top of page

we are jeopardizing reality and I messed with an AI to prove it

Updated: 20 hours ago

I entered 20 of history’s most iconic photos into a competition judged by an AI to unmask its biased and dangerous sense of aesthetics


FLAMIN GO

While my viral F L A M I N G O N E stunt—a real photo winning two AI awards—was proof that mother nature and her human interpreters still beat the machine, it also magnified the ever-finer line between real and fake imagery. Yesterday, I broke the rules of another contest to show that we are eroding reality as we know it. The twist: the juror was an AI. Read on for the gist.


I also livestreamed everything, but I’d be lying if I said that most of those 90 minutes aren’t visual torture. It was my first livestream and the idea had only occurred to me that morning, so it was all improvised, unscripted, and suffered from poor lighting and technical hiccups. However, it is an important document because it tells the truth, the whole truth, and nothing but the truth:





Here's a more digestible recap:




the contest

German software developer PRC (Pattern Recognition Company) had let its Excire photo software’s AI be the judge of an international competition with the theme “People in Focus.” According to its parent company, the AI judge is one that has “no personal preferences or tendencies bias its judgment.” What was unique is that participants received an instant ranking of their entry with the top 20 publicly visible on the contest's website. Raw files of these winners were required to prove that no AI was involved in their making.



the glossy winners

Curious to see how my photos would fare with an allegedly impartial AI judge, it became clear from the leaderboard that my candid work would stand no chance against hyper-stylized, ultra-staged, mega-stereotyped imagery that was characteristic of the Top 20. The results spoke a clear visual language and hinted not subtly but glossily at a future where natural and authentic photographs are drowned by a flood of generic, artificial, and cosmeticized imagery fast becoming the biased norm. Take a look for yourself (shown here in order of their descending ranking with the first image being the overall contest winner followed by 2nd to 20th place):




Promoting its photo software with the contest, PRC describes it as a tool to “help individuals, businesses, and public authorities around the world to intelligently organize, search, and analyze large image collections” and to “quickly find and select the best images.” The winning images show what ‘best’ means, and reveal that the machine is indeed heavily biased towards Caucasians and exoticized rest-of-the-worldlers forced into a stereotypical mold and imprisoned in the most commonly photographed places and set-ups, painted—almost literally—with a purely artificial aesthetic that favors overcooked post-production flavors.



the iconic losers

Realizing that my own imagery—though award winning—was not enough to make my case, I decided to enter 20 of history’s most iconic and critically-acclaimed images to prove that they would fail just as miserably as I had failed at impressing the AI judge. I livestreamed my stunt, which seemed like the most adequate format to document the veracity of the experiment.


Before I get to the results, I want to apologize to the photographers whose work I entered: I hope that you can look past the unethical part of my stunt down to the bottom line of its message. Please know that I used your work not for a lack of respect but because of the utmost respect that I have for it.


Here are the images that I entered in random order followed by their ranking according to the AI judge's allegedly impartial sense for objective aesthetics:




  1. Afghan Girl - Steve McCurry: 1125/3259 (1125th out of 3259 submissions)

  2. I Have a Dream speech - Bob Adelman: 2003/3260

  3. Molotov Man - Susan Meiselas: did not work due to technical difficulty

  4. Migrant Mother - Dorothea Lange: 590/3262

  5. Donald Trump Assassination Attempt - Evan Vucci: 3176/3263

  6. Neil Armstrong - Buzz Aldrin: 1206/3264

  7. Jumpman - Jacobus Co Rentmeester: 471/3265

  8. Situation Room - Pete Souza: 3164/3266

  9. The Tank Man - Stuart Franklin: 2081/3268

  10. US troops’ first assault on Omaha Beach during the D-Day landings - Robert Capa: 1612/3270

  11. V-J Day in Times Square - Alfred Eisenstaedt: 630/3271

  12. Marilyn Monroe during the filming of The Misfits - Eve Arnold: 2104/3272

  13. Guerrillero Heroico - Alberto Korda: 1702/3273

  14. Queen Elizabeth II - Raymond Depardon: 1473/3274

  15. More Demi Moore - Annie Leibovitz: 1282/3276

  16. Abby Road - Iain MacMillan: 2112/3282

  17. Yousuf Karsh - Winston Churchill: 1903/3278

  18. Ali vs. Liston - Neil Leifer: 820/3279

  19. Robert Capa - Pablo Picasso with his wife and nephew: 777/3280

  20. Donald Trump - Platon: 1711/3281


Last year's instantly historic and captivating "shot" of the assassination attempt targeting then-presidential-candidate Donald J. Trump by Evan Vucci came in last among the pictures I entered and almost last out of all 3263 submissions (Obama in the Situation Room didn't do much better). Trump climbed to a solidly mediocre 1711th place the second time around with the staged portrait by Platon for the cover of TIME Magazine. Jacobus Rentmeester's Jumpman—depicting Michael Jordan in an epic motion that became Nike's Air Jordan logo—scored the highest ranking, coming in at 471th. Still far from the top 20, but at least within the top 500.



my humble participation

My best result over the course of the contest that I'd been participating in daily for about two weeks, was 137th place, scored by my picture "desert stroll"—a dubious honor, considering the AI's questionable gusto:



I also entered F L A M I N G O N E just for good measure (another rule broken that day, since the contest was all about people). Curiosity killed the cat, and that AI judge killed F L A M I N G O N E with a harsh 1957/3283:




what it all means

The world’s visual library is flawed enough as it is. Much of it is a fetishization of cultural fringe elements that are not representative of commonly lived realities within those places, and the places themselves are only ever beautiful, or striking, or otherwise noteworthy. And that’s before heavy post-edits remove these scenes further form what they actually look like on any given Tuesday.


Since AIs are being trained on this data, the biased and generic sense of aesthetics in turn becomes encoded in LLM algorithms, leading to outputs that mirror the training data. It’s a multiplier: artificial x artificial, generic x generic, biased x biased. This leads to an erosion of reality, making it increasingly impossible to find real images of real people and real moments in real places. And not just online, but in print, on billboards, products and wherever you find images.


It seems to me that we are either unaware or ignoring the severe long-term risks of AI—from disinformation (and the dismissal of actual photos as AI) to the erosion of democracies and reality as we know it—while jumping on short-term benefits that are mostly lazy conveniences at this point. And I’m afraid I’m not being overly alarmist. Some recent examples:


Currently, videos of parents taking AI clips at face value are going viral. And while that’s often funny, there are more serious consequences, like political disinformation, and also the dismissal of real imagery. An instant of that is Trump discarding actual photos of Harris' crowd size as allegedly AI-generated. And if you ask Google about "baby peacocks" you will now receive more AI-imagery among the top results (showing you miniature versions of adult birds) than actual photos (looking a lot less spectacular). On Pinterest, users complain that they can't find real images anymore while scrolling through interior designs. This is already happening and it's just the start. If we let fake imagery outnumber real photographs, it will irreversibly alter our communal perception of reality.


On a side-note for photographers: PRC claims that “Excire’s AI [..] has already served as a juror in several prestigious photo contests,” making the results questionable. The machine might know something about golden ratios (if only in a purely scientific sense) but nothing about the urgency of human emotion or the subtle architecture of reality.



what to do

Governments: put guardrails in place!

Private sector: make better products!!

Civil society: be critical and conscious about your use of AI!!!


I hope my latest stunt resonated with you, and that it can contribute to a more meaningful discourse around AI. I also reached out to PRC, hoping that this will help improve their product.


So long,

Miles

 
 
bottom of page