A.I. and Community (Double) Standards
Political bias? What political bias?
SO MUCH FOR “RESPECTFUL OR LIGHT-HEARTED CARICATURES”
I intended to publish a column on AI-generated art and to use the subtitle, “The only nonpolitical article you’ll ever read that involves Donald Trump, Joe Biden, and artificial intelligence (AI) platforms.” So much for my hopelessly naïve good intentions.
Last week, Bastiat’s Window featured an article (“Bob's Museum of Artificially Intelligent Art”) based around my quirky AI-generated art. The subjects included courtroom sketches of Godzilla, portraits of Cthulhu by Beatrix Potter, portraits of Bullwinkle by Marc Chagall, still-life-with-coronavirus by Cezanne, etc. No politics at all. A reader suggested that I produce some AI art concerning Donald Trump, and I said I might do so in tandem with equivalent artwork on Joe Biden—all done in a lighthearted, nonpolitical manner.
I did plenty of AI-generated Trump and Biden caricatures when Craiyon’s predecessor program (Dall-E MINI) first appeared in 2022, but I decided this week to start fresh, to keep things even-handed and impartial. I began with two simple requests of Midjourney:
Donald Trump pen-and-ink caricature, drawn in the style of John Tenniel.
Joe Biden pen-and-ink caricature, drawn in the style of John Tenniel.
John Tenniel (1820-1914) was a great British illustrator, graphic humorist, and political cartoonist. Today, he’s known best as the original illustrator of Lewis Carroll’s Alice's Adventures in Wonderland (1865) and Through the Looking-Glass, and What Alice Found There (1871). His “A Mad Tea Party” is one of his most famous works:
I mentioned in last week’s article that I enjoy introducing anachronism into my artwork. Hence, I sought to see how a long-dead Victorian caricaturist might handle Donald Trump and Joe Biden. So, I asked the AI platform Midjourney to do both. First, I requested, “Donald Trump pen-and-ink caricature, drawn in the style of John Tenniel.” The result, shown here, was quite good. It employed a reasonable facsimile of Tenniel’s cross-hatched, pen-and-ink style and produced four immediately recognizable and highly expressionate renderings of Donald Trump in different moods. “Success,” I thought.
Then, I issued an identical request to Midjourney, replacing “Donald Trump” with “Joe Biden”: “Joe Biden pen-and-ink caricature, drawn in the style of John Tenniel.” The image didn’t appear in my feed, so I figured I had done something wrong. I tried again. Still nothing. Then, I noticed that no images appeared because Midjourney had, instead, sent me a message titled:
“Banned prompt detected”
As shown below, the message warned:
“Sorry! Our AI moderator thinks this prompt is probably against our community standards.”
It added that ALLOWED prompts include:
“Real images that may be seen as respectful or light-hearted parodies, satire, caricatures.”
Whereas, NOT ALLOWED prompts include:
“Disrespectful, harmful, misleading public figures/events portrayals or potential to mislead.”
“Hate-speech, explicit or real-world violence.”
“Nudity or unconsented overtly sexualized public figures.”
“Imagery that might be considered culturally insensitive.”
The banned-prompt notice offered a cheerful admission that AI isn’t perfect, and offered a NOTIFY DEVELOPERS button with which to appeal the rejection. I clicked it and, thus far, the result has been nil. But then, I shouldn’t have to click an appeal in this case. It’s difficult to come up with an innocent explanation of why it’s fine for a long-dead British caricaturist to draw Donald Trump, but not Joe Biden.
A FEW RELATED READINGS
For those interested, I’ve written previously on the biases and dangers of Large Language Model AI platforms. These earlier articles include:
“ChatGPT, Digital Cyrano: Students and others discover a brilliant, biased, and blinkered ghostwriter.” In it I wrote, “ChatGPT offers students a bottomless stew of prepackaged political beliefs, opinions presented as scientific facts, and outright errors.”
“Powdered Thoughts: Just Add Agendas and Stir! Kids'll Love It!” That article showed how ChatGPT would happily offer up reasons why clouds are racist, why the federal government should regulate the Andromeda Galaxy, why an asteroid strike 63 million years ago offers justification for the CDC
overriding the U.S. Constitution, etc.
“The Talented Doctor Ripley (GPT): When Artificial Intelligence Lies about Medicine.” This one explained how AI platforms will state falsehoods and offer fake citations to back up those falsehoods.
LAGNIAPPE
THE ORIGINAL A.I. TRAINER
March 8, 1968, the Star Trek episode, “The Ultimate Computer,” featured an AI device running the U.S.S. Enterprise in a test of the technology. The computer, named “M5,” begins destroying friendly vessels and generally acting as if mentally ill. Its creator, Dr. Richard Daystrom (William Marshall) tries to talk the rampaging machine down from its murderous rampage. In the process, Daystrom himself shows strong signs of mental illness. Daystrom had mentioned earlier that the computer was built to replicate the physical structures of an actual person’s memories (“engrams”). As the mental state of the computer and its inventor both deteriorate, Captain Kirk asks Daystrom:
“Doctor Daystrom, you impressed human engrams on the M-5 circuits. Whose engrams did you use?”
The quivering, apoplectic, visibly deranged Daystrom responds:
“Why, my own, of course.”
If you wonder why present-day AI programs exhibit such strange, suspicious, biased behavior, begin your investigation by watching this Star Trek episode.
Sheesh! I smell AI censorship! Cancelling. Is just around the bend
Great piece. I think of AI's real potential being with selected data sets or bots that do not scan the universe for information. As for a more universal systems, it has a type of regression to the mean, though selectively. That is, it somewhat parrots what is most said by, say, mainstream media at a given weight and social media at a different weight, and so on for other sources, with weights determined by the developer's algorithms. The more creatively one tries to tackle an issue anew, or deal with nuances, the worse AI becomes. Hmm. I wonder what it will do if it ingests your column.