If it were AI, would you complAIn?

Fabio Galiana Martínez

Generative Artificial Intelligence is capable of creating new content or ideas based on other existing content or ideas. Generative AI for images, in particular, is improving so quickly that it’s becoming increasingly difficult to distinguish between its creations and real photographs.

According to Mónica Ballesta, a Department of Systems Engineering and Automation professorat the Miguel Hernández University of Elche (UMH), image generation through AI is based on deep learning models. One example of this is Generative Adversarial Networks (GANs), which consist of two networks that compete against each other. One tries to “deceive” the other while learning in the process. This way, it can generate synthetic images with such realism that they are difficult to distinguish from real images.

In general, these types of models work with a large number of images, from which they learn complex patterns such as shapes, textures, styles, etc. The more images they are trained on, the better the results can be. On the other hand, the internal structure of these networks consists of different layers interconnected through which the information they are fed passes.

A determining factor in the performance of AI tools, says Ballesta, is the prompts (requests made by the user to the AI). Depending on how precise we are when making a request, the AI will provide better or worse results. However, there are more variables involved in the process. Those images that are more demanded by users, such as those of attractive people, are closer to reality than those that are less requested. This leads to one of the disadvantages of AI: bias.

Let’s imagine a painter with no artistic aspirations who only paints portraits by commission. His patrons are all wealthy people in good health and well-dressed. Due to professional habits, if the artist wants to paint something else, it’s likely that he will tend to replicate the appearance of the people he has painted for years. This is what happens with AI: it prefers to generate images it is used to generating and has more examples of. Therefore, if the user requests and image banks are biased, its creations will be too.

Although AI has made great advances and is used in many fields (advertising, multimedia content, etc.), it’s expected that the level of realism and detail will be higher in the not-so-distant future. Furthermore, the images it generates have flaws that we can identify if we look closely.

Generative AI and People

The people created by AI often have malformations in their extremities, such as more than five fingers on each hand, and impossible hairstyles that mix with one another. On one hand, their skin is too perfect, as if they were using the ‘beauty mode’ on a mobile camera at full power. On the other hand, their eyes have no soul and show no emotion. Regarding animals, their appearance tends to be too ideal, as if, instead of real animals, they were models designed to resemble the mental image we have of them.

Generative AI and Inanimate Objects

Unlike what happens with images of people or animals created with AI, those containing landscapes or inanimate objects have fewer flaws and may go unnoticed. A trick to detect them is to pay attention to whether the image includes text, as AI is not a good graphic designer. If it is well-trained, it may create very simple and brief text, but most of the time it will fail in either the form or content: made-up words, spelling mistakes, blurry letters, etc.

Generative AI and Scientific Illustrations

Another example of limitations of AI-generated images would be scientific illustrations: the generated images may look correct to the general public, but they may not be scientifically rigorous at all. In fact, they may contain significant errors, such as drawings and texts with no meaning.

Generative AI and Geometry

Symmetry and geometry are also not its strong suit. It is common for it to generate images with impossible perspectives or buildings with nonsensical architectures, such as columns that do not connect to the ground. Distant figures, whether humans or animals, become distorted when zoomed in. In summary, as the saying goes, «the devil is in the details.»

Put your knowledge into practice with the following game, where you’ll have to identify which image in each pair is real and which one is generated by artificial intelligence.

In collaboration with the Spanish Foundation for Science and Technology – Ministry of Science and Innovation.

Did you like these activities? You have many more available, for young and old, in the special ‘Science in your living room’