As AI content floods the internet, Pangram raises $9M to detect it

New York-based AI detection startup Pangram is on a mission to combat the AI slop infestation spreading across the internet, and it just raised $9 million on a bet that demand for tools that distinguish human-generated content from AI-generated text will only grow.

Pangram’s fundraise — led by Menlo Ventures, with participation from Haystack, ScOp, Script Capital, and Cadenza — comes as the startup also launches its next-generation AI text detection model, Pangram 4, and an AI image detection model, Pangram Image.

Pangram says the new text detection model is over 99% accurate at finding AI-assisted writing and mixed human-AI content, plus it can more easily detect AI humanizer programs. The AI image detector is only available via research preview for now; Pangram plans to release it more widely in the coming weeks.

Stanford AI and machine learning grads Max Spero and Bradley Emi launched Pangram about two years ago, after the launch of ChatGPT opened the floodgates for an internet full of bots, AI-generated SEO slop content, and what Spero calls “LLM-powered Russian disinformation campaigns and UAE-influenced campaigns on Twitter.”

“I think it’s just incredibly valuable to know whether what you’re looking at is something that’s AI-generated or not,” Spero told TechCrunch. “Especially text that you’re reading, because it changes how people approach the text. 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?”

Pangram’s AI detection system is essentially a large machine learning model that was trained on tens of millions of known human documents. The startup then created a “synthetic mirror” for each document, replicating the topic, length, and tone of voice, but written by a frontier LLM.

“Our model is learning the stylistic differences and the choices that AI makes consistently and is able to use that to learn what makes something AI-generated with high confidence,” Spero said, adding that the AI detector isn’t relying on copy-paste metadata or hidden watermarks.

For Pangram, AI detection isn’t just about whether or not a piece of text was written entirely by AI. It’s also about distinguishing between levels of AI assistance — like in the case of someone who writes something themselves, but then asks AI to edit or clean it up. Spero believes AI assistance can be acceptable, just so long as the writer discloses their use of AI.