When Google Earth Breaks, It Accidentally Becomes Art

New York–based artist Clement Val­la has been steadi­ly col­lect­ing images of the Earth, but not the Earth as we know it. Through his ongo­ing series, Post­cards from Google Earth (2010–present), Val­la has com­piled a dig­i­tal archive of topo­graph­ic anom­alies that defy the real­i­ty they are meant to rep­re­sent: visu­al arti­facts that expose the lim­i­ta­tions of ear­ly machine learn­ing and qui­et­ly antic­i­pate today’s debates over the accu­ra­cy and author­i­ty of arti­fi­cial intel­li­gence.

Bridges plunge at impos­si­ble angles into the water below, high­way inter­changes col­lapse into one anoth­er, and tree­lines stretch and warp into unnat­ur­al forms. For more than a decade and a half, Val­la has gath­ered screen­shots from loca­tions around the world, pri­mar­i­ly the Unit­ed States and Switzer­land, as Google Earth sees them, each image cap­tur­ing a fleet­ing moment before the soft­ware inevitably receives new data, cor­rects itself, and restores the illu­sion of coher­ence.

The dis­cov­ery was entire­ly acci­den­tal: “The first time I dis­cov­ered this kind of error, I was stunned. I had com­mis­sioned paint­ings from Xia­men, Chi­na, for anoth­er art project. Out of curios­i­ty, I searched for the city on Google Earth and saw an image of a bunch of build­ings upside down or even com­plete­ly dis­tort­ed. … I start­ed look­ing for oth­er exam­ples. I found the most spec­tac­u­lar ones near roads and bridges. That’s where the risk of incom­pat­i­bil­i­ty between the 3D mod­el and the satel­lite pho­to is great­est. I would some­times sit for hours in front of my screen fol­low­ing roads, know­ing that I would even­tu­al­ly find a hilly sec­tion with a bizarre shape,” explains Val­la.

In anoth­er inter­view, Val­la explains the rea­son for this being due to “two com­pet­ing visu­al inputs here—the 3D mod­el and the map­ping of the satel­lite photography—and they didn’t match up. The com­put­er is doing exact­ly what it’s sup­posed to do, but the depth cues of the aeri­als, the per­spec­tive, the shad­ows, and the light­ing were not align­ing with the depth cues of the 3D Earth mod­el. I fig­ured that this was not a unique sit­u­a­tion in Google Earth, and I start­ed look­ing at obvi­ous sit­u­a­tions where the depth cues would be off—bridges, tall sky­scrap­ers, canyons. Soon I noticed the pho­tos being updat­ed, and the aer­i­al pho­tographs would be ‘flat­ter’ (tak­en from less of an angle), or the shad­ows below bridges would be more mut­ed. Google Earth is a con­stant­ly chang­ing, dynam­ic sys­tem, so I had to cap­ture these spe­cif­ic moments as still images.”

Google Earth infers height and form from shad­ows, cam­era angles, and pho­togram­met­ric data. When those cues mis­align, par­tic­u­lar­ly around bridges, ele­vat­ed roads, steep canyons, and dense urban struc­tures, the soft­ware “wraps” the image tex­ture over the ter­rain mesh, pro­duc­ing grotesque dis­tor­tions. Roads sag. Rivers appear to climb, fly­overs melt into ravines, and sub­ur­ban hous­es fuse with the trees behind them. The effect inten­si­fies when man-made struc­tures are exclud­ed from the ter­rain mod­el, leav­ing only nat­ur­al topog­ra­phy beneath. The more oblique the source pho­to­graph and the deep­er the under­ly­ing land­scape, the more extreme the defor­ma­tion.

The evolv­ing nature of these maps echoes the evo­lu­tion of paper car­tog­ra­phy, where once-unknown ter­ri­to­ries were com­piled from total hearsay and took on more accu­rate forms over the cen­turies as infor­ma­tion became more read­i­ly avail­able. Here, in a dig­i­tal equiv­a­lent, we wit­ness Google’s under­stand­ing of the Earth slow­ly con­verg­ing toward some­thing we can more eas­i­ly rec­og­nize as our own, but until then, we are wit­ness­ing a sur­re­al glimpse of what the world is thought to be, but not accord­ing to any human mind mind.

Because Google Earth’s land­scapes are gen­er­at­ed by arti­fi­cial intel­li­gence draw­ing from mul­ti­ple aer­i­al and satel­lite inputs, the project rais­es broad­er ques­tions about how place is rep­re­sent­ed in the dig­i­tal era. Oth­er vir­tu­al worlds, par­tic­u­lar­ly in video games, often recre­ate real loca­tions at the expense of scale or accu­ra­cy, com­press­ing cities and omit­ting unnec­es­sary zones by design. But this sim­u­lacra is cre­at­ed so as a cre­ative choice, both for the sake of flow and to cut down on devel­op­ment time, as these are prod­ucts made by human hands. With Google Earth, how­ev­er, “the soft­ware edits, re-assem­bles, process­es and pack­ages real­i­ty in order to form a very spe­cif­ic and use­ful mod­el” and this cre­ates inad­ver­tent lim­i­ta­tions in the rep­re­sen­ta­tion of a giv­en space.

Cru­cial­ly, these images are tem­po­rary. As Google updates its datasets with flat­ter aer­i­al pho­tog­ra­phy, sub­dued shad­ows, and refined 3D mod­els, the anom­alies dis­ap­pear. “Because Google Earth is con­stant­ly updat­ing its algo­rithms and three-dimen­sion­al data, each spe­cif­ic moment could only be cap­tured as a still image. I know Google is try­ing to fix some of these anom­alies too. I’ve been con­tact­ed by a Google engi­neer who has come up with a clever fix for the prob­lem of droop­ing roads and bridges. Though the change has yet to appear in the soft­ware, it’s only a mat­ter of time,” Val­la notes.

His prac­tice is there­fore as archival as it is aes­thet­ic. Val­la pre­serves a dis­ap­pear­ing phase in the evo­lu­tion of machine vision. In doing so, he is also cre­at­ing an active chron­i­cle of how the sim­u­lat­ed world with­in is being shaped, arguably a con­cur­rent time­line with our own real world, but one where land­scapes shift and mate­ri­al­ize much more dras­ti­cal­ly than the ones in our real­i­ty. In addi­tion, this allows us, as Val­la sees it, “focus our atten­tion on the net­work of algo­rithms, com­put­ers, stor­age sys­tems, auto­mat­ed cam­eras, maps, pilots, engi­neers, pho­tog­ra­phers, sur­vey­ors and map-mak­ers that gen­er­ate them”

In an essay for Rhi­zome, Val­la dis­cuss­es Google’s “Uni­ver­sal Tex­ture,” a pro­pri­etary sys­tem that maps an immense, con­tin­u­ous­ly updat­ed pho­to­graph­ic col­lage onto a three-dimen­sion­al mod­el of the globe. “The Uni­ver­sal Tex­ture promis­es a god-like (or drone-like) unin­ter­rupt­ed nav­i­ga­tion of our plan­et—not a tiled series of dis­crete maps, but a flow­ing and flu­id expe­ri­ence.” This aspi­ra­tion mir­rors the expan­sive claims made by con­tem­po­rary tech­nol­o­gy com­pa­nies, such as Ope­nAI, sug­gest­ing a form of plan­e­tary knowl­edge and inter­con­nect­ed­ness that is total, seam­less, and con­sum­able, with one com­men­ta­tor sug­gest­ing this pre­sen­ta­tion reframes the earth as itself a prod­uct to be con­sumed.

Val­la fur­ther com­pares this seam­less­ness to the inven­tion of the esca­la­tor: “No inven­tion has had the impor­tance for and impact on shop­ping as the esca­la­tor… the esca­la­tor accom­mo­dates and com­bines any flow, effi­cient­ly cre­ates flu­id tran­si­tions between one lev­el and anoth­er, and even blurs the dis­tinc­tion between sep­a­rate lev­els and indi­vid­ual spaces.” To Val­la, akin to the infi­nite scroll of social media (what we would now call the doom­scroll), there is no start or fin­ish to the esca­la­tor, and sub­se­quent­ly there is no start or fin­ish to the flow of infor­ma­tion, and no dis­tinc­tion of where it comes from, be it the pub­lic, Google’s employ­ees, or gov­ern­ment agen­cies. Instead, there is an inter­minable flow of data points and a pro­gram cre­at­ing an image using only the best ones. It deliv­ers what would seem to be the ulti­mate per­spec­tive: fric­tion­less, omnipresent, and always in day­light.

How Google deter­mines which images are “best” remains opaque. Angle, sea­son, and light­ing all play a role. There is, notably, no night in Google Earth. Clouds are sys­tem­at­i­cal­ly removed. High-con­trast, clear-weath­er imagery is favored. The out­come is a world opti­mized for leg­i­bil­i­ty rather than fideli­ty, a sta­tis­ti­cal­ly curat­ed Earth that sup­press­es atmos­pher­ic noise, tem­po­ral dis­rup­tion, and visu­al ambi­gu­i­ty.

A great essay address­es this fact, stat­ing that rough­ly every part of the world you see is from a pho­to tak­en dur­ing the day and in spring­time. The essay goes on to under­line the impor­tance of “smooth­ness” as a fac­tor, i.e. images which lack ele­ments that may play a role in cre­at­ing a flawed mod­el, such as clouds, smoke, or var­i­ous dis­as­ters, yet this often means that many unfor­tu­nate real­i­ties are ignored or “air­brushed” from the record, one notable exam­ple is in the wake of Hur­ri­cane Kat­ri­na, when Google opt­ed to use pre-hur­ri­cane pho­tos of the south­east­ern US. In eras­ing these real­i­ties, the world is still imper­fect, only this world’s imper­fec­tions are due to a lack of per­fect­ly accu­rate data for the pro­gram to prop­er­ly process, with the implied expec­ta­tion that even­tu­al­ly these imper­fec­tions will be total­ly done away with, a promise that could not be rea­son­ably made for the real world we occu­py.

Seen from the van­tage point of more than fif­teen years since its incep­tion, Valla’s project antic­i­pates con­tem­po­rary anx­i­eties sur­round­ing AI’s role in cul­ture and knowl­edge pro­duc­tion, facets of our lives com­mon­ly asso­ci­at­ed with the human touch. Where paper maps were com­piled by hand, and the best routes and land­marks were often left to the sug­ges­tions of locals or those with lived expe­ri­ence of a giv­en area, this now comes as one of the first ele­ments of the human expe­ri­ence to fall to AI, long before any con­cerns that art could be tak­en over as well.

The images are not manip­u­lat­ed. There is no com­posit­ing, no retouch­ing. Their aes­thet­ic emerges entire­ly from auto­mat­ed process­es—“atyp­i­cal ephemera” pro­duced by advanced sys­tems at plan­e­tary scale. As Val­la observes, “There is very lit­tle direct human hand in these arti­facts… it’s about fram­ing them, allow­ing them to be seen, and reveal­ing a strange byprod­uct of these super-func­tion­ing sys­tems.” In time, these dis­tort­ed land­scapes may be remem­bered much like the myth­i­cal con­ti­nents and phan­tom islands of ear­ly car­tog­ra­phy.

By call­ing them “post­cards,” Val­la also ges­tures toward a form of arm­chair tourism: a world appre­hend­ed remote­ly, through a pol­ished and min­i­mal inter­face, engaged and under­stood through the com­fort of one’s own home, echo­ing the man­i­fes­ta­tions of opti­mistic glob­al­iza­tion from the 90s such as the world music boom and the Soft Colo­nial Wan­der­lust aes­thet­ic, only this time pre­sent­ed in a stream­lined, min­i­mal image of the earth as it is (or as it could be).

“Satel­lite pho­tos aren’t made to be beau­ti­ful,” Val­la notes, “They are beau­ti­ful by acci­dent.” His act of archiv­ing and high­light­ing util­i­tar­i­an car­to­graph­ic data trans­forms it into some­thing con­tem­pla­tive and uncan­ny, clos­er to sur­re­al land­scape paint­ing than to tech­ni­cal imag­ing. In this sense, Post­cards from Google Earth belongs to a broad­er tra­di­tion of glitch art and crit­i­cal media prac­tice, along­side artists who inter­ro­gate how soft­ware shapes per­cep­tion: Trevor Paglen’s sur­veil­lance imagery, Hito Steyerl’s essay on the “poor image,” or Mish­ka Henner’s use of Google Street View. Yet Valla’s con­tri­bu­tion is dis­tinct in its quiet­ness. These images are not spec­tac­u­lar fail­ures, but sub­tle rup­tures in a sys­tem designed to hide its own mechan­ics.

Google has acknowl­edged the project, and at times even respond­ed. The engi­neer that con­tact­ed Val­la also sought per­mis­sion to use his images in an inter­nal com­pa­ny pre­sen­ta­tion to help solve the very prob­lems the artist has attempt­ed to doc­u­ment. In anoth­er inter­view, he states that Google’s only offi­cial response was their request that he retain the logo and copy­right infor­ma­tion in his screen­shots. While the com­pa­ny has worked to rem­e­dy the sur­re­al­ist land­scapes, notably in high-inter­est loca­tions such as the Gold­en Gate Bridge and the Hoover Dam, more remote areas often retain their warped geom­e­try to this day, despite new satel­lite images hav­ing been over­laid onto them in the inter­ven­ing years.

Despite what the view­er may believe, Val­la main­tains the old adage that the pur­pose of a sys­tem is what it does and that the images we are see­ing are not glitch­es, for a glitch would imply that the pro­gram is not run­ning the way it should. “They are the absolute log­i­cal result of the sys­tem. They are an edge condition—an anom­aly with­in the sys­tem, a non­stan­dard, an out­lier, even, but not an error.” In the end, Post­cards from Google Earth is less about error than about expo­sure. It makes vis­i­ble the hid­den infra­struc­ture of con­tem­po­rary vision: algo­rithms, sen­sors, pilots, sur­vey­ors, servers, and soft­ware, all col­lab­o­rat­ing to pro­duce the illu­sion of a sin­gle, sta­ble, know­able Earth.

And in those rare moments when the illu­sion frac­tures, we are offered some­thing that seems more like a sober­ing reminder of our own time rather than a strange nov­el­ty, that many of the new tech­nolo­gies we engage with on a near-dai­ly basis, despite draw­ing from seem­ing­ly lim­it­less resources and infor­ma­tion, are not omnipo­tent and cer­tain­ly not per­fect, even if they are run­ning exact­ly as they were pro­grammed.