An artificial intelligence-generated influencer named Janie B accumulated more than 120,000 views on TikTok during the University of Alabama's annual sorority recruitment season, in an experiment designed to measure how audiences respond to synthetic content that mimics real people.
The account, @jani.b086, posted its first video on Aug. 8, showing a curly-haired, bespectacled redhead introducing herself from a dorm room and listing her outfit of the day — a pleated skirt from Abercrombie, a knit top from J. Crew, socks from Target and sneakers from Nike. The comment section filled with users calling her «adorable.»
But Janie does not attend the University of Alabama. She did not rush a sorority. She did not buy any of the clothes she described. Every element of every video — the dorm room, the outfits, the person herself — was generated by artificial intelligence.
The experiment was created by Olivia Moore, an investment partner at the venture capital firm Andreessen Horowitz, who spent about $100 in credits to produce a week's worth of videos of a fake influencer participating in «RushTok,» the annual TikTok phenomenon in which thousands of potential new members document their sorority recruitment experience.
Moore, who works on the consumer investing team and focuses on companies building consumer AI products, said she has long been fascinated by how people create, interpret and understand AI-generated content.
«The image, video, audio models are getting very human-like, very fast. I think we probably cross the uncanny valley in the next six months,» Moore said. «And so it means that how we as humans are able to detect AI-generated or AI-assisted content and how we interact with it is just going to become even more important because we're going to have a flood of this content, and it's going to be, in many cases, indistinguishable across all of our channels.»
To build Janie, Moore primarily used ChatGPT and the AI video model MiniMax. She pulled a real video of a University of Alabama dorm room from the internet and uploaded it to ChatGPT, then prompted the system to create a character with frizzy red hair, glasses and a style inspired by the 1960s. ChatGPT produced a still image of Janie placed inside the real dorm room photo.
For each TikTok video, Moore animated Janie by feeding the still image into MiniMax H3. She used ChatGPT to generate prompts that shaped how Janie would move, what she would sound like and what scripts she would follow.
Moore chose RushTok as the testing ground after watching the phenomenon for five years. She said it was «the perfect testing ground» because Janie would not be selling anything or hurting anyone, and because RushTok naturally draws high traffic and has a low barrier to entry for new creators. With other potential new members posting similar content, audiences would encounter Janie alongside real humans.
Reactions to Janie were split. Moore said viewers under roughly 25 immediately identified her as AI, while many in the older demographic — a significant portion of the Bama Rush audience — were fooled and wished her luck.
By the second day, much of the audience had caught on. People discussed Janie on Reddit, posted commentary videos on TikTok and corrected anyone in her comment section who believed she was real. Yet some viewers who were told Janie was AI responded that they did not care.
«She seems so sweet. Like I want her to succeed,» Moore recalled them saying.
As the audience caught on, Moore changed her strategy. She began dropping hints in the videos and watched as digital sleuths proudly pointed out clues in the comments. The experiment demonstrated that people were still willing to watch AI-generated content as long as it was entertaining enough.
Moore said humans are getting better at detecting AI faster than AI generation is improving, describing the situation as two competing lines on a graph with different slopes. She said the rapid improvement of image, video and audio models means the ability to detect and interact with AI-generated content will become increasingly important as such material spreads across all channels.
The experiment also raised questions about what people feel when they do not know content is AI, and what they feel when something they liked or hated turns out to be generated. Moore said those philosophical questions were part of what motivated the project.