With the AI models being trained using Reddit data, do you think by now someone somewhere would have gotten shittymorph’ed?

Exploring AI’s Knowledge of Niche Reddit Subcultures: Have Language Models Been Indicted by Internet Memes?

In recent discussions about artificial intelligence, a recurring question revolves around the datasets used to train these models. Given that many AI language models are developed using vast amounts of data scraped from platforms like Reddit, it’s natural to wonder: Have these models been exposed to the most obscure and niche internet subcultures—and, more intriguingly, have they ever encountered content that could be considered offensive or outlandish?

This curiosity was sparked anew when I decided to test a popular AI assistant, Google’s Gemini, by prompting it to respond in a “shittymorph” style—an internet slang and meme-heavy speaking pattern originating from densely populated Reddit communities and online meme culture. The result? It did not disappoint, delivering responses that stayed true to the gritty, meme-laden tone.

This experiment raises a broader point: can we gauge the depth of these models’ understanding of the more clandestine corners of Reddit? By exploring more obscure and less mainstream Reddit lore, could we uncover the extent of what AI models truly know—and perhaps identify gaps or biases in their training data?

If you have ideas or experiences about probing AI’s familiarity with internet culture’s most esoteric elements, I’d love to hear your insights. Understanding how these models navigate the vast and varied landscape of online communities can shed light on their capabilities—and their limitations—when it comes to recognizing the complex and colorful language that shapes digital subcultures.