You’ve probably seen the fake screenshots floating around. Tim Cook as CEO of Xiaomi. A SpaceX acquisition of Cursor. Claude opening to China. You laughed. You shared. You didn’t think twice.
But here’s the thing: some of them are real. And you can’t tell which ones anymore.
That’s not a glitch. That’s the point.
GPT-Image-2 has done something terrifying. It’s not just that it can generate fake screenshots—it’s that it can generate them so perfectly, so flawlessly, that the very concept of ‘seeing is believing’ is now dead. A single sentence. That’s all it takes. “Generate a tweet from Elon Musk announcing a new Tesla model.” Boom. Done. The font is right. The layout is right. The logo is right. The shadows are right. Every single pixel matches the real thing.
I tested this myself. I compared the fake ‘Luo Yonghao tweet’ against a real X screenshot. The fake used a Chinese font. The real one used a different system font. That tiny difference—a font mismatch—is still a tell. But for how long? The model is learning. It’s already matching logos, avatars, and interface elements perfectly. We are racing against a machine that is getting better at lying every single day.
This isn’t new. Fake screenshots have been a problem since the dawn of the internet. In 1999, a Bloomberg report was faked so convincingly that a company’s stock surged 30% before anyone realized it was a hoax. In 2018, a fake ‘Pony Ma WeChat screenshot’ fooled even Tencent’s own tech journalists. But those fakes had costs. They required skill. They required Photoshop knowledge. They required hours of meticulous work to get the spacing right, the fonts right, the shadows right.
GPT-Image-2 has cut that cost to zero.
Think about what that means. Every screenshot you see from now on is suspect. Every tweet, every news article, every company announcement—it could be real, or it could be a single prompt away from a lie. The burden of proof has shifted entirely. You can no longer trust your eyes. You must now trust the URL. The text. The metadata. The things you can click on, not the things you can see.
And here’s the uncomfortable truth: we can’t stop this. We can’t cripple the model. We can’t hide the technology. The genie is out of the bottle, and it’s not going back in. Every attempt to ‘restrict’ AI image generation will be bypassed by the next open-source model. Every safety filter will be broken by a clever prompt. The tools are getting better, faster, and more accessible. It’s a race to the bottom of trust.
So what do we do?
The answer is not technical. It’s punitive. We can’t stop people from making fake screenshots, but we can make them regret it.
We need to shift the focus from restricting the AI to punishing the user. If you create a fake screenshot that causes real harm—stock manipulation, reputation damage, public panic—you should face severe consequences. Not a slap on the wrist. Not a fine. Real, credible deterrence. The kind that makes someone think twice before typing ‘Generate a fake news headline.’
This is the only way. The tool is too powerful. The user is too fallible. The only thing that can bridge that gap is fear of punishment.
I wrote this article using GLM-5.1. I’m admitting it because I want you to know that even the content you’re reading right now might be partially generated. Trust is dead. We killed it with a text prompt.
The question is: what are you going to do about it?
FAQ
Q: Can't we just add watermarks or metadata to AI-generated images?
A: Watermarks are trivial to remove. Metadata can be stripped. The cat-and-mouse game is unwinnable at the technical level. The focus must be on deterring the human abuser, not the AI.
Q: What's the practical implication for me as a regular user?
A: You must adopt a zero-trust policy for every screenshot you see. Verify the URL. Check the text. Cross-reference with official sources. Your eyes can no longer be trusted. This is the new normal.
Q: Isn't this just fear-mongering? Won't people adapt?
A: Yes, people will adapt. But the adaptation requires a fundamental shift in how we consume information. The 'old' way of trusting visual evidence is over. The new way requires active verification. That's a massive cognitive load, and it's precisely why punitive measures are necessary.