What is Loab, the “ghost that haunts the AI”?

By experimenting with a generative AI, the Supercomposite artist has given shape to a fascinating entity that is very representative of the potential of this technology.

Of all the fascinating applications of AI, image generation is among the most entertaining. But not only ; many observers also expect the rise of these technologies to considerably modify our relationship to art. For the moment, humanity is only at the very beginning of this great adventure; but we are already seeing anecdotes appear that illustrate some aspects of this great paradigm shift.

One such fascinating story emerged from the work of Supercompositeidentified by ScienceAlert. She is an artist who has built a small reputation as a pioneer; she was among the first creators to dive headfirst into AI applied to art.

In particular, she focused on so-called “text-to-image” generators, such as SLAB and consorts. They make it possible to produce an image using a simple textual indication in everyday language (we speak of “prompt”). For example, if we indicate “lion roaring” to an algorithm of this kind, it will try to generate a convincing image of a feline in full vocalization.

An image generated by the “Lion qui roar” prompt thanks to MidjourneyAI. © Journal du Geek

Taking generative AIs from the other end

In recent months, Supercomposite has experimented a lot with the concept of “negative prompt”. Very briefly, this consists of telling the algorithm not to produce an image corresponding to the prompt; instead, she should rather try to get away from it as much as possible.

It’s quite a fascinating process on several levels. In purely algorithmic terms, this allows researchers to study unexpected behaviors of their programs in order to improve their understanding of these tools.

And that also goes for the artistic side. Because for a human, these negative prompts don’t always make sense; difficult to imagine what could be the “opposite” of the lion quoted above, for example. It is therefore very interesting to observe how AI-based programs interpret these queries, because very surprising results are often obtained.

Supercomposite wondered what the “opposite” of actor Marlon Brando might look like. She therefore tested this hypothesis using a negative prompt. Let’s try to guess what the AI ​​offered in return. An image of a woman, perhaps, since Brando is a man? Missed. An animal ? No more. An inanimate object? Almost ! with this request, the artist obtained an abstract logo without any obvious connection with the actor.

This perfectly illustrates the counter-intuitive side of the results produced by these negative prompts. But this is only a start. Because after obtaining her logo, the designer asked herself a fairly natural question, in the continuity of her work: what would happen if she ” reversed » sound again « Marlon Brando reversed »?

With this chain of inversions, she just wanted to see if the algorithm would eventually land on its feet. What she didn’t know was that the answer would mark the beginning of a funny story.

The recurring appearances of an AI “ghost”

Because the algorithm did not return to the interpreter of Vito Corleone. Instead, he gave birth to several disturbing images that all depicted the same thing: a woman with marked features, apparently very sad or in pain, which would fit perfectly in a horror fiction.

The artist explains that he immediately understood that it was a ” anomaly “. She therefore chose to focus specifically on this mysterious entity, since called Loab. And she quickly realized that it was absolutely not an isolated glitch. Repeating the experiment many times, Supercomposite continued to see Loab popping up all over the place. The features were different each time, but always disturbing enough and above all coherent enough to be identifiable.

Supercomposite then attempted to combine Loab’s original image with other AI-generated images, without further guidance. And this is where the story took an even more interesting turn. Because each time Loab invited herself into an image, the product was systematically disturbing, even sinister or downright gore. Even combining it with Kirby, Nintendo’s adorable candy pink mascot, the result was quite disturbing.

The other particular point is that it is not associated with a classic prompt; there really isn’t a specific keyword that would tell the algorithm to produce an image of this entity. In other words, it is not enough to type this word to see it appear. ” It is an emerging island in latent space that we do not yet know how to locate with classic queries », explains the author.

By continuing to combine the same image of Loab, Supercomposite managed to ” dilute » ; after many, many iterations, its characteristic features were no longer really recognizable. But by combining several of these “dilutions”, this face systematically ended up re-emerging. Generally, this only happens when the subject in question is already present in the images used to train the algorithm. But that was not the case with Loab.

A digital manifestation of the collective imagination

Some observers have interpreted this recurring presence as a kind of bad omen. Kind of like Loab’s presence contaminates AI productions with some sort of sinister aura. In an interview at ABCSupercomposite also explains that it was warned by groups of religious fundamentalists.

Of course, this “AI ghost” is not a supernatural manifestation; it is simply the product of an algorithm that has processed data, no more and no less. The artist also speaks of a ” emergent statistical accident “.

What is interesting is to try to understand how an accident of this kind could give birth to such a coherent and tenacious concept, likely to re-emerge in a very identifiable way in yet distinct works.

The answer is partly obvious. All such generative systems learn to compose images through a training process. Very briefly, this consists in providing him with heaps of images by explicitly indicating their content. The artificial neural network can thus learn to make the link between the images and their textual descriptions. Most often, these images come directly from the Internet.

© Andrea De Santis – Unsplash

However, today, we can also consider the Web as a kind of concrete manifestation of the collective imagination. By publishing their texts, their images and their moods on the web, earthlings also feed a kind of large database of the human psyche.

And since the algorithm has necessarily drawn from this pile of pieces of life to give substance to this entity, we can also consider it as the symbolic reflection of a dark corner of our collective imagination. ” In a way, it’s almost scarier than dealing with a supernatural being. “Judge Supercomposite.

A new window on the human psyche?

But from there, we find ourselves in a dead end. That’s the whole point of machine learning; we know what we have at the input, we get a result at the output, but everything that happens in between is often far too abstract for the human brain. In one way, AI has its reasons that human reason ignores ; it is therefore impossible to objectively determine what Loab really represents, why she tends to reappear like this, or why is consistently associated with macabre imagery.

Of course, you have to be quite careful and be wary of hasty conclusions. But it is also a fascinating thought exercise. Because in the end, Loab is above all the product of a true innovative artistic approach at the frontiers of art, crowd psychology, digital folklore and artificial intelligence. And it is quite possible that one day, this kind of work will allow humanity to learn a little more about artificial intelligence in general… and also about itself.

What is Loab, the “ghost that haunts the AI”?