In April 2022, artist Steph Maj Swanson asked an AI image generator to produce something as far from Marlon Brando as possible. After a couple of iterations, the system started generating the same woman over and over: an older figure with hollow eyes and triangular patches of rosacea on her cheeks. Swanson named her Loab.
Loab didn’t just appear once — it’s that she kept appearing. When Swanson combined Loab’s image with other pictures and fed them back into the AI, she’d resurface in the results, often in increasingly violent and disturbing contexts. Attempting to dilute her out of the system would work temporarily, but crossbreeding those diluted images would eventually conjure her back.
Swanson theorized that Loab occupies an isolated region of the AI’s “latent space” — the mathematical representation of everything the model knows — that happens to sit adjacent to gore and horror imagery. She’s become known as the first AI-generated cryptid, a ghost in the machine that emerges unbidden from the statistical depths of image generation.
Researchers say this is just a quirk of how these models work, not evidence of anything spooky. Others have pointed out that labeling a woman with a skin condition as inherently horrific says more about human bias in training data than about AI consciousness. Swanson agrees Loab isn’t “really haunted,” but notes that with so many people now sharing her image online, she’s becoming embedded in the training data of future models. If we wanted to get rid of her, it’s already too late.
Previously:
• Authors leaving AI prompts and responses in their novels
• I had AI write a memoir based on my life
• People get mad at writer who ‘loves AI’
Discover more from Earlybirds Invest
Subscribe to get the latest posts sent to your email.
