Let's look at the teeth.
When the image saturation is increased to the maximum, the teeth in AI-generated portraits become very strange, with blurred edges.
The overall color of the image remains normal, but the microphone part is particularly odd.
In contrast, ### real human photos should look like this.
The teeth are clear, and the color blocks in the image are uniform and consistent.
This tool is now available for everyone to try with their photos.
Even a single frame from an AI-generated video can't escape this method.
Photos without visible teeth also reveal problems.
By the way, this tool ### was created by Claude. Using AI to crack AI, a curious loop.
To be fair, the recent hyper-realistic AI portraits have sparked quite a discussion. For example, a viral set of "TED speaker videos" doesn't feature a single real person.
It's not just faces that are hard to distinguish; even AI's previous weakness - handwriting - can now completely pass as real.
More crucially, generating such AI portraits is not expensive. It can be done in as little as 5 minutes, at a cost of $1.5 (about 10 RMB) per 20 seconds.
This has prompted netizens to start an AI fact-checking competition.
Nearly 5,000 people came to discuss which of these two images is a real person.
The reasons given are varied. Some noticed abstract details in text and patterns, while others felt the person's gaze was empty...
The patterns in portrait generation by the most advanced AIs are gradually being figured out by people.
01 It's already difficult to distinguish without looking at details
In summary, adjusting saturation might be the quickest method of identification currently.
AI group portraits are exposed even more thoroughly under this method.
However, it has one problem. If the image has been compressed using the JPEG algorithm, this method may fail.
For example, this photo is confirmed to be a real human photo.
But due to image compression and lighting issues, the person's teeth are also a bit blurry.
So netizens have listed more methods to determine if a portrait is AI-synthesized.
The first method, simply put, relies on human knowledge to judge.
Since AI learns images differently from humans, it inevitably can't 100% grasp visual information from a human perspective.
The result is that AI-generated images often contain ### inconsistencies with the real world, which provides a starting point for image identification.
Using the image at the beginning as an example.
Overall, the person's skin is too smooth, with no visible pores. This overly perfect feature actually increases the sense of unreality.
Of course, this "sense of unreality" doesn't entirely equate to "fake", as images processed with skin smoothing also don't show pores.
But this isn't the only factor for judgment, as AI may leave multiple inconsistencies with common sense in an image.
In fact, just by looking at the details of this image, you can see a rather obvious feature - ### the strange connection method of the hook above the badge.
And the microphone that reveals flaws in high saturation mode can be seen with the naked eye when enlarged.
More subtle is that the position of a few strands of hair at the end is very unreasonable, but such features might require Leeuwenhoek-level vision to see.
However, as generation technology improves, finding features is becoming increasingly subtle, which is an inevitable trend.
Another method is to look at the text. Although AI is gradually overcoming the problem of "scribbling" in font rendering, it still has some difficulties in correctly rendering text with actual meaning.
For example, some netizens noticed that on the badge worn by the person in the photo, the last line under the Google logo has the letters "CA", representing California, USA, with a long string before it that should be a city name.
But in reality, there is no city in California with such a long name.
Besides these details of the objects themselves, information like lighting and shadows can also be used to judge authenticity.
This image is extracted from a video, which also contains this frame.
To the right of the microphone, there's a very strange shadow. This shadow corresponds to one of the person's hands, clearly showing AI's inadequate processing here.