In the wake of DALL-E, the artificial intelligence image generator designed by OpenAI, many tools using AI to produce images from text queries have emerged. The results have often been stunning, raising questions about our ability to tell right from wrong. These concerns are not new, they were already very present when deepfakes appeared a few years ago. Over time, AIs will undoubtedly grow and become more difficult to unmask. In this article, we offer you some tips for spotting images produced by AI!

1. Check if the use of an AI is indicated

While it may be tempting to hide that an image has been generated by artificial intelligence, some users still have the honesty to indicate the real source of their image. In some cases, the objective of the operation is to demonstrate the performance of the tool, in which case it is necessary to indicate on which software the image was produced. The first reflex to adopt is therefore to check if a credit is present on the image. If it is posted on a social network, this credit can be indicated in the comments of the publication. Some image generators even offer a symbol to sign their work. DALL-E, for example, adds a small multicolored watermark to the bottom right of images (see photo above). For its part, Craiyon integrates its logo in the same location.

2. Look for inconsistencies or distortions in the image

Artificial intelligences are far from perfect, and most of them still make gross mistakes. Certain elements of the image may be out of focus, blurry or out of proportion, and it is not uncommon to see faces without noses or 6-fingered hands depicted. It is also the human representations that seem the most difficult to reproduce: the teeth and (especially) the eyes are often distorted on the images or on the deepfake videos.

If you want to practice spotting faces created by AI, you can play the game which face is real, in which you have to guess which of the two faces actually exists. You will see, it is not very difficult!

3. Check if the image elements are consistent with each other

In many cases, artificial intelligences produce rather successful images, but on which the elements are poorly associated with each other. For example, different components of an image can take on slightly different drawing styles.

Here are some details that we suggest you analyze:

• Are the shadows and lighting realistic?
• Are the colors of the image consistent with each other?
• Do the different elements all have the same style or level of realism?

4. Use image recognition software

As with textual analyzes (following the emergence of ChatGPT), certain solutions are being put in place to detect whether an image has been produced by artificial intelligence. At this time, not all of them are efficient: the GAN detector from Mayachitra Inc, a company initially specializing in AI image recognition, still has a long way to go to be reliable. But these tools will benefit from being perfected by dint of being trained. In addition, many renowned companies, such as Microsoft or Intel, have been working for a few years on the recognition of deepfakes, and their expertise could prove useful in the development of new tools. For its part, the company Hive AI offers companies an API aimed at detecting and moderating works of art generated by AI.

5. Distinguish AI styles

A priori, artificial intelligence image generators are supposed to be able to reproduce a wide variety of styles. It is indeed possible to ask them to create an image in a particular style of illustration, or by reproducing the style of a great painter. Yet, it can be seen that, by default, the tools tend to adopt a particular style that may be recognizable to a trained eye. By producing your own images on the different generators, you will train yourself to identify registers of drawings or color palettes that come up regularly.

Discover AI image generators