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

Spero said roughly one in 10,000 human documents are incorrectly labeled as AI with Pangram’s model, so I decided to put it to the test. The text detection model was very impressive but not perfect. It easily flagged entirely AI-generated news articles written by both ChatGPT and Claude, and was rarely fooled by my attempts to edit the AI-generated text into sounding more human. At the same time, Pangram did flag sentences that I completely rewrote as being AI-written. Pangram also wasn’t at all fooled by my attempts to prompt ChatGPT and Claude into evading AI detectors when generating content.  

I also gave ChatGPT and Claude one of my own articles and asked them to polish it up. Pangram gave it a 13% AI assisted score, which was probably close to accurate, but the model was able to detect subtle word-choice changes in some sentences and ignored them in others. It also flagged some sentences as AI-assisted when they were human written. That was notable because when I gave Pangram that same article in its entirety, as I had written it, it got a 100% human score. 

Maybe the problem was that news articles can be a bit dry and could easily sound like AI. So I tried a different tactic. I tested Pangram on my own more voicey, personal Substack newsletter content, pasting the first half of the text into Pangram and then asking ChatGPT and Claude to copy my style and write the second half. For the most part, Pangram easily detected human-written text versus AI-written text.

My limited testing of Pangram’s new image detection model turned out to be equally impressive. 

Pangram’s AI image detection system promises to spot AI-generated images across AI models, unlike OpenAI’s or Google DeepMind’s watermark-based checks, which mostly detect their own output. It works on pixel-level distributions, learning subtle statistical differences between real photos and AI-generated images. Spero says the model can even detect an AI image that appears inside a real-world photo. 

In my testing, the model easily detected AI-generated imagery, whether it was photorealistic or cartoonish. I can also confirm the model could detect an AI image appearing in a real-world photo — the heat map Pangram provides clearly lighting up over the image — though in one instance it incorrectly labeled a photo of an AI-generated image as human content. 

Spero says he doesn’t want his technology to fuel a witch hunt against people using AI for writing, but that there needs to be some sort of mechanism to push back against the slop. 

“The future that I see is that AI content just continues to proliferate,” Spero said. “We’re getting new GPUs faster than new people are being born. If we do not actively discriminate in favor of human content, then we’re just gonna see more and more AI, and it’s just gonna drown out any human signal that we have.”