Future Tense

The Weird Failures of Algorithm-Generated Images

When a style can be infinitely generated, it’s not much of a creative style anymore.

Three airplanes flying
The view from Microsoft Flight Simulator. Microsoft

Last week, New York Times columnist Farhad Manjoo published a curious column on Microsoft’s video game Flight Simulator, which uses a trove of data sourced from OpenStreetMap, filtered through Microsoft’s Bing Maps, to create a 3D rendering of the entire world. The data has been algorithmically translated into an enormous environment, every home, skyscraper, or mountain in the dataset made interactive. You can fly a virtual plane past a virtual replica of your house.

Manjoo casts the game as utopian. The experience of playing it, no pilot license necessary, shows you that the boundaries of our cities or countries don’t really exist: “The tech giant has done something uncanny here: It has created a virtual representation of Earth so realistic that nearly all sense of abstraction falls away.” Computers can give us “a view of the world that is more real than the one we can see outside,” he writes.

The argument reminds me of the old fictional fragment by Jorge Luis Borges, “On Exactitude in Science.” In the story, mythical mapmakers make bigger and bigger maps until they reach the limit: “The Cartographers Guilds struck a Map of the Empire whose size was that of the Empire.” Later generations finally realize that the giant map is useless and leave it to decay in the desert. The underlying message of the story, if there is one, is that no map can be completely accurate, or that abstraction is part of the purpose of maps. Some form of simplification is what gives them meaning.

Flight Simulator is a virtual Earth as large as the Earth, but it isn’t as good a map as it appears. Since the data has to be automatically translated into 3D models on a vast scale, there are plenty of glitches, as the game’s players have discovered. The Washington Monument got turned into an office building, for example:

The city of Bergen, Norway, became a surreal hash of landscape and construction, like an avant-garde architect’s rendering. Hills morph into skyscrapers before decaying into patchworks of pixels:

No human could render the entire world, nor could anyone investigate every corner of this virtual world to make sure it’s all correct. Whatever code Microsoft used failed to apply the appropriate surface to the appropriate 3D object. In the case of the Washington Monument, it decided that any particularly tall and skinny building must be a generic corporate office tower. This is how algorithms work. If certain factors are present, then a blanket solution is applied, which might be mostly right but is still wrong in a handful of circumstances.

Manjoo uses the game as a way to critique problems like disinformation and radicalization that happen via algorithmic platforms, including Facebook and YouTube. He writes that the social networks’ feeds aren’t close enough to reality: “It often feels like society is being shaped by the algorithmically defined sensibilities of online echo chambers and anonymous bots and trolls rather than the nuanced ideas of living and breathing people.”

Yet that’s exactly what’s happening with Flight Simulator. An automated, unchecked process is warping the (virtual) world around us, leading to these weird errors and aberrations. Bergen isn’t some post-apocalyptic semi-underground Hong Kong, but that’s how the data was interpreted. These glitches are evidence of the kind of algorithmic sensibility that Manjoo is critiquing.

There’s an aesthetic evolving out of algorithmic visuals like Flight Simulator. It reminds me of Google’s DeepDream filter, which is supposed to visualize how a machine learning system is perceiving an image, but ends up turning everything into hallucinatory dogs, because dogs are what the system was trained to recognize.

Some artists are intentionally adopting and playing with this algorithmic aesthetic. A video game called Townscaper lets you build tiny islands of Scandinavian architecture that automatically stretch and adapt as you add new elements. Unlike Minecraft, where changes are made by the player one by one, the structures adjust by themselves, adding in bridges and slopes so the whole island stays stylistically consistent, even though it might end up looking kind of weird or illogical.

A service called Blush is similar. It lets you generate that cliché kind of startup-branding illustration with bright colors and hard black outlines without commissioning an artist or interacting with a human. You can generate as many randomized Silicon Valley cities as you want and then use them to advertise your new co-living service on a landing page.

These tools of algorithmic style aren’t bad, necessarily, but they are uncanny, because what they do is extend a set of rules or a pattern that began as aesthetically pleasing over too large an area, too wide a swath of culture. The visuals are the equivalent of A.I.-generated writing: It might look OK at first, but it’s ultimately nonsensical and could be destructive in the wrong context. When a style can be infinitely generated, it’s not much of a creative style anymore. In the case of Blush, it also puts human illustrators out of work, and that’s not cool. Just hire someone!

We can appreciate strange glitches as artifacts of our time and hints of what the technological future might look like. But the automatic application of rules in digital space is never going to result in a more “real” reality (in regard to Flight Simulator) or a more “authentic” artistic creation. As with disinformation, we have to watch out for where the human stops and the algorithmic process takes over.

Future Tense is a partnership of Slate, New America, and Arizona State University that examines emerging technologies, public policy, and society.