We introduce introduce "slop forensics", a novel technique developed by Samuel J. Paech to determine the lineage and training influences of Large Language Models (LLMs). This method analyzes the unique linguistic quirks, or "slop," in an AI's output to identify patterns resembling those of other models, similar to DNA sequencing in biology. By creating "slop profiles" and comparing them using algorithms, researchers can construct a phylogenetic tree showing how models relate to each other. A notable case study with DeepSeek R1 demonstrates how the technique revealed a shift in the model's "linguistic parents" from OpenAI to Google's Gemini after an update. This research is crucial for transparency, intellectual property concerns, and AI safety, helping to track potential biases and avoid a "model collapse" where AIs trained on other AIs become less diverse and more flawed.

Slop Forensics Unravels LLM Family Trees

delimiterbob