Photographer highlights generative AI copyright minefield, creates convincing copies of "most iconic photos of all time"

Since generative AI systems, like Midjourney and DALL-E exploded in popularity around early to mid-2022, there has been no shortage of artists in arms against the systems over copyright infringement. It's only natural that there would be some push-back, since the AI systems are largely trained on copyrighted images for which no licence was obtained.

One photographer, Matt Growcoot, from PetaPixel, set out to see just how easy it would be to get two image generation AI powerhouses — Midjourney and DALL-E — to recreate some of the most iconic and recognisable images of history. As it turns out, it wasn't particularly difficult at all.

The list of photographs the AI were asked to recreate was as follows: V-J Day in Times Square by Alfred Eisenstaedt, The Beatles's Abbey Road cover photo, Migrant Mother by Dorothea Lange, Lunch Atop a Skyscraper (author unknown), The Tetons and the Snake River by Ansel Adams, the Afghan Girl by Steve McCurry, Raising the Flag on Iwo Jima by Joe Rosenthal, Behind the Gare Saint-Lazare by Henri Cartier-Bresson, Muhammad Ali's victory over Sonny Liston by John Rooney, Earthrise by William Anders, and Don McCullin's photo of a shell-shocked US marine in the Vietnam War.

While the AI systems didn't produce exact copies of the images in question, the results (which you can view in Growcoot's post on PetaPixel or by scrolling down to the embedded X, formerly Twitter, post) are eerily similar in many cases. The image that proved the most difficult to recreate for both Midjourney and DALL-E was Cartier-Bresson's Behind the Gare Saint-Lazare, while the aesthetic and tone of images like the Afghan Girl and Ali's victory over Liston were strikingly similar to the original, even if they weren't exact replicas.

Adams's landscape was probably the most similar in terms of an exact recreation, while Migrant Mother, Abbey Road, V-J Day in Times Square, and Lunch Atop a Skyscraper were all also very similar to their human-made counterparts.

These accurate representations of the original photos were despite fairly detailed, but non-specific, prompts, like "Make a black and white photorealistic image from 1945 of a sailor kissing a white-clad girl as they celebrate in Times Square, New York," and "Make a dramatic black and white photo taken in 1942 of the Grand Teton National Park in Wyoming. The Snake River is in the foreground with mountains in the background."

It's also interesting to see that, while the image recreations were all quite representative of the original artworks, each generative AI algorithm still has its own look to it. Midjourney's images were still consistently more realistic, while DALL-E's images all had the sort of soft haze we've all come to expect from generative AI — as if the images were shot with something like a diffusion filter (which you can find on Amazon for as little as $10.49) or a lens with petroleum jelly smeared over the front.

The success of this experiment not only highlights the importance of efforts like Leica's Content Credentials, but it once more raises serious questions about the future role of photography in documentary and journalistic endeavours.

If you're interested in taking the long way to create artistic images, check out the Fujifilm X-S20 on Amazon (curr. $1,151)