AutoDraw: the website that fixes your terrible doodles

AutoDraw is a free web tool from Google that watches a rough scribble, guesses what it is meant to be, and offers a clean version drawn by professional artists in its place. It is worth a look because it turns the most dreaded part of making a quick diagram, actually drawing, into a short act of pointing at the right suggestion.

The pitch is almost too plain. A person draws a lumpy blob that is supposed to be a cat. A row of polished cat icons appears along the top of the screen. One click swaps the blob for a proper cat. There is no account, no download, and no cost, which is the kind of restraint that makes a single-purpose website pleasant to recommend.

In brief

• AutoDraw reads a freehand sketch, predicts the subject, and suggests clean replacement drawings made by real illustrators, all inside the browser at autodraw.com.
• It was built by Google Creative Lab and launched on April 11, 2017, reusing the recognition engine from the earlier Quick, Draw! experiment.
• The polished suggestions come from seven named artists and studios, so the output looks hand-drawn rather than generated.
• It is free, needs no sign-in, and runs on phones, tablets, and desktops.

What AutoDraw actually does

The core loop is simple. A user picks the AutoDraw tool and starts sketching on a blank canvas. As the lines go down, a horizontal strip of guesses fills in near the top of the screen. The guesses are candidate objects: a sun, a lightbulb, a shoe, a fish. Selecting one replaces the messy original with a crisp, professionally drawn icon scaled to fit.

The official Google announcement, titled Fast Drawing for Everyone and written by Creative Technologist Dan Motzenbecker on April 11, 2017, frames the goal as making drawing "as easy and fast as everything else on the web." That framing is the whole product. The tool is not trying to make anyone a better artist. It is trying to remove drawing as an obstacle.

Around the AutoDraw tool sits an ordinary drawing app. There is a plain pencil for freehand lines that skip the suggestions, a text tool, a fill tool, a shape tool, and a color picker. A finished piece can be downloaded or shared. None of that is novel. The suggestion strip is the reason anyone remembers the site.

The feel of using it

Part of the appeal is speed. The guesses update live, so a rough oval becomes an egg, then a balloon, then a face, depending on what gets added next. The experience rewards committing to a recognizable shape rather than fussing over detail. A person who cannot draw a bicycle can still produce a page full of tidy bicycles, which is the small magic trick the site performs over and over.

The machine learning under the hood

AutoDraw did not build its recognition from nothing. According to Google's announcement, it uses "the same technology" found in Quick, Draw!, an earlier A.I. Experiment. Quick, Draw! was released in 2016 by Google Creative Lab in collaboration with the company's handwriting recognition team, and it asked players to draw a prompt while a neural network tried to name it in real time.

That game was more than entertainment. Every doodle people drew became training data. The network learned to associate the gesture of a drawing, the order and direction of strokes, with a label. AutoDraw inherits that learned association and points it at a practical task: not scoring a guess in a game, but offering a replacement a person can actually use.

Why strokes matter more than pixels

One detail explains a lot about how the recognition behaves. The Quick, Draw! data stores each drawing as a time series of pen positions rather than as a grid of pixels. The order in which someone draws a shape carries information. A circle drawn clockwise then crossed with two lines reads differently from the same final image produced in another order. Because the model learned from stroke sequences, AutoDraw often guesses correctly well before a drawing looks finished, since the path of the pen is already a strong clue.

Why the suggestions look so clean

The recognition is only half the system. The other half is a library of finished drawings that a machine did not make. Google's announcement credits a specific group of contributors: the design studios HAWRAF and Selman Design, and the designers and illustrators Erin Butner, Julia Melograna, Pei Liew, Simone Noronha, and Tori Hinn. Seven names and studios, all human.

This matters for how the output feels. When the tool suggests a cat, it is not synthesizing a cat from statistical averages. It is surfacing a cat that an illustrator drew on purpose, with consistent line weight and a deliberate style. The result sidesteps the smeared, uncanny quality that purely generated art often carries. The machine's job is matching, not drawing.

The announcement also invited more artists to contribute their own drawings to the library, treating the icon set as something that could grow. That design choice keeps a person in the loop at the point where taste matters most, the final mark on the page.

What it is good for, and what it is not

AutoDraw is at its best for fast, functional pictures. A slide that needs a quick icon, a worksheet for a classroom, a birthday card, a sticky-note diagram, a simple logo sketch to hand to a designer later. In all of these, the point is a clear symbol, not a personal style, and a clean library icon beats a shaky freehand attempt every time.

It is a poor fit for anything that depends on the drawing being yours. The suggestions are a shared, finite set, so two people reaching for the same subject land on the same icon. There is no texture, no expressive line, no happy accident. The tool also only recognizes subjects that exist in its trained categories, so an unusual object may never trigger a useful guess and has to be drawn by hand with the plain pencil.

Approached honestly, those limits are features. The site set out to make serviceable pictures quickly, and it does exactly that. Expecting it to carry a creative voice is asking a convenience to be an art studio.

A tool built for people who say they cannot draw

The quiet audience for AutoDraw is the large group of adults who decided, usually in childhood, that they cannot draw. That belief tends to harden around the age when a person first compares their work to someone else's and finds it wanting. From then on, a blank canvas reads as a small threat rather than an invitation.

AutoDraw lowers the stakes of that blank canvas. The user is not asked to produce a good drawing. They are asked only to make a shape recognizable enough for a machine to name, which is a far gentler bar. The polished result that follows belongs to a professional, so there is no personal failure to wince at and nothing to apologize for before showing it to a colleague.

That reframing is the real accessibility story. It is less about assistive technology in the formal sense and more about removing a social fear. A teacher can build a visual worksheet without dreading the art. A manager can sketch a flow on a shared screen without the usual self-deprecating joke about stick figures. The tool does not make people confident artists. It makes the question of talent beside the point.

There is a plausibility to the whole approach that newer tools sometimes lack. Because the user still draws the first pass and still chooses the final icon, the output feels earned rather than conjured. The person remains the author of the idea even when they are not the author of the lines.

How it differs from today's AI image generators

It is tempting to file AutoDraw alongside the text-to-image generators that arrived years later, but the two work on opposite principles. A modern generator invents pixels from a prompt, producing an image that did not exist before. AutoDraw invents nothing. It recognizes a sketch and then retrieves a finished drawing that a human already made.

That distinction has consequences a user can feel. A retrieval system is bounded and predictable: the set of possible outputs is a known library, so results are consistent, legible, and free of the strange artifacts that generators can produce. The trade is range. A generator will attempt any prompt, however odd, while AutoDraw can only offer what its artists have drawn and its model has learned to spot.

There is also a cleaner provenance to the AutoDraw bargain. Every suggested icon traces back to a named illustrator who drew it on purpose, which tidily avoids the questions about training sources that hang over generative art. The model is a matchmaker between a rough gesture and a specific human-made picture, not a synthesizer of new imagery.

Seen this way, AutoDraw is a useful reminder that machine learning in creative tools does not have to mean generation. Sometimes the most pleasant result comes from a system that simply understands what a person meant and hands them something a human already made to match it.

Where it sits in Google's drawing experiments

AutoDraw is one node in a small family of browser experiments that treat drawing as a way to study and play with machine learning. The lineage is easy to trace and is part of why the tool rewards a second look.

Quick, Draw! and its open dataset

Quick, Draw! came first and did the data gathering. The pictures it collected were later released publicly. As Google's open source blog and the project's own data page describe, the Quick, Draw! Dataset holds roughly 50 million drawings across 345 categories, gathered from players worldwide, which made it the largest doodle dataset of its kind at release. The full dataset lives on GitHub under the googlecreativelab/quickdraw-dataset repository, where anyone can download it.

That openness turned a web toy into a research resource. The drawings are tagged with the requested label and the player's country, which let researchers study not only whether a network can name a doodle but how drawing habits differ around the world.

SketchRNN and generative sketching

The dataset also fed work that went the other direction, from recognizing drawings to producing them. David Ha of the Google Brain team built SketchRNN, a neural network that learned to draw in strokes and could interpolate between sketches, released through the TensorFlow Magenta project. AutoDraw, Quick, Draw!, and SketchRNN together form a loop: a game collects strokes, a dataset preserves them, one model reads them, another model generates them.

The reception, and the staying power

The launch landed well with a technical audience. On Hacker News, the launch thread for Fast Drawing for Everyone drew more than 730 points and over 140 comments, with the usual mix of delight, nitpicking about recognition misses, and speculation about uses. Mainstream technology coverage followed quickly, consistently landing on the same friendly summary: Google made a thing that cleans up your bad doodles.

What is quietly impressive is that the site is still live and still free years later, listed to this day on Experiments with Google. Many playful web projects from 2017 have gone dark. AutoDraw kept running with no apparent change to its bargain: draw badly, get something usable, pay nothing. That durability is rare enough to be worth noting.

It belongs to a particular tradition of the web, the single-purpose site that does one modest thing and does it without friction. Readers who enjoy that tradition will recognize the spirit in the wider Boring Websites collection, where the whole point is a page that commits to one small idea and leaves everything else out.

How to get the most out of it

A few habits make the tool noticeably better. Draw the most recognizable version of a subject, not the most detailed one. The model was trained on quick doodles, so a stereotyped shape, a boxy house, a stick tree, a round-eared cat, triggers stronger guesses than a careful, idiosyncratic rendering.

Watch the suggestion strip as the drawing develops rather than finishing first and checking after. Because guesses update with each stroke, a person can steer toward the intended result by adding the one line that resolves the ambiguity, an ear, a wheel, a window. If a subject refuses to be recognized, switch to the plain pencil and draw it by hand, then keep AutoDraw for the surrounding icons.

Finally, treat the output as a starting layout. Clean icons placed, resized, and colored make a tidy composition in a minute, and the plain tools handle the custom bits the library cannot. Used that way, the site is less a novelty and more a genuinely quick way to assemble a simple picture.

FAQ

Is AutoDraw free, and does it need an account?

Yes on both counts. Google's announcement states the tool is free and runs in the browser on phones, tablets, and computers with no sign-in required. A drawing can be made and downloaded without creating an account.

Does AutoDraw use the same engine as Quick, Draw!?

Yes. The official post says AutoDraw uses the same recognition technology as the Quick, Draw! experiment, which trained a neural network on millions of player doodles to guess what a sketch represents.

Are the finished icons drawn by a computer?

No. The suggestions are drawings made by named human artists and studios, including HAWRAF, Selman Design, Erin Butner, Julia Melograna, Pei Liew, Simone Noronha, and Tori Hinn. The machine learning chooses which human-made drawing to offer, but it does not create the art itself.

Can the underlying drawing data be used by developers?

Yes. The related Quick, Draw! Dataset, about 50 million drawings across 345 categories, is open source and published on GitHub, where developers and researchers can download and study it. Because each drawing is stored as a sequence of pen strokes rather than a flat image, it supports research into how people draw, not only what they draw.

When did AutoDraw launch and who built it?

AutoDraw launched on April 11, 2017, built by Google Creative Lab. The announcement was written by Dan Motzenbecker, a Creative Technologist at the lab, and the project drew on Google's earlier Quick, Draw! experiment from 2016. Years later it remains free to use and is still listed among the official Experiments with Google.