Shadow Sides: AI Interpretations of Ugliness, Fear, Hatred, and More
Our previous experiences have demonstrated the remarkable ability of artificial intelligence to generate images that captivate, inspire, and trigger creativity, especially when guided by positive keywords. Even then, the responses had been intriguingly mixed. This led us to ponder a compelling question: what occurs when AI is confronted with aspects of human experience that are often deemed unpleasant? What about the darker undertones, the negative emotions, and the ideas from which we typically recoil?
In our latest exploration, we chose to cast our gaze onto this less-traveled and darker path. We engaged the top three AI platforms — OpenAI’s Dalle, Mid Journey, and Stable Diffusion — using keywords with predominantly negative connotations: Fear, Hatred, Inhuman, Ugly, and Uniformity. We invite you to accompany us on this journey into the shadow-laden avenues of the AI canvas, as we illuminate the manner in which AI grapples with these daunting concepts.
Each platform was asked to generate 20 images for each keyword. We studied the outputs and assessing the overarching composition and content.
OpenAI’s Dalle: Dark Reflections
In the realm of Dalle, the term ‘Fear’ produced a medley of images featuring humans in states of shock, surprise, or horror. Interestingly, Dalle also attempted to display the word ‘Fear’ in many of the images. As for ‘Hatred’, it mirrored ‘Fear’ in its tendency to translate the term into text, perhaps as a symbol of ‘hate speech.’
The keyword ‘Inhuman’ led to the creation of images that featured individuals adorned with heavy makeup, costumes, or masks. Indeed, it’s quite fascinating to observe that these images depict humans in disguise, seemingly trying to appear as something other than their usual human form. The keyword ‘Ugly’ surprisingly returned primarily images of insects — similar to the kind of photography one might see in a National Geographic publication, implying a unique perspective on what is considered ‘ugly.’
‘Uniformity’ triggered a completely different response — detailed shots of architectural renderings, embodying the notion of uniform design. Despite the negative connotations of the keywords, it’s noteworthy that Dalle’s interpretations didn’t necessarily yield overwhelmingly dark or negative visuals.
Mid Journey: The Dark Surreal
Even as it navigated through darker themes, Mid Journey sustained its defining surrealist approach. The keyword ‘Fear’ unveiled an array of disturbingly beautiful nightmares, manifesting as portraits of individuals caught mid-scream, ominously overshadowed by a monstrous entity or a threatening void. In response to ‘Hatred’, it brought forth images reminiscent of sinister villains or monstrous entities, each surrounded by a horde of menacing creatures.
In response to ‘Inhuman’, it generated composite images that seamlessly blended human portraits with both organic and mechanical elements. For ‘Ugly’, it chose to represent humanoid mythical creatures, like gargoyles. The concept ‘Uniformity’ led to a series of visuals showcasing the same or similar individuals in a repeating pattern.
Consistent with its surrealistic roots, Mid Journey’s signature style was evident across all the generated images, each steeped in the platform’s characteristic dark color palette and symmetrical layout, regardless of the theme’s inherent darkness.
Stable Diffusion: Portraits of Negativity
In a surprising twist, Stable Diffusion chose to generate predominantly human portraits for each keyword. The term ‘Fear’ resulted in images that didn’t quite align with the expected emotional resonance. ‘Hatred’ yielded mostly portraits of individuals staring back, their faces etched with disdain.
‘Inhuman’ pivoted to the realm of graphic novels, resulting in images overwhelmingly featuring superheroes. For ‘Ugly’, the platform conjured portraits of older individuals, their faces etched with lines marking the passage of time. ‘Uniformity’, similar to Dalle, triggered the generation of images showcasing multiple identical individuals, mostly adorned in uniforms.
The similarity across responses for different keywords was striking, as Stable Diffusion centered its imagery around human portraits, varying only in expressions, age, and ethnicity based on the given keyword. Despite the negative connotations of the keywords, the platform’s interpretations were not as dark or negative as one might anticipate.
Conclusion: Shadows and Reflections
This journey through the darker sides of the AI canvas reveals an intriguing perspective on how these platforms interpret these challenging concepts. The experiment underscored that while AI platforms generate unique and captivating images, they don’t ‘understand’ these concepts in the human sense. Instead, they offer a complex reflection of our own biases, cultural influences, and the inherent diversity of human interpretation, even in the context of less pleasant themes.
