June 24, 2026
Bot or not? The comments erupt
Exploring the internal representations of Pangram 3.3.2
AI detector peeks inside its own brain — and commenters are already poking holes in it
TLDR: Pangram says its newest AI-writing detector is better at telling human text from bot text and is now studying the clues it uses internally. Commenters are split between being impressed by its accuracy and worrying it may really be tracking writing style — while others say clever prompts can still game it.
Pangram, a company that builds tools to spot machine-written text, just pulled back the curtain on how its latest detector, Pangram 3.3.2, "thinks" when deciding whether something was written by a person or a bot. The company says its model is better at catching newer AI writing while making fewer mistakes on human writing, especially from non-native English speakers. To do that, it looked at thousands of writing samples — from news stories and product reviews to Reddit posts and books — and mapped the hidden patterns inside the system. In plain English: Pangram is trying to figure out what vibes its own AI detector is picking up on.
But the real fireworks are in the reactions. One commenter immediately went full sci-fi thriller, wondering if the model isn’t just detecting AI, but quietly learning to recognize a person’s unique writing voice too. That’s the kind of idea that makes readers go, "Wait... is this thing profiling writers now?" Another user, who says they actually use Pangram a lot, gave the company a rare mixed review: praise for its super-low false alarms, then a sharp jab that AI can still be prompted to sound "100% human." Ouch. In other words, fans are impressed, skeptics smell loopholes, and the research crowd is already chanting for a sequel — specifically Sparse Autoencoders, the ultra-nerdy follow-up one commenter is begging for. Even Pangram’s own jokes about dodging "em-dashes" and the word "delve" gave the whole thing a very online, very self-aware energy.
Key Points
- •Pangram Labs published an interpretability-focused article on Pangram 3.3.2, its 2026 AI text detection model.
- •The company says its detection system is a fine-tuned LLM for sequence classification and does not rely on handcrafted features such as perplexity or burstiness.
- •For the analysis, Pangram built a balanced 5,000-document held-out dataset split evenly between human and AI text and examined activations across 20 even-numbered layers.
- •The dataset includes outputs from multiple LLM families, including Claude, GPT, Gemini, DeepSeek, and Qwen, across domains such as news, scientific abstracts, reviews, books, Wikipedia, and ESL writing.
- •Pangram analyzes 5,120-dimensional hidden activations from the EditLens architecture using dimensionality reduction methods such as PCA rather than focusing on the final ai_assistance_score.