The Digital Toybox: Exploring AI-Generated Nostalgia and the Frontier of Generative Hallucinations
Last Revised: June 22nd, 2025
In the rapidly evolving landscape of generative artificial intelligence, the line between creative tool and unpredictable collaborator is becoming increasingly blurred. For Mike, a dedicated blogger and tech enthusiast, what began as a whimsical project to memorialize the memories of the 2025 "Flower Power Cruise" has transformed into a complex case study on the capabilities and persistent "hallucinations" of OpenAI’s latest models.
By leveraging ChatGPT to create custom "Action Figures" of friends, family, and iconic musical figures from the 1960s and 70s, Mike has documented the evolution of AI prompting. His project serves as both a nostalgic tribute to the golden era of rock and roll and a technical exploration of how AI interprets, misinterprets, and occasionally reconstructs reality.
The Genesis: A Cruise Through Time and Tech
The project was born out of the 2025 Flower Power Cruise, a seven-day voyage departing from Fort Lauderdale, Florida. The itinerary, which included ports of call in Falmouth, Jamaica, and the Cayman Islands, was defined by a soundtrack of legendary acts. Passengers were treated to performances by The Cowsills, Jay and the Americans, Ron Dante, and Jefferson Starship, among others.
To immortalize these experiences, Mike turned to ChatGPT to generate custom, retro-style action figures for his fellow travelers. However, he soon discovered that while the AI could capture the spirit of the request, it struggled with the constraints of physical reality—a phenomenon known in the AI industry as "hallucination."
Understanding the AI "Hallucination" Factor
As Mike’s project grew, so did the irregularities in the outputs. Despite advancements in OpenAI’s latest o3 and o4-mini models, internal tests suggest these models still exhibit hallucination rates between 30% and 50%. These errors occur when an AI generates information or visual elements that are logically inconsistent or factually incorrect, often based on a misinterpretation of the underlying data.
"If you see a word misspelled, or multiple ‘Groovy Bears’ or ‘Milk Shakes’ where there should only be one, that is the AI hallucinating, not me," Mike explains. The unpredictability of these models turns every prompt into a gamble, requiring the user to act as both a director and a debugger.
Chronology of an AI Artist: The "Martha" Experiments
The learning curve of AI prompting is best illustrated by the creation of the "Martha" action figure. Initially, the goal was simple: a figure of Martha T. Briseno, dressed in a white shirt and a pink poodle skirt.
- Version 1 (The Glitchy Start): The AI struggled with the prompt, repeatedly inserting unintended elements such as multiple "Groovy Bears" and chocolate milkshakes.
- Version 2 (The Band Invasion): Upon tweaking the prompt to remove the extra items, the AI hallucinated the band "The Archies" into the background. While the inclusion of The Archies—whose lead singer, Ron Dante, was a featured performer on the cruise—was contextually relevant, it was not the intended output.
- Version 3 (The Cinematic Drift): At one point, the AI abandoned the "Soda Fountain" aesthetic entirely, pivoting to a visual style reminiscent of the movie Grease.
- Refinement: Through iterative prompting, Mike eventually achieved a proportional, realistic-looking action figure. The process highlighted a key trend: the more an AI is asked to regenerate an image, the more it tends toward a "cartoonish" aesthetic, often mimicking the exaggerated style of Who Framed Roger Rabbit.
Supporting Data: From Concert Artifacts to Digital Heirlooms
The most successful iterations of this project occurred when the AI was provided with "ground truth" data. For his Peter Noone (Herman’s Hermits) action figure, Mike utilized a unique set of real-world artifacts.
During a concert, Noone had interacted with the crowd, tossing a fan’s purse and a badge onto the stage, and teasing a gentleman wearing gold sparkle shoes. By providing the AI with photographs of these specific items—the purse, the badge, and the shoes—alongside a headshot of Noone, Mike was able to generate a highly personalized, accurate action figure that incorporated "true-life artifacts" as digital accessories.
This method proved effective for other figures as well:
- Dr. Neil V. Hitch: To memorialize Dr. Hitch’s presentation on the archeological excavation of a fire pit at the site of the 1969 Woodstock festival, Mike combined a portrait of the doctor with images of the recovered artifacts (steel soda cans, sandwich packaging) and iconic photos of the original festival. The result was a nuanced, historical action figure.
- The Three-Pack: Using a video frame from a concert, Mike successfully created a three-pack figure featuring himself, Denise, and Keith, demonstrating the AI’s growing ability to process group compositions.
Official Responses and Industry Implications
The project also served as a training module. Mike taught a class at his workplace, USAA, on the practical applications of AI prompting. During these sessions, he and his colleague, Eric Carlson, tested the boundaries of the technology using corporate branding and spokespersons like Rob Gronkowski.
The results were a mixed bag of professional-looking output and bizarre hallucinations. For example, when creating a figure of his director, Brandon, the AI inexplicably generated two Jeeps in the background. In another instance, Eric Carlson applied a "The Purge" movie theme to his own project, creating a "frighteningly" detailed figure that tested the AI’s content moderation filters.
The IP Wall
The limitations of AI became most apparent when Mike attempted to generate a figure of Jar Jar Binks from the Star Wars franchise. The system repeatedly flagged the request as a violation of intellectual property (IP) standards. This forced the AI to suggest "fantasy" alternatives, ultimately resulting in a generic "space alien" figure. This underscores a critical implication for the future of AI: as models become more sophisticated, the enforcement of copyright and licensing is being baked directly into the generation process.
The Human Element: Why We Continue to Experiment
Despite the technical frustrations—the extra teeth on grandchildren, the missing faces on robots, and the persistent obsession with "Groovy Bears"—the project remains a labor of love.
Whether it is his childhood crush, Susan Cowsill (the drummer he once idolized), or his grandson Milo (whose missing baby tooth the AI kept "fixing" against the user’s wishes), the project captures more than just pixels. It captures the human desire to preserve memories in a format that feels tangible.
As Mike notes in his updates, the process is an ongoing journey. "I still have more to add, I just need to find them all," he says. The project serves as a reminder that while AI may be prone to hallucination, the human intent behind the prompt is what gives these digital toys their meaning.
Conclusion: The Future of Generative Nostalgia
The evolution of Mike’s blog from a simple cruise recap to a complex AI-prompting masterclass reflects a broader societal trend. We are entering an age where personal history can be rendered in the style of mass-produced collectibles, provided we have the patience to navigate the "hallucinations" of the machine.
As OpenAI continues to refine models like o3 and o4-mini, we can expect these glitches to decrease. However, for now, the "errors" are part of the charm—a digital fingerprint of a technology still in its infancy, attempting to make sense of our world one "Groovy Bear" at a time.
For those looking to replicate these results, the key remains in the data: the more specific the input, the more grounded the output. Until then, we continue to iterate, refine, and occasionally laugh at the strange, beautiful, and sometimes nonsensical results generated by the silicon brain.
