Quick, Draw!: Google's website that guesses your doodles

Quick, Draw! is a free game from Google that hands you a word, gives you twenty seconds to draw it, and sets a neural network guessing out loud while your pen is still moving. It is worth opening for one slightly absurd reason: a machine is trying to read your scribble faster than you can finish it, and most of the time it wins. There is no account, no download and no price, only six short rounds and a computer thinking aloud at quickdraw.withgoogle.com.

In brief

• Quick, Draw!, at quickdraw.withgoogle.com, gives you six prompts and twenty seconds each to sketch while a neural network calls out guesses in real time.
• Google Creative Lab released it on 14 November 2016 as one of its A.I. Experiments, and players have since drawn more than a billion doodles.
• Fifty million of those drawings were tidied up and published as an open dataset of 345 categories under a Creative Commons licence, one of the largest public collections of sketches anywhere.
• The same heap of scribbles went on to train other Google tools, from the AutoDraw sketch helper to the Sketch-RNN research model that draws on its own.

What Quick, Draw! actually does

The game opens with a dare. It names an object, 'a helicopter', say, or 'a kangaroo', counts you down and gives you twenty seconds to draw it with a mouse or a fingertip. As you draw, a voice reads out its running guesses: 'I see a circle, I see a clock, oh, I see a helicopter.' The round ends the instant the network is sure, or when the time runs out. According to Wikipedia, a full match is six rounds, twenty seconds apiece.

What makes it stick is the commentary. The machine does not wait politely for a finished picture. It blurts out everything the half-drawn lines remind it of, often wrongly, sometimes hilariously, then lands on the answer just as you add the last stroke. A rushed drawing that looks like nothing to you can still be enough for it to shout the right word.

At the end of the six rounds, the game lays your drawings out and tells you which ones it recognised and which defeated it. Click any prompt and it shows what the network thought you had drawn instead, and lets you browse a grid of how thousands of other people tackled the same word. That last screen is quietly the best part: a wall of strangers' attempts at 'octopus', no two alike.

The A.I. Experiment behind the game

Quick, Draw! did not arrive as a standalone toy. Google's Experiments with Google gallery lists it as one of the studio's A.I. Experiments, a run of small, playful projects meant to show what machine learning can do without a lecture. It was built by the Google Creative Lab and Data Arts Team, with Jonas Jongejan, Henry Rowley and Nick Fox-Gieg among the credited makers, and released on 14 November 2016.

The friendly surface hides a deliberate plan. Writing in Fast Company, reporters described these silly-looking experiments as a serious strategy: a way for Google to make an abstract technology feel legible and even fun to ordinary people. A doodle game that talks back is a soft introduction to the same pattern recognition that powers far less whimsical software.

It also had a job to do. Every drawing a player makes is a tiny, labelled training example, a picture with a known answer attached. The game is therefore both a demo and a data-gathering machine, dressed up as twenty seconds of fun. That double life is the whole point of it.

A billion doodles, and the world's largest sketch dataset

The scale is the surprising part. Google reported on The Keyword that people have drawn more than one billion doodles in Quick, Draw! since it launched, from players all over the world. Very few browser games can claim a billion of anything.

A slice of that has been given away. The Quick, Draw! Dataset on GitHub gathers fifty million drawings across 345 categories, released under a Creative Commons Attribution licence so that anyone can use them. Google's Cloud team explained in a blog post that the drawings come in several formats, from raw timestamped strokes to simple 28-pixel bitmaps, hosted on GitHub and Google Cloud Storage.

Each doodle is stored as a set of pen strokes rather than a flat image. Wikipedia notes that the strokes are simplified with the Ramer-Douglas-Peucker algorithm, a standard method for thinning a line down to its essential points, which keeps the files small while preserving the shape and the order in which the picture was drawn. The result is not just what people drew, but the sequence in which they drew it.

What the machine learns from a rushed scribble

The recognition trick is less magical than it looks. The network was trained on huge numbers of human doodles, each tagged with the word it was meant to be, so it learns the rough visual grammar of a cat or a sailboat from the crowd rather than from any single perfect example. The more people play, the more examples it has seen, and the weird, wrong and lazy drawings turn out to be as useful as the tidy ones.

It also learns something about the world. Because each drawing is tagged with the player's two-letter country code, as Wikipedia records, the dataset has let researchers compare how the same object is drawn in different places, from the direction people sweep a circle to whether a chair is drawn face-on or in profile. A billion labelled sketches is a strange but genuine record of a shared visual habit.

The privacy footprint is deliberately light. The game keeps only the drawing itself, the word you were asked to draw, whether the network guessed right, and that country code, with no names or accounts attached. In an era of heavy tracking, a viral toy that files away almost nothing about you is its own small novelty.

From guessing game to AutoDraw and Sketch-RNN

The doodles did not stay in the game. Google has said the same data powers AutoDraw, a companion tool that watches your rough sketch and offers clean, professional versions to swap in, effectively an autocomplete for drawing. The billion-doodles announcement on The Keyword names AutoDraw as a direct beneficiary of the collection.

Researchers took it further. In the 2017 paper A Neural Representation of Sketch Drawings, David Ha and Douglas Eck of Google used the Quick, Draw! sketches to train Sketch-RNN, a model that does not just recognise drawings but produces its own, building a cat or a pineapple one stroke at a time. The work became part of Google's Magenta project on machine creativity.

The recognition engine has had quieter uses too. Wikipedia records that the underlying technology fed into other Google products, including the app Spoken, which turns drawings into synthesised speech, and work on character and handwriting recognition in Google Translate. A game about drawing a frog in a hurry ended up as a small building block in several other tools.

Why a doodle game still pulls people in

Quick, Draw! keeps getting rediscovered, and the reason is not the technology. It is the comedy. A confident machine voice insisting your lopsided shape is 'a bandage, a hospital, a light switch' before finally getting it is funnier than most things built to be funny, and it never plays the same way twice.

It also asks for nothing. There is no sign-up, no score to chase across sessions, no notifications and no clutter, just a prompt and a timer. That restraint is rare enough that the game has found a second life in classrooms, where teachers use it as a five-minute, low-stakes way to talk about how machine learning actually works.

Most of all it is honest about being an experiment. It does not pretend the network is clever or infallible, and half the fun is watching it fail out loud. In a moment when artificial intelligence is usually sold as either a miracle or a menace, a toy that lets you watch one fumble a drawing of a toothbrush is a healthy corrective.

Where it sits in the quiet web

Quick, Draw! belongs to a small family of sites that turn a machine-learning demo into a plaything. Its closest cousin already covered here is This Person Does Not Exist, which runs the idea in reverse, generating faces instead of reading drawings, and shares its watch-the-algorithm-work appeal.

It also sits happily beside the draw-with-your-mouse toys, like Silk, This is Sand and Drawing Garden, all of which reward a few idle minutes and ask nothing in return. The same unhurried spirit runs through the wider Boring Websites network, including Today's Rock, which asks only that you look at one rock for a while.

For the broader genre there is The Useless Web, which flings you toward a random oddity, and a longer roundup of sites to visit when bored. Quick, Draw! earns its place among them by being useful and useless at once: a genuine research tool that also happens to be a very good way to lose ten minutes.

FAQ

What is Quick, Draw!?

Quick, Draw!, at quickdraw.withgoogle.com, is a free browser game from Google. It gives you a word and twenty seconds to draw it, while a neural network guesses aloud what you are sketching. A full match is six rounds, and at the end you can see which drawings it recognised and how other players drew the same things. There is nothing to install, buy or sign up for.

Who made it, and when?

It was built by Google's Creative Lab and Data Arts Team and released on 14 November 2016 as one of the company's A.I. Experiments, according to Wikipedia and the Experiments with Google gallery. Credited makers include Jonas Jongejan, Henry Rowley and Nick Fox-Gieg.

Is my drawing saved, and is it private?

Your drawings are saved, but with very little attached. As described on Wikipedia, the game stores the drawing, the word you were asked to draw, whether the network guessed correctly, and a two-letter country code, with no name or account tied to it. Those anonymised doodles are what feed the public dataset.

What is the Quick, Draw! dataset used for?

The open dataset of fifty million drawings is used by researchers, developers and artists for machine-learning projects. Google itself used the doodles to build the AutoDraw sketch helper and to train the Sketch-RNN model described in the paper A Neural Representation of Sketch Drawings.

More boring websites like this: Generative art websites.