Detecting AI Through Stylistic Analysis

Pangram, founded by Stanford AI and machine learning graduates Max Spero and Bradley Emi, has raised $9 million to address the rise of "AI slop"—low-quality, automated content flooding the internet. Unlike detection methods that rely on hidden watermarks or metadata—which are easily stripped or bypassed—Pangram’s approach uses a large machine learning model trained on millions of human-authored documents.

To train the system, the company creates a "synthetic mirror" of human documents, replicating their length, topic, and tone using frontier LLMs. By comparing these mirrors against human originals, the model learns to identify the consistent stylistic choices and patterns unique to AI generation. This allows the system to detect AI-assisted content even when it has been lightly edited by a human.

Multi-Modal Detection and Real-World Application

Beyond text, Pangram is expanding into image detection. Its image model analyzes pixel-level statistical distributions to distinguish between real photos and AI-generated imagery, regardless of the model used to create the image. This is a significant departure from watermark-based checks, which are typically limited to content produced by a specific company's own models (e.g., Google or OpenAI).

Pangram offers several ways to integrate this technology:

• Consumer Tools: A $20/month web subscription and a browser extension that provides real-time "health scores" for content on platforms like X, LinkedIn, and Substack.
• API Integration: The company provides an API for platforms like Substack, schools, and publishers to verify the provenance of content.

The startup claims a false positive rate of approximately one in 10,000 for human documents, though real-world testing indicates that while the model is highly effective at identifying pure AI content, it can occasionally misidentify heavily edited or specific human-written styles as AI-assisted.