New York-based startup Pangram has raised $9 million in a Series A round led by Menlo Ventures, with participation from Haystack, ScOp, Script Capital, and Cadenza. The fundraise coincides with the launch of Pangram 4, a next-generation AI text detection model the company claims is over 99% accurate, and Pangram Image, an AI image detection model currently in research preview.
Stanford AI graduates Max Spero and Bradley Emi launched Pangram roughly two years ago. Their pitch: as autonomous agents and LLMs generate content at scale across the internet, someone needs to build the detection infrastructure that tells publishers, recruiters, and institutions what was written by a human and what was not.
How the Detection Works
Pangram’s system is a large machine learning model trained on tens of millions of known human documents. For each document, the team created a “synthetic mirror,” replicating the topic, length, and tone but written by a frontier LLM. The model learns the stylistic differences between human and AI writing to classify content with high confidence, according to co-founder Spero in an interview with TechCrunch.
“Our model is learning the stylistic differences and the choices that AI makes consistently,” Spero told TechCrunch. The detector does not rely on copy-paste metadata or hidden watermarks, and it can distinguish between fully AI-generated text and human-written content that was edited or cleaned up by AI.
Institutional Demand is Growing
The startup’s integrations reflect where demand is sharpest. Substack uses Pangram for author disclosure. Quora and multiple universities use it for content verification. Recruiters use it to screen job applications. The open-access archive arXiv introduced an enforcement policy this year banning submissions with evidence of unreviewed LLM output, including hallucinated references or meta-comments like “Would you like me to make any changes?”
Consumers can access Pangram via a $20-per-month subscription or a Chrome extension that automatically labels posts in real time on X, LinkedIn, Substack, Reddit, and Medium. It provides a feed health score showing the percentage breakdown of human versus AI content on screen.
A Crowded Detection Market
Pangram competes against Winston AI, Originality.ai, Copyleaks, and GPTZero, all building their own detectors. The proliferation of agent-generated content, from autonomous SEO tools to AI writing assistants, makes the category’s growth trajectory clear. The question is whether detection accuracy can keep pace as models improve. Pangram’s bet on a $9 million war chest: that enterprises and publishers will pay for authenticity infrastructure the same way they pay for cybersecurity.