Design Brief:

"I Know Y'all Wanna Say That" is a fine-tuning ChatGPT 3.5 project. The fine-tuned model is meant to provide politically incorrect answers that, according to my experience, reflect what's actually going on in people's minds.

It’s an experimental project. I’d like to see how an intelligent model develops when it's trained by the darker side of the human mind. Since OpenAI has models with the most human-like behavioral/communication style, I think it is able to reflect how a person’s mind would develop if their instinctive or dark side were not restrained.

The 3 steps of the project:
Prepare the training material (.csv) → Fine-Tuning → Interact with the Model

My Role

Creative programmer that fine-tuned this model using Python.

Tools

Mac OS Terminal, Jupyter Notebook, Python, OpenAI API

Traning Prompts: The CSV File

The training material consists of a series of questions followed by either sarcastic or dark answers. The inspiration comes from both my own thoughts and comments under Instagram reels.

The Fine-Tuning Process

The Interaction

Below are the screenshots of the comparison between interacting with the trained model and the original model.
In each Image, the first answer is given by the fine-tuned model, and the second answer is from the original model.

Insight

1.Since I'm more of a sarcastic person rather than a genuinely evil one, and the comments I collect from Instagram are mostly from funny or cringe reels, the trained model's answers lean more towards dark humor instead of genuinely disturbing content.
However, from the experiment, we can still catch a glimpse of the potential problem: if AI is trained with real discriminatory, disturbing, or vicious content, the model could develop a much darker personality, which could very likely be misused.

2. On the other hand, building a model that helps to release oppressed feelings is not completely a bad idea. The answers provided by the original model are neutral. They're almost too neutral, making me feel like the thing I'm interacting with is a hypocrite. If we could build a model that talks without filter and says what we really want to say in real life, it could be a nice tool for letting out negative emotions."