AI-generated images: how to tell if they’re real or fake?

Images created artificial intelligence shocked the whole world, since in recent days supposedly photographs showing Pope Francisco with elegant white diamond or former US President Donald Trump being arrested by the police, all with a level of detail that has baffled millions of people because of how close to reality it is.

These images were created using artificial intelligence-based systems capable of processing user text descriptions to produce impressive results. While platforms like DALL-E, middle of the road And stable diffusion they started with various mishaps, such as adding too many fingers to the hands of their creations, the line between reality and fiction becoming more and more blurred with each new version available.

What to do when real and fictional are so similar?

Image generators using AI they don’t produce the same results, and depending on the platform, they can be more or less convincing. However, recent system updates such as Mid travel V5 they reach a level that makes them difficult—if not impossible—to distinguish from real photographs, some experts point out.

“It’s amazing what AI imaging devices can do. There has been a giant leap in terms of image processing capabilities over the last year or so.”It says S. Shyam Sundarresearcher Pennsylvania State University who studies the psychological impact of media technologies.

sundar points out that one of the factors explaining this leap in the capabilities of these systems is the greater number of images available to train them. AI, as well as advances in data processing infrastructure and interfaces that make the technology available to ordinary Internet users. As a result, artificial images are everywhere and can be “nearly undetectable”according to the expert.

The experiment demonstrated the ability of AI to deceive. Sophie Nightingalepsychologist Mr. Lancaster UniversityDigital technology specialist England co-authored a study that tested whether online volunteers could distinguish ID-card-like portraits created by an artificial intelligence system called StyleGAN2 and real pictures. The results were disappointing even at the end of 2021, when the researchers conducted the experiment.

“On average, people were more or less random. Basically, we’re at the point where it’s so realistic that people can’t reliably tell these synthetic faces from real faces, these are the faces of people that actually exist.”said nightingale. While humans have provided some assistance to AI by classifying images generated by StyleGAN2 in order to select only the most realistic ones, the author states that whoever wants to use such a program for nefarious purposes will probably do the same.

Artificial intelligence to distinguish artificial intelligence

Ironically, as image processing technology continues to improve, the best thing people can do to avoid being fooled by one AI system is another image processing system. AI, one that is already trained to detect artificial images. As AI imaging evolves, experts say, algorithms are better equipped than humans to detect some of the tiny pixel-scale quirks present in these robot creations.

In accordance with Yong Jae Leecomputer scientist University of Wisconsin-Madisoncreating these recognition programs AI it works just like any other machine learning task. “You collect one set of real image data and one set of AI generated images. You can then train a machine learning model to distinguish between them.”explains Lee.

Despite this, these systems have significant drawbacks. Most of these algorithms are trained on images from a specific AI generator and cannot identify fakes created by other algorithms. Most detectors also lack the easy to use interfaces that have attracted so many people to test drive systems. AI generating.

In addition, the detectors AI they will always have to struggle to keep up with imaging devices, some of which use similar detection algorithms but use them to learn how to make their false positives less detectable. “The battle between AI systems that generate images and AI systems that detect AI generated images will be an arms race. I don’t see a victory for either side any time soon.”It says Wael Abd AlmagidAssociate Professor, Department of Computer Science University of Southern California.

Then what do we do?

Obviously, given the progress of this type of technology, there are several strategies that we have to deal with the flood of images created from scratch or modified by AI. However, we can still adhere to certain conditions in order to successfully exit the routine photo inspection. Here are our recommendations:

Check the message. Look at the title, description, or user comments themselves when such a photo appears on Instagram or Twitter. While we can’t take internet discussion for granted, it can help us discover interactions with the photo’s creator or a more specialized eye. This includes a profile check, if necessary, to determine the type of content you are sharing.

Look at watermarks. Some services add brands to the image with some prominence so that the final product can be differentiated. DALL-E 2, for example, adds very small colored squares in one of the corners. However, this is not the case in Midjourney. In this case, you should refer to the overview of graphic details that can be used to “sign” the image.

Anomalies. This is the key to very popular photos created with Midjourney or other types of media. From a distance, the image can be very real, but it’s important to take a close look at some of the details that the AI ​​can’t quite pick out. This is more noticeable in texts – for some reason, different neural networks find it difficult to design a T-shirt with messages or posters with letters – but blurring of the background, some over-derived forms, glasses that are not correctly aligned, hair, teeth, arms and hands, mismatched earrings, unbalanced background. Everything is working.

We recommend you METADATA, an RPP technology podcast. News, analytics, reviews, recommendations and everything you need to know about the world of technology.