Is This How ChatGPT Actually Perceives Me?

A grumpy, clean-shaven individual with nothing happening in life apart from a roommate and a decorative artificial pumpkin? Is that how ChatGPT envisions me?

You might have seen the recent viral trend involving ChatGPT images: “Produce an exact duplicate of this image, leave everything unchanged.” The concept revolves around inputting an image into ChatGPT (using an image generation tool) and then repeating that process 99 additional times — using the previous image as the new input each time. The outcome? A gradual but consistent transformation that grows stranger with each iteration.

Individuals have been experimenting with their own pictures, resulting in outcomes that range from amusing to truly odd. One notably popular instance is a transforming image of Dwayne “The Rock” Johnson that has been circulating on Reddit and X (formerly Twitter). It resembles a visual game of telephone — and by the 100th iteration, the image frequently becomes unrecognizable.

As Rick once remarked in a memorable episode of Rick & Morty: “And far enough down the line… there be monsters.”

Critics of AI argue that this trend reveals a significant issue: when AI models generate outputs based on their own results, quality diminishes — a kind of digital inbreeding that results in gibberish. Is that genuinely what’s occurring?

Intrigued, I decided to test it out myself. I utilized the GPT-4o model on ChatGPT and requested it to duplicate a picture of me. However, the chatbot wasn’t cooperating. “I can’t create an exact replica of the image,” it replied. “But I can generate a new, highly similar version using AI based on its content.” I pressed further, yet the answer remained unchanged: “I understand your request, but I can’t produce an exact replica of a photo.”

Eventually, I bypassed the limitations by employing the OpenAI API and a thoughtfully constructed prompt. That finally succeeded — and the results were what one would anticipate from this trend: a slow, surreal distortion of the initial image.

I’ve reached out to OpenAI for insights regarding this trend and will amend this article if I receive a response. But from my experience, it appears that ChatGPT isn’t genuinely attempting to create flawless copies of images — at least not through the conventional interface.

So, is the AI ouroboros legitimate? Perhaps. But it’s also plausible that this trend reflects more on the constraints of current tools rather than an inherent flaw in AI itself.

Disclosure: Mashable is owned by Ziff Davis, which filed a lawsuit against OpenAI in April, alleging the company used copyrighted material to train its AI models.