it thought it was a knee.

    I also learned that when asked to draw a bridge in 20 seconds most people don’t draw suspension bridges. ?

    Edit: Also, apparently no one draws a book with proper perspective. Or at least a single person did but from a different perspective than I did so the AI didn’t recognize it.

  15. RanDomino says:

    I wonder what Wittgenstein would say about this game.

  16. Caiman says:

    Yeah, it’s clearly comparing it to its existing data, because several times it successfully guessed my drawing before it was clear. My street lamp could just as easily have been a walking cane, or the start of an umbrella, or a particularly inquisitive snake. If you click on each image afterwards, it shows you how it works, basically comparing what you’ve done with previous drawings. It’s still impressive though.

    • froz says:

      That is exactly how human brain works and i don’t know what could be more impressive.

  17. Poppis says:

    I drew a tiger and it guessed an aircraft carrier.

  18. Harlander says:

    The little images of the things other people drew that it used to infer what your picture is are pretty interesting – particularly how so many people tried to represent a hockey puck in the same titled perspective that shows its thickness…