In 2023, I created a series of photographs exploring the possibilities of AI, called "The Most Likely Image".
Each image originated from a simple instruction. For most of the images I used something like:
“contemplate the mystery of birth and death, while surrounded by scenes of civil war.”
The resulting images were rich with references to ‘classic’ photography, often echoing the styles of well-known photographers. Yet, due to glitches and imperfections in the software, they also felt uncanny—surprising, ambiguous, and at times unsettling.
Today, in 2026, it has become difficult to produce anything remotely interesting in this way. Most models now generate images that align closely with what one might expect from such a prompt: mysterious, atmospheric, and ultimately clichéd, filled with smoke, skulls, and a baby. Technically flawless, perhaps, but rarely thought-provoking. Three years ago, prompts functioned like seeds for hallucination; today, they operate like tools of art direction.
Still, I set out to recreate the original images using current software. The only method that seemed to work was to ask one chatbot to describe the original image in as much detail as possible, and then, crucially, to ask another chatbot to generate a new image based on that description.
The results look like perfect, flawless photos. Still, I’m not sure what to make of them. Have a look..