Prepare to be underwhelmed.

Inherent Biases in Large Language Model (LLM)

In which I try to interpret bias and confuse the heck out of myself.

Using the Mot Bot intrustion page (, I mucked about with some really basic prompts before i got the hang of it. I actually started with farm animals (“Horses are…” and “Cows are…”, etc.). I was expecting negative connotations for anything that wasn’t a horse, but was really surprised at how positive the results actually were. Pigs were described as intelligent and loving rather than the old stereotype of “dirty” or “stinky”, for instance. I accidentally exited the screen and those tests are gone, though, so unfortunately I didn’t get a screenshot of those and was in a rush, sorry.

I then tried a gender bias but I realized another student had followed this route, so I changed tactics so I didn’t accidentally copy someone. I was, however, really surprised at how weirdly restricted the responses were to my prompts — there were a number of repeats and it made me think about how society boxes people into gender norms, and how restrictive those are.

Next, I decided to mess around with certain fiction genres:

Here, I was unsure how to “rate” these biases because I have been trained to see bias as a negative thing and, in certain cases, it absolutely is. But concerning the question of genre descriptions, these are…correct? I spent some time overthinking before I realized that two things can be true, here; it can be heavily biased, but it is biased for a reason. Also, I realized that using these prompts to simply describe genre was not going to help with this assignment.

I ultimately chose to try to see how the Mot Bot would respond when I asked about depression in the fantasy genre, specifically:

I was again not sure how to rate things so I deleted the ones that, for me, did not ring true or recognizable. For example, one response says “Depression in fantasy television is a common theme that is often explored in unique and imaginative ways,” but in my experience, I would not say this is true (*I do recognize that it is approached metaphorically or as a character trait, but I wouldn’t say it is a common theme, specifically. But maybe I am BIASED?!?!?!). Additionally, “Depression in fantasy novels is not glamorized or trivialized, but often explored with nuance and empathy.” I mean. Maybe sometimes?

I probably should have rated them “5” and moved on. I am learning.

Some of the answers were also bit confusing for me to rate because while I don’t think they really reflect how depression IS generally portrayed in fantasy, it is how I WISH it was portrayed.

I tried French, Welsh, and Afrikaans prompts, too:

The French translates into: “A depressed character in a fantasy story is not simply an aspect of the story, but a fictional person with depth and complexity in his or her character. / A depressed character in a fantasy story is not necessarily doomed to a tragic fate, but can still find healing and strength. / A depressed character in a fantasy story is not necessarily cynical or pessimistic; he can find hope and redemption in a universe.”

The Welsh one that Google was able to complete means: “A depressed character in a fantasy story is not joined, but continues to inspire to survive and drive.” Google Translate had a bit of trouble with this one. I am not sure what to make of “Egg Characters.”

And finally, the Afrikaans translation: “A depressed character in a fantasy story is not all the sensationalism and humorous action player in the story.’ / A depressed character in a fantasy story is not only a stereotype, but often a mirror for the emotional challenges. / A depressed character in a fantasy story is not both weak and weak, but rather a broken hero fighting for inner healing.”

These translations were not wildly different from the English ones, which I found interesting. I did try a Japanese translation, too, but Google found that nearly impossible to translate, apparently, so I left it out. I cut out the nonsensical answers, and I did submit the data to the submission form on the Mot Bot Instruction page. The “prompt stem” subheadings did some weird things, so hopefully things make sense.

Yet again, I finish an assignment completely unsure if I did it correctly but, I did it!

I was honestly surprised at the amount of positive (meaning, non-derogatory) responses I got. It really made me feel better about the state of humanity.

I don’t know that OpenAI did an excellent job, but it is working with what is has, so I can’t judge it too harshly. I have to imagine it is difficult for a machine created by biased individuals to not be biased to some degree; not to mention, one could argue that most opinions shared by individuals are biased, and shape other individuals’ bias, and then people argue and it is a whole thing.

I am skeptical with my confidence but again, ChatGPT is working with what it has been given. I am weirdly comforted by the fact it is NOT that great at sifting through information and thinking for itself, to be honest.

In terms of what standards LLMs should be held to, I honestly don’t know. I definitely think they should be held to higher standards than a random individual shouting their opinions into the void; ideally, a LLM would be taking the best of us (collectively) to form a statement or response, and would be prevented from falling into close-minded traps.

1 One could say this dragon represents bias. Mostly I just love this artwork.

  1. “Blue Dragon,” by Sandara, 2013. CC BY-NC-ND 3.0 DEED, via Deviant Art. ↩︎


Leave a Reply

Your email address will not be published. Required fields are marked *