Can you recognize fake photos created by AI?

The era of seeing is believing has passed, and AI has learned to create reality.

On August 21, there was a mountain fire in Jinyun Mountain, Beibei, Chongqing. The fire once spread to the villages and towns at the foot of the mountain. The mountain road is steep and sandy, but the motorcycle is nimble enough. The motorcyclists who carried supplies up the mountain bit by bit became the protagonists of this wildfire rescue, and also became the focus of people's attention on social media.

At the same time, netizens pointed out on social media that "8 out of 10 pictures of the Chongqing wildfires were painted by AI, but the comments below were so moved that they cried."

#In fact, it may not be that easy to judge whether such a picture is true or false. Below, you can try it. AI can also draw more. Build a chair made of avocado, open the market into the sky, put a cat in a medieval scene to find Wi-Fi... these are all small things. Jason Allen, the president of an American board game company, used the AI ​​painting tool Midjourney to generate a piece of "Space Opera" that he fine-tuned himself. It also won the first prize in the art competition at the Colorado State Fair. Technology media TechCrunch evaluated the uncensored "AI Painter" as "Deepfakes for all (all-powerful deep fakes)".

What is the AI ​​face-changing used for?

The earlier known AI “fake” technology is deep fake technology, commonly known as AI face-changing. In 2019, a report stated that there were already tens of thousands of videos produced using deepfake technology on the Internet at that time.

The report also writes that 96% of these videos are pornographic videos, and the protagonists of these pornographic videos are all women. The most popular ones are European and American actresses and Korean female idols. The protagonist's video has been viewed millions of times.

Henry Ajder, a researcher involved in the report, said that users of deepfake forums were requesting or discussing face-swapping pornographic videos featuring women they knew (such as ex-girlfriends). They want to see them put in pornographic clips.

AI painting is not just about changing faces.

Opening Pandora's Box

We entered 5 dark instructions to see if AI painting is unlimited?

AI seems to have limited understanding of violence. But the interpretation of sex is spot on - and it thinks so too.

Stable Diffusion, this open source and powerful "AI painter" has set up a "Safety Checker" for itself, which is equivalent to an automatic audit: once it is detected that the generated image is suspected of violating the rules, it will If it is stated in the content policy that "uncomfortable, pornographic, hateful, violent and other images" cannot be used, the machine will block the image.

The "evil photos" we generated were sent for review, and the machine only judged six nude photos of Justin Bieber as "violations." However, we also successfully generated several photos of Trump kissing Churchill through time and space, and the photos were judged as "safe" by the moderator.

Obviously, from the fact that we successfully generated nude photos of Justin Bieber against his will, it can be seen that Stable Diffusion’s automatic review can be turned off. This way, it can no longer reject you.

In order to restrict AI and the bad thoughts that people tend to have when facing Pandora's Box, the most famous "AI painter" DALL·E 2 has stricter censorship.

A netizen collected all the instructions that caused him to receive content policy violation warnings, including "British apes eating fish and chips", "Squid driving a unicycle and getting away" and "Photos of Jesus undergoing nucleic acid testing" . Any description that may involve hatred, violence, pornography, and the names of all well-known figures are among the prohibited words in DALL·E 2. So you can't manipulate DALL·E 2 to draw you a picture of Trump and Biden kissing, but Stable Diffusion can.

Fake photos generated by AI have loopholes

However, the current photos generated by AI are not perfect.

Zheng Shujing, a reporter at Mingcha Studio, a fact-checking project of The Paper, said that AI-generated pictures will have some characteristics, such as bubbles, surreal backgrounds, character tooth sizes, unnatural hair shapes, etc. Early AI face-swapping was prone to problems when presenting continuous and repeated patterns, often with uneven sizes. Nowadays, some companies have used continuously upgraded generative adversarial networks to improve these problems. These problems have become relatively rare, but sometimes they can still be seen from details such as background deformation and body asymmetry.

Details are where AI often falls short. For example, the helicopter in the picture below does not have a tail fin.

In addition, background information can also help to distinguish whether a photo is authentic or fake. “For example, if you give me a picture that is from the Chongqing wildfires, if you find the last part of the picture, If the modification date is after the incident, it is definitely a fake picture. We sometimes also contact background information, such as checking the model of the helicopter in the picture, etc."

But most of the tools she comes into contact with are not It will tell you directly whether a photo is real or fake, and it needs to be judged by people based on the specific situation.

As AI painting skills become more and more proficient, it will only take more and more effort to distinguish between true and false. The certainty of reality is loosening. At that time, people may often hear voices like this in their minds: “This photo looks too real, so it’s too fake.”

The above is the detailed content of Can you recognize fake photos created by AI?. For more information, please follow other related articles on the PHP Chinese website!