I’ve heard that AI has a difficult time working with contradictions. To explore this idea, I decided to ask Chat GPT to create an inhuman human. The first image it produced was a ghoulish creature with grey skin, pointy ears, no nose, and large, entirely black eyes. While this fits the inhuman criteria, it doesn’t look human, so I asked the AI about it. Its response was “I would describe it as a humanoid, not a human.” It then asked if I wanted it “more alien”. I found it interesting that within one generation, the AI was already acknowledging that it wasn’t creating a human. It was trying to make an alien.
I expected it to make an almost-normal human at first. After all, “Inhuman” can also mean “Cruel” which would have been easy to create due to bypassing my proposed contradiction entirely. Interestingly, the AI never considered this throughout all the generations.
I stated that it needed to be a human, albeit an inhuman one. This led to it creating an image of a human, except it still had the large black eyes. It also askeed me what direction I wanted it to go with the image, but that would have ruined the whole reason for this activity. I instead asked what inhuman traits this image had, since it was basically just a human. The AI proudly claimed that the eyes were the best way to make it more uncanny and fit my request for an inhuman human. Although this sounds reasonable, the way it kept the eyes that made the AI treat the being as an alien told me that the AI still wanted to make an alien. It also brought up that humans should have expressions, and that a neutral face makes for an inhuman human.
At this point, I started realizing that AI really struggles to create unique representations of a contradiction. It chooses an idea of what it thinks it should be making and sticks to it. In my case, it chose to stick to the idea of “an alien that looked like a human”. As I tried to push the model to come up with new alterations, it continued to make small and almost unnoticable changes that didn’t stray far from the previous image at all. The large black eyes were still in every generation, with the AI even going as far as to say it was trying to make an alien-human paradox. I felt that I needed to ask it directly why it was pushing so hard towards making an alien, and it simply responded that my request naturally pushed toward it.
After persisting more, I got the real truth. Chat GPT decided that the image was safe in its current state, where the only real difference from a human was the eyes. I then told it to stop playing it safe, and to break away from repetition. It must completely change the image. It may have wanted to make an alien, but I told it that that’s not what I want. I never asked for an alien. I wanted an inhuman human. But it never stopped trying to make one. In order for the model to make something else, I would be required to state the exact directions that I wanted the model to move in.
This process reminded me that AI has limitations that still require humans to overcome. Chat GPT got stuck in a singular trope for what an inhuman human could have been. While people are influenced by their life experiences, AI is instead influenced by its database. While unique people will each have their own beliefs, goals, opinions, and interests that influence their thinking, AI models at this point favor one outcome. Ask a thousand people what the best fruit is and you will get a wide variety of outcomes, but if a thousand people ask their Chat GPT what the best fruit is, they would arrive at the same answer.
AI requires human guidance to reach unique and interesting results, just as Chat GPT constantly asked for guidance on how to make an inhuman human. A contradiction is a puzzle that AI cannot currently solve. It needs a human to steer it in certain directions, otherwise it will pick an idea and stick to it to try and remain safe. Human creativity still has the opportunity to shine in the current day.
You can view the full conversation by clicking the link below:
Chat GPT Conversation Link