Now, it's really getting harder and harder to distinguish between AI images.

Given a few seconds, can you tell which of the four images below is AI-generated?

Do the question first, don't hesitate to draw a peek at the answer!

Actually, there are only hereBottom leftIt's a real photo. I don't know if you guessed it right, anyway, the friends in the editorial department think it's very difficult.

As it is, AI images have become increasingly difficult to distinguish between real and fake, and even many AI detection tools have failed.

Let's put it this way, some of the pictures generated before are comic-style, some have strange limbs and facial features, and unreasonable backgrounds, in short, they are full of loopholes, and sometimes they are quite scary.

But a moment ago, GPT-4o was upgraded, and the Wenshengtu ability of the large model was directly superb. For example, the "selfie" in the upper right corner of the gang is generated by the following prompts:

The large model can even understand abstract requirements such as "mediocrity", "carelessness", "ambiguity", "overexposure" and so onThe resulting picture is like a casual shot in our lives, and there is no sense of disobedience at all.

Specifically, how these models make AI graphics look real and fake, the official has not yet open-sourced their training architecture.

However, on the official website of OpenAI, we found some clues.

Officials say that when they train the model, they can make the model better understand the association between language and images. Coupled with the mysterious "post-training", the generated results can look smooth.

Therefore, when we give some abstract words, such as "careless", the model can know that the angle of the image should be a little crooked, the picture should be a little mushy, the expression should be natural, etc., and it can be perfectly displayed.

With the rapid development of technology, we really can't do anything about carbon-based organisms.

But what is even more desperate is that the experimental results show that this time the silicon base is also indistinguishable.

We first tried to see if the spear of the large model could break through its own shield. Not surprisingly, the original kind of fake AI image, like us, can be easily distinguished. But now, the same picture, tossed to Doubao and GPT, they both think it's a real selfie.

Bean Bao can't tell that this image is AI-generated

In addition to testing with large models, we also found two free AI image detectors that recommend the highest rankings.As a result, each of them collapsed in their own way.

We tested eight AI portrait images that were completely invisible to the naked eye. There are four of them that they agree on, but they agree that they are all real photos...

There are four more, and the opinions of the two detectors are completely opposite. I thought they were copying each other's homework, but now I don't doubt it, because this time the wrong questions are all different.

In short, it's just to do it

This is just a relatively simple portrait, the picture is focused on the front face of the person, and the background is relatively simple.

The next test of some complex scenes is even more miserable, with a large number of people or too detailed backgrounds, or even simple landscape pictures, which make the detector almost completely annihilated. If the detector is a little skeptical about AI selfies, it really believes them when it comes to these images.

It's just that if you can't recognize it, there is a detector that has accidentally injured and judged a real photo to be an AI image.

There is one thing to say, the sky of online lovers has fallen, and I really can't tell whether it is a photo or a photo in the future.P-pictures may leave traces, but the current AI pictures really make people wonder if this is not an Internet celebrity who is going to fall in love with me.

So why are AI detection tools not working now?

When searching, we found that although the development of Wensheng graph technology is like riding a rocket, AI image detection is still riding a convolutional neural network bicycle for so many years.

Since most tools don't open source their source code, we've found several AI image detection projects on GitHub as references.

We found that the architecture of these AI detection tools is still in the stage of dataset + convolutional feature recognition + classification.

Those who are familiar with computer vision may know that this process has been followed for N years: first label each image in the dataset with whether it is or is not AI-generated, and then leave the rest to the neural network to learn the corresponding image features of the label, and finally classify it.

AI mapping technology has been updated one after another, and all these tools do is label new AI graphs, add them to old datasets, and retrain them.Even the CvT-4 model used for one of the tools is already an old thing 0 years ago.

It can be said that the magic is one foot high, the road is one inch high, the technology itself has not been updated, and the accuracy rate certainly cannot go up.

CvT-13 architecture

Although there are some academic studies related to AI image recognition, the research speed, quantity, and attention are not comparable to those of large model Wensheng diagrams.

However, instead of time-consuming and laborious post-separation, it is better to solve the problem at the source.

For example, the C2PA organization, which is jointly advocated by major AI companies, encourages the development of relevant standards to make it easier to verify the source of information and avoid the proliferation of AI content.

Among them, OpenAI said that it will try to watermark the generated images. Google has also proposed synthID, which can embed digital watermarks into AI-generated text, images, videos, and audio. This watermark does not affect our perception, but it can be recognized by the software.

而且，在今年 3 月國家頒布的《 人工智慧生成合成內容標識辦法 》中明確表示，從 2025 年 9 月起，所有 AI 生成的內容都必須添加顯式或隱式標識。

So why do we have to distinguish between AI graphs? Isn't it a good thing that you can't tell the difference between technical power max?

The picture is really powerful, but we have to look at both sides of everything.Because when AI images shocked the world, news of using AI to commit fraud and crimes was still frequently exposed. The more realistic the AI, the higher the probability that we will be scammed.

After all, some people don't think about how to use AI to generate cute Ghibli-style images, but instead use the most realistic images to attack everyone's weakest points.

In general, it is now difficult for us to distinguish the authenticity of AI images on our own.

Whether it's identifying tools or tagging AI content at the source, the current technology is a bit backward, but the need is urgent.

In this way, differentiating AI content will be a constant battle. When major companies are engaged in sketch technology and show off their muscles, it is time to consider the upgrade of AI recognition technology.