Artificial Insights? Understanding Political Narratives through AI Imagery
And the Consequences of Midjourney’s Policy Shift on Political Imagery
Starting on June 15, 2023, as a part of my radical sabbatical, I began a project that combines two of the things I tend to focus on when I’m teaching at Stevens Institute of Technology, political communications and artificial intelligence. AI products burst onto the scene right as I was leaving the classroom, so during my time away I wanted to see how these new tools could augment the way we understand and visualize political data. I gave a talk in Kansas City around that time about AI generate images and started with one of my first, an awkward presentation of Joe Biden and Donald Trump holding hands across a table.
For this project I took the most used words from the official e-newsletters of Congressional Democrats and Republicans hosted at DCinbox, and used them as the prompt for image generation using Midjourney once a day. This project has revealed fascinating (to me at least) insights into the thematic differences in communications between the two major political parties in the United States. Midjourney generated images are a visual representation of political discourse, and by having the process update each day, I was able to watch our politics change over time. These pictures show things about our politics in a way that is far more interesting (and somewhat easier to understand) than traditional word lists or wordclouds.
Early on I tinkered with signaling the partisanship of the e-newsletters by asking Midjourney to render things with either red or blue hues, but I found that that often meant that the Republican images seemed more menacing, while images rendered from Democratic words had a calmer look — perhaps because these two color cues are important when constructing a perception of an image — or perhaps because the words used by the GOP were more menacing and those from the Democrats were more dull. One goal of the project was to continue doing this each day until the 2024 election and see what I could learn both about communications and AI tools.
It’s been an interesting ride, and I have learned neat things, but recently the project took a big change when Midjourney decided to “hammer” or ban images using the prompts “Biden” or Trump”. Both the project and the change teach us things about political communications and how AI tools intersect with this arena.
The most obvious difference between images rendered from Republican and Democratic daily word list is that Republicans in Congress talk about Biden. A lot. In nearly every day his name is a top used word and his image shows up in the pictures, oftentimes in amusing ways. On July 23, he was pictured praying in a way that is reminiscent of Christian imagery, in August as Republicans wrote about fentanyl and president he becomes a mad scientist, in October as things turned to spending bills and funding the government he’s dancing with cash flying all around him, as Israel became a topic his likeness (and sometimes that of former president Trump) were rendered in conflict settings, in December Biden became a sort of Santa.
Democrats don’t often write about Biden in these e-newsletters, they didn’t really do that for Obama either when he was in office. Instead they tend to use more general words like “community”, “health care”, or they call out other political actors like Putin. In my opinion — and others’ — the images from Democrats are a little boring, and much more scattered in attention. They do tend to have more people of color than images generated from Republican e-newsletters. There are clear markers celebrating Juneteenth and Dr. Martin Luther King Jr. from the images rendered from Democratic e-newsletters.
The Biden and Trump “Hammer”
In late February or early March, Midjourney introduced a policy change that significantly impacted this project and changes the broader conversation around AI and politics. The service decided to prohibit the generation of images using the names “Biden” or “Trump,” with company leaders indicating a move towards banning political images altogether for the upcoming year. This decision changed my project — which is fine — but also raises new questions about the role of AI tools in political discourse.
Now when prompted with more generic terms like “president,” — which is also often a top word in official Congressional e-newsletters from Republicans, Midjourney’s AI defaulted to images of Barack Obama or Donald Trump, reflecting a peculiar sort of interpretation or bias in its algorithm. This shift and the resulting renderings reveal much about the challenges and sensitivities inherent in the use of AI within political contexts. When prompted to imagine “our most recent president”, Midjourney renders this:
And when asked to imagine “the current US president”, it returns:
And as for “the upcoming US president”:
Everyone is trying to get this as right as possible, so I don’t fault Midjourney for trying to implement some guardrails against a slew of “deepfakes” heading into the election — but the work arounds also have things to consider. Midjourney’s policy change highlights a n ongoing debate about the realms of AI and creativity: where do we draw the line between censorship and responsibility? While the intention might be to prevent misinformation or reduce political polarization, such actions raise concerns about who controls the narrative in an age dominated by digital platforms and algorithms.
The peculiar behavior of Midjourney, favoring certain former presidents over others when given ambiguous prompts, is an example of the issue of algorithmic bias. This phenomenon not only affects how political figures are represented but also influences public perception, demonstrating the need for transparency and accountability in AI development.
People will want to play with AI to make pictures of politicians. Midjourney CEO David Holz told the Guardian about his plans to ban Biden and Trump prompts and the paper wrote,
I noticed. And I don’t think I’ll be alone. Researchers like myself will want to understand how AI influences politics and how policies within AI companies reverberate into political discussions. As my sabbatical heads towards a close, I’ve been grateful for the time to think about these issues. This project, the internal Midjourney policy change and the resulting limited, rendering universe should give interested parties lots to work through over the coming months. I’ll continue to be in awe of the time we are in, and happy to mix the paths of politics and AI generated pictures.
