Pangram Raises $9M to Hunt AI Slop, Claims 99% Accuracy on Mixed Content

The internet’s AI slop problem just got a $9 million countermeasure. Pangram, a New York startup founded by two Stanford AI grads, raised a seed round led by Menlo Ventures to fuel its mission of distinguishing human writing from the rising tide of machine-generated text. The funding coincides with the launch of Pangram 4, a new detection model the company says is over 99% accurate at flagging AI-assisted writing and mixed human-AI content, along with a research preview of an image detector.

Co-founder Max Spero frames the tech as a tool for skepticism. “Is this something that I’m going to have to look out for hallucinations and jump in skeptically, or is this something that I trust was well-researched from an actual journalist?” he told TechCrunch. The underlying model was trained on tens of millions of human documents, then taught to spot an LLM’s stylistic tics by comparing them against a “synthetic mirror” of AI-written copies matching the topic, length, and tone. Spero insists it’s not leaning on metadata or watermarks, just the consistent choices frontier models make.

Demand for this kind of triage is moving from niche forums to institutional policy. The academic paper repository arXiv now threatens a one-year submission ban for authors who fail to review LLM output—like leaving in hallucinated citations or a stray “Would you like me to make any changes?” The consequences range from a Canadian politician’s public embarrassment to lawyers facing sanctions for citing fake ChatGPT cases. Pangram is chasing a market alongside GPTZero, Copyleaks, and others, already finding a home on Substack where readers can see if their favorite writers are using AI.

In a hands-on test, the model proved stubbornly good. It easily flagged articles from ChatGPT and Claude, rarely fell for my humanizing edits, and wasn’t fooled by prompts designed to evade detection—though it did occasionally mislabel completely rewritten sentences as AI, and Spero admits roughly one in 10,000 human documents gets flagged incorrectly. The $20 monthly subscription also powers a Chrome extension that slaps a real-time label on posts across X, LinkedIn, and Reddit. It’s not a magic bullet, but it’s the kind of arms race the internet needs right now.