AI detection can identify text from new LLMs because most models share a common "ancestry." They are either trained on the same foundational corpora, like Common Crawl, or fine-tuned with synthetic data from major models, giving them a detectable shared statistical fingerprint.

Pangram Labs' detector isn't hard-coded. It's a deep learning model trained on millions of examples. For each human text (e.g., a Yelp review), it sees an AI-generated equivalent, learning the subtle, often inarticulable, differences in word choice and structure that separate them.

To determine if one AI model has been trained on the output of another (a process called distillation), analysts use semantic similarity charts. These charts compare the diction and phrasing of different models. High correlation between a new model and existing ones, like Chinese model Kimi K3 and Anthropic's models, suggests potential distillation, though it isn't definitive proof.

A new form of analysis compares the semantic similarities (e.g., diction, phrasing) of outputs from different AI models. This technique is being used to create 'fingerprints' that can suggest if one model was illicitly 'distilled' or trained on the outputs of another, a key concern in the AI arms race.

To distinguish between light AI assistance (like Grammarly) and heavy generation, advanced detectors analyze the "cosine difference"—the distance in a multidimensional space between the original human text and the AI-edited version. This quantifies the degree of AI influence.

AI-generated text often uses devices like em-dashes or structuring ideas in threes. These aren't random; they're patterns learned from scraping skilled human writers like C.S. Lewis. This creates a paradox where the stylistic habits of good writing can now be misinterpreted as tells for AI.

While the em dash is a known sign of AI writing, a more subtle indicator is "contrastive parallelism"—the "it's not this, it's that" structure. This pattern, likely learned from marketing copy, is frequently used by LLMs but is uncommon in typical human writing.

Heuristics for spotting AI writing, like the overuse of em dashes, are becoming obsolete as models learn from human feedback. For instance, ChatGPT now uses em dashes *less* frequently than human writers at The Economist, flipping the old tell on its head and complicating detection efforts.

The AI detection arms race now includes "humanizers": specialized LLMs that rewrite AI-generated text to evade detection. This is the modern, sophisticated version of the old plagiarism tactic of running copied text through a thesaurus to change just enough keywords to fool detectors and teachers.