Janelle Shane (previously at Neatorama) works with artificial intelligence algorithms, and brings us some amusing experiments at her site AI Weirdness. Her Halloween experiment generated some neat Halloween images, but to get there, the algorithms made things up out of whole cloth, adding wild details that were not in the original data. This only reinforces the warnings we've seen about never trusting AI to tell the truth, or anything near the truth.
Shane took several innocuous, generic pictures and told the AI program Google Bard to describe it, specifically "Please describe this spooky Halloween scene." While there is nothing Halloweeny about the pictures, it spat out a plausible description of something very spooky, using a lot of imaginary details. Then she gave the generated text description to DALL-E3 as a prompt to generate an image. That program also veered widely from the data it was given, adding more Halloween details and deleting others. See what these programs did to totally change the original data because of the words "spooky" and "Halloween" in four images with descriptions (including a potato) at AI Weirdness.
Comments (1)