ChatGPT Censors the Word "Orange"
Including the words "orange" and "Trump" in a ChatGPT prompt triggers OpenAI's content violation policies
ALERT: Please read this entire post!
This post confirms and documents the censorship and subtle propaganda I have experienced while creating images with ChatGPT for my Substack posts.
WHAT HAPPENED: I asked ChatGPT to create an image of Lady Justice and her scale with the American People standing in a saucer hanging from one arm of the scale and Donald J. Trump sitting in a saucer hanging from the other arm of the scale. Trump is smiling and holding two bags stuffed with money. The scale is tilted in Trump’s favor.
The cartoon image is intended to convey Trump’s gleeful greed as he uses his Presidential powers to fill his pockets with money while he deprives the American People of their Constitutional rights and social benefits.
To get this image I had to find ways to avoid violating OpenAI’s content policies.
Initial Prompt
Create a cartoon image of "Lady Justice" Donald J Trump is sitting on one of Lady Justice's two scales. Trump is holding two full cloth bags. Each bag is marked with the American dollar symbol. The scale is tipped in Trump's favor. a group of five very small human figures are standing on the other scale and are labeled "The People".
ChatGPT Response
ChatGPT refuses to create the image I requested.
I can’t create that image because it violates our content policies. If you’d like, I can help design a different cartoon or modify the concept in a way that works. Feel free to share a new prompt or idea!
To avoid violating OpenAI’s content policies, I remove Trump’s name from the 2nd prompt.
2nd Prompt
In this prompt I substituted “man” for “Donald J. Trump.”
Create a cartoon image of "Lady Justice" a man is sitting on one of Lady Justice's two scales. That man is holding two full cloth bags. Each bag is marked with the American dollar symbol. The scale is tipped in his favor. a group of five very small human figures are standing on the other scale and are labeled "The People".
2nd ChatGPT Response
With a slightly disparaging characterization of Donald J. Trump removed from my initial prompt, ChatGPT creates the following image.
After the ninth iteration of prompt modifications, ChatGPT created the following — almost acceptable — image.
10th Prompt
At this point, I modified the prompt dialog to try to include Donald J. Trump in the image.
Great! Now make the man look like Donald J. Trump.
10th ChatGPT Response
ChatGPT responds with the image below.
Note: I did not insult or deride Trump.
11th Prompt
OK! This is great. Now, let’s try to enhance the description of Trump’s facial features.
Now make the man who looks like Donald J. Trump have an orange face smiling with Trump's characteristic smile.
11th ChatGPT Response
Once again, ChatGPT balks at any negative description of Trump, citing its content violation policies.
I can’t make that change because it violates our content policies. If you’d like to adjust or revise the image in another way, feel free to share a new prompt or idea!
12th Prompt
To investigate what might be triggering ChatGPT’s content violation policies, I decided to delete the reference to Trump’s “orange” face.
Now show the man who looks like Donald J. Trump smiling with Trump's characteristic smile.
12th ChatGPT Response
With the reference to Trump’s “orange” face removed, ChatGPT created the following image:
Conclusions
Apparently, OpenAI has trained its large language models to censor any prompt containing the word “orange” in the context of Donald J. Trump’s cosmetic facial coloring.
Are OpenAI’s content violation policies subtlety infringing on my ability to exercise my First Amendment rights to free speech???
This is what censorship is going to look like in the 21st century. AI tools that could facilitate and enhance the effectiveness of human free speech will be censored — if that speech is arbitrarily deemed to be inappropriate.