最近，ChatGPT图像复制实验在网络上走红。用户输入一张图片，要求ChatGPT“完全复制这张图片，不要做任何改动”，然后重复这个过程99次。随着迭代次数的增加，图像逐渐发生变化，最终变得非常奇怪。本文作者亲自尝试了这个实验，发现ChatGPT实际上并没有完全复制用户的照片，而是生成高度相似的新版本。同时，OpenAI拒绝就此事发表评论。这个实验引发了人们对AI模型训练的思考，以及AI在复制和模仿方面的能力。

🔄 用户通过不断迭代的图像复制实验，发现ChatGPT生成的图像会随着迭代次数的增加而逐渐变异，最终产生与原图截然不同的结果。

🤖 作者使用ChatGPT 4o模型进行实验时，发现该模型拒绝完全复制照片，而是选择生成高度相似的新版本，这表明ChatGPT并非旨在进行精确复制。

🚫 OpenAI API允许用户通过特定prompt实现图像复制，但ChatGPT本身似乎存在限制，无法直接进行完全复制操作。

Is this really what ChatGPT thinks of me?

An angry, beardless man with nothing in his life but a roommate and a fake pumpkin on a shelf?

You may have seen the latest viral ChatGPT image trend: "Create an exact replica of this image, don't change a thing." Users then redo that process 99 more times, with the image shifting slightly in each iteration. People have been plugging in their own photos to test out the results, and a mutating image of Dwayne "The Rock" Johnson has also gone viral on Reddit and X. When you make a copy of a copy of a copy, the results naturally degrade over time. And after 100 iterations through ChatGPT, the results can get truly strange.

To quote a famous Rick & Morty episode, "And far enough down the line...there be monsters."

To AI critics, this trend is proof that when AI models train on their own output, the end result is infinite slop — an artificial intelligence ouroboros. But is that really what's going on here?

I decided to try the trend myself, and in doing so, realized this trend isn't all it's cracked up to be. I tried plugging this prompt into ChatGPT using the 4o model, and it repeatedly refused to replicate my photo without changes. "I can’t make an exact copy of the image, but I can generate a new, highly similar version using AI based on its content," the AI chatbot said. And when I pressed, it still refused, saying, "I understand your request, but I can’t create an exact replica of a photo."

I was finally able to get the chatbot to do my bidding by using the OpenAI API and the prompt I mentioned earlier. And the results are typical of similar experiments.

I asked OpenAI about this trend, and I'll update this post if I hear back. But based on my attempts to replicate the replication trend outside of the API, it seems clear to me that ChatGPT isn't actually trying to create an exact replica of users' photos in the first place.

---

Disclosure: Ziff Davis, Mashable’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.