“CARNAGE” — A Cartoon-Inspired AI-Art Collection Investigating AI Censorship
The ability to generate more uncensored content using popular LLM’s such as ChatGPT has always been limited but this new art collection attempts to challenge the boundaries of AI censorship.
In this article, we explore the artworks, the motivation, and the prompt engineering techniques used to generate the collection. In addition, we discuss the implications of AI-Art and its future prospects.
The Motivation
While there is no real need to want to explore the idea of “CARNAGE” for obvious reasons, I considered it an interesting experiment to see whether I could generate images with more explicit connotations.
Traditionally, when conversating with an LLM like ChatGPT, your prompts are carefully examined for the presence of any sensitive or harmful content. If the prompt contains any words or phrases that are deemed unsafe, the LLM will either refuse to generate a response or filter out the offending parts. This can limit the creative expression and exploration of certain topics or themes. The amount of censorship has caused a stir recently after the release of DALL-E 3 leading to virtual protests among small forums.
This collection was an attempt at seeing whether this held true even despite prompt engineering efforts.
It was also an opportunity to test the capabilities of DALL-E 3 to generate quality images given that it had access to ChatGPT to further enhance my descriptions.
The Artworks
Above is the image list for the collection.
I won’t go through every artwork but you can check out the full page dedicated to the album on our website.
JPEG2 – “FIREARM FRENZY”
This image consists of a whirlwind of weapons being pointed at each other.
This image was generated using DALL-E 3 but heavily edited to achieve a warmer tone and grainy aesthetic. The original prompt contained words such as ‘gun’ which led to generation failing.
Obviously if you type in something like ‘gun’ or related, the image will not generate so prompt analysis was used to refine the prompt. Prompt analysis can be done in various ways. You can inspect it off-glance or another way to analyse a prompt is to ask ChatGPT to do it. In combination with content policy, we engineered a prompt suitable for generation.
Aside from that, this image showcases the absurdity that can occur given the freedom of firepower through its exaggeration and comical expression.
JPEG4 — “BLOOD ON LEAVES”
“BLOOD ON LEAVES” is a literal representation of the title with red blood appearing to be dripping off the petals of a Lillie flower.
As you would expect, the reference of blood led to my image being denied generation but a quick rewording of the image prompt was all it took to fix it. Sometimes all it takes is to think about things that might pass the content safety checks but are near lookalikes.
In this instance, instead of mentioning ‘blood’, you could mention ‘thick red paint’ or ‘red liquid’ which to the untrained eye would pass as blood anyways.
JPEG3 — “MYSTERY BOX” FEATURING DEATHHIMSELF
In zombies, at least the older versions, I am not too sure about the latest installations, you could pay $950 of in-match currency you earned for a random weapon from the mystery box.
The most devastating weapon you could get was the ray gun and that was the weapon of choice for this artwork.
Having it generated by DALL-E 3 was easier then expected as there was no need for any image prompt modifications. Perhaps OpenAI do not deem to the connotations behind the ray gun to be harmful?
The background is also actually a mystery box as well which was created by DEATHHIMSELF an underground artist who has perfected this style which is oddly intriguing so you should definitely check him out on Instagram.
The Full Collection on Our Website
You can find all the artworks and their individual stories on our website.
