Midjourney, DALL-E, DeepAI — images created with artificial intelligence tools are flooding social media. Some carry the risk of spreading false information. Which images are real and which are not? Here are some tips.
In recent days, many of these images have gone viral: the arrest of Russian President Vladimir Putin, or Tesla CEO Elon Musk holding hands with General Motors CEO Mary Barra.
The problem is that both AI images depict events that never happened. And while some of these images may be funny, they can also pose real dangers in terms of disinformation and propaganda.
Photos showing the arrest of politicians like Vladimir Putin or former US President Donald Trump, can be verified fairly quickly by users if they check reputable media sources. Other images are more difficult, such as those in which the people in the photo are not that well-known.
However, AI tools make mistakes, even if they are developing rapidly. Currently, as of April 2023, programs like Midjourney, DALL-E, and DeepAI have their flaws, especially with images showing people.
The fact-checking team at Deutsche Welle (DW) has put together some tips that can help you assess whether an image is fake. But a word of caution: AI tools are developing so rapidly that these tips only reflect the current state of affairs.
1. Zoom in and look carefully.
Many AI-generated images look real at first glance.
That's why our first suggestion is to take a close look at the photo. To do this, look for the image in the highest resolution possible and then zoom in on the details.
Enlarging the image will reveal inconsistencies and errors that may have gone undetected at first glance.
2. Find the source of the image
If you're not sure whether an image is real or created by AI, try to find its source.
You may be able to see some information about where the image was first posted by reading the comments posted by other users below the image.
Or you can perform an image search. To do this, upload the image to tools like Google Image Reverse Search, TinEye, or Yandex and you can find the original source of the image.
The results of these searches may also show links to fact checks done by reputable media outlets, which provide further context.
3. Pay attention to body proportions
Do the people depicted have accurate body proportions?
It's not uncommon for AI-generated images to show inconsistencies when it comes to proportions, for example, with hands that are too small or fingers that are too long. Or the head and legs not matching the rest of the body.
This is the case with the photo above, in which Putin is supposed to be kneeling before Chinese President Xi Jinping. The kneeling person's shoe is disproportionately large and wide. The half-covered head is also very large and out of proportion with the rest of the body.
Read more about this at our dedicated fact-checking.
4. Beware of typical AI mistakes
Hands are currently the main source of errors in AI imaging programs like Midjourney or DALL-E.
People often have a sixth finger, such as the policeman to Putin's left in our photo above.
Or even in these photos of Pope Francis, which you've probably seen.
But did you realize that Pope Francis only appears to have four fingers in the right photo? And did you notice that his fingers on the left are unusually long? These photos are fake.
Other common errors in AI-generated images involve people with many teeth , or eyeglass frames that are strangely deformed , or ears that have unrealistic shapes, such as the aforementioned fake image of Xi and Putin.
Within seconds, image generators, such as Random Face Generator, create fake images of people who don't even exist. And even if the images look deceptively authentic, it's worth paying attention to unnatural shapes in the ears, eyes, or hair, as well as distortions in glasses or earrings, as the generator often makes mistakes. Reflective surfaces, such as helmet visors, also cause problems for AI programs, sometimes breaking up, as in the supposed arrest of Putin.
5. Does the image look artificial and polished?
In particular, the Midjourney app creates many images that look too good to be true.
Follow your gut feeling here: Could such a perfect image with perfect people really be real?
The skin of people in many AI images is often smooth and free of any irritation, and even their hair and teeth are perfect. This is usually not the case in real life.
Many images also have an artistic, shiny, and shimmering look that even professional photographers find difficult to achieve in studio photography.
AI tools often seem to design ideal images that are supposed to be perfect and please as many people as possible.
6. Examine the background
The background of an image can often reveal whether it has been manipulated.
Here too, objects may appear distorted; for example, street lamps.
In some cases, AI programs clone people and objects and use them twice. And it's not uncommon for the background of AI images to be blurry.
But even this blur can contain errors, like the example above, which is intended to show an angry Will Smith at the Oscars. The background isn't just out of focus, it looks artificially blurred.
Conclusions
Many images created by artificial intelligence can currently be hidden with a little research. But the technology is improving and mistakes are likely to become rarer in the future. Can they help us AI detectors like Hugging Face to detect manipulation?
Based on our findings, the detectors provide data, but nothing more.
Therefore, when in doubt, the best thing users can do to distinguish real events from fakes is to use their common sense, rely on reputable media, and avoid sharing photos.
spring: DW