Shrimp Jesus plunged AI-generated viral images into the big mainstream sea two years ago with a full net of images of the Christian son of God mashed up with the crustacean that drove Red Lobster into bankruptcy.
The images of a deep-sea prophet were nearly impossible to ignore when they washed up on our virtual shores. This made creating those fake AI images highly profitable for social accounts willing to generate this special brand of AI slop, which is low-quality content produced by AI.
Shrimp Jesus is part of a larger ecosystem of fake AI images. Cuihua (Cindy) Shen, a UC Davis professor of communication who studies online misinformation, said this ecosystem has pushed us into uncharted territory. She describes AI slop and misinformation as a systemic problem a lot like environmental pollution. Dealing with it individually shifts the cost of misinformation from online platforms to each of us with a need for constant skepticism.
“In the past, ‘seeing is believing’ is a good thing because it allowed us to conserve our cognitive resources rather than constantly questioning every visual we encountered,” said Shen. “Instead, we have to make skepticism our default. This is so inefficient because it puts the burden on individuals to constantly question what they see, when the problem is really systemic”
How to spot AI-generated images
Not all AI-generated images are as bad as Shrimp Jesus. Some fake AI images are so realistic it’s hard to spot them if you don’t know what to look for.
In a new study, Shen and her co-authors describe the features that give away most AI-generated images as “algorithmic surrealism.” They analyzed more than 28,000 AI-generated images on Instagram and found that even the most photo-realistic ones differ from reality in three specific ways:
- Physical surrealism: unnatural or implausible physical characteristics
- Behavioral surrealism: impossible or supernatural behaviors that violent physical laws, biological limitations or sociocultural expectations
- Contextual surrealism: implausible or fantastical settings that violate normal expectations of time, space, narrative coherence, or plausibility
The volume of fake AI images and AI-generated content in general has steadily grown since ChatGPT, the first publicly available AI chatbot, came online in 2022. A 2025 report estimated that 52% of all articles posted online were generated by AI. Another report that year estimated that 74% of the 900,000 new web pages analyzed contained AI-generated content.
An objective scale for media literacy
The ability to spot fake images is part of media literacy, which is one way to maintain a sense of what is true and false online. However, Shen pointed out that even media literacy has been taught and measured in many different ways that don’t all translate to real-world skills.
In another new study, Shen and her co-authors evaluated several approaches to media literacy that come from across research disciplines, including education, communication, media studies, information science and others. The paper’s first author is Sijia Qian, an assistant professor at the North Carolina at Charlotte who received her Ph.D. in communication from UC Davis.
“Media literacy is clearly very important, but there are so many different dimensions and people keep inventing more,” said Shen. “They’re all trying to capture important competencies, but we don’t have a unified way to measure them. That makes it difficult to compare findings across studies or even know which dimensions really matter.”
Shen and her co-authors analyzed the most prominent models for teaching media literacy to find out if they teach the same skills or different ones. One finding is that some measurements of media literacy are based on self-reports, meaning that learners say whether or not they feel they have media literacy skills.
But self-reports of skills are notoriously inaccurate. A number of studies in psychology have documented what is known as the Dunning-Kruger Effect, in which people consistently overestimate their knowledge or ability in a specific area.
Shen and her colleagues developed the Digital Media and Information Literacy Scale, a new measure of media literacy that separates digital media from news media and knowledge from skill. It also measures media literacy through subjective measures as well as the much less-common objective tests. They developed and validated this scale with participants from both Western and Asian countries.
Ways to deal with AI slop and misinformation collectively
Shen teaches media literacy but sees it as an insufficient solution to a systemic problem within the information ecosystem — one that places the burden on individuals rather than addressing the underlying structural issues.
Online media are a resource we all share just like a water reservoir, said Shen. With a reservoir, the resource is clean water. For social media, the resource is quality information. For both a reservoir and social media, it only takes a handful of bad actors to ruin the resource for everyone. When online media are polluted with bad information, people have to respond by becoming skeptics of everything they see.
Media platforms hosting AI slop and misinformation can take action to reduce the spread and visibility of low-quality or inaccurate content, said Shen. However, platforms may profit from the engagement that misinformation generates. A 2024 University of Washington Study documented how influencers made millions from COVID misinformation.
Governments can also develop policies that encourage greater accountability within the information ecosystem. Singapore, where Shen’s Fulbright U.S. Scholar award enabled her to collaborate with colleagues to develop the Digital Media and Information Literacy Scale, has stricter laws about online content that hold people responsible for posting misinformation — a stark contrast to the more limited regulatory approach in the U.S.
Singapore’s 2019 Protection from Online Falsehoods and Manipulation Act imposes penalties for people and platforms that share content deemed “false” or “misleading.” However, the law has also been criticized as a means for government ministers to silence or punish critics.
Shen said that any lasting solution will look beyond individual users and consider the broader systems that shape what people encounter online.
“Individual-level solutions can only go so far in addressing a systemic problem,” said Shen. “When someone falls for a scam, for example, we often focus on whether they had enough knowledge or training. But emphasizing training alone can obscure the broader systemic factors that create these problems.”
This research was partially supported by the National Science Foundation and a Fulbright Scholarship.
YOU MAY ALSO LIKE THESE STORIES
What Makes a Hit? On TikTok and Spotify, Listeners Only Partly Decide
A new study in communication examined how the data-driven business models of TikTok and Spotify shape both the music artists make, and the songs people choose.
What Explains the Increase in Online Hate Speech?
Over the past few years, there have been extensive changes in content moderation policy on social media platforms. As companies like X and Meta loosen these policies, users have reported an increase in hate speech. So what caused this shift?