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

AI detection startup in New York Pangram is on a mission to combat the AI ​​contagion spreading across the internet, and 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 fundraising – led by Menlo Ventures, with participation from Haystack, ScOp, Script Capital and Cadenza – comes as the startup is also launching its next-generation AI text detection model, Pangram 4, and an AI image detection model, Pangram Image.

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

Stanford AI and machine learning graduates Max Spero and Bradley Emi launched Pangram about two years ago, after the launch of ChatGPT opened the floodgates to 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 incredibly valuable to know whether what you’re looking at is something AI-generated or not,” Spero told TechCrunch. “Especially the text you’re reading, because it changes the way people approach the text. Is this something I should watch out for hallucinations and approach with skepticism, or is this something I trust has been properly researched by a real journalist?”

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

“Our model learns the stylistic differences and choices that AI makes consistently and can use that to learn with high confidence what makes something AI-generated,” Spero said, adding that the AI ​​detector does not rely on copy-and-pasted metadata or hidden watermarks.

For Pangram, AI detection is not just about whether or not a piece of text is written entirely by AI. It’s also about distinguishing between levels of AI support, as in the case of someone writing something themselves, but then asking AI to edit or clean it up. Spero believes that AI assistance can be acceptable, as long as the writer makes their use of AI public.

The rise of Pangram comes at a time when the use of AI is becoming increasingly common. In some cases, such as the Canadian politician who read an AI prompt out loud in a speech to lawmakers, the mistakes result in ridicule. In other cases, such as with certain lawyers arguing their case using fake quotes created by ChatGPT, the consequences may be the same sanctions And fines.

This counter-reaction not only costs individuals shame or sanctions, but is also starting to become visible in institutional rules.

The open-access archive arXiv introduced a new enforcement policy this year, stating that submissions containing evidence that authors have not reviewed the LLM output (such as hallucinatory references or meta-comments such as: “Do you want me to make changes?”) may trigger a one-year submission ban.

Pangram isn’t alone in betting that AI detection will become increasingly popular. Competitors such as Winston AI, Originality.ai, Copyleaks and GPTZero are pursuing the same question and are each building their own detector.

Pangram’s technology, while not perfect, could fuel resistance to accepting the AI-generated content flooding the internet, courtrooms and academic papers.

Users can access Pangram through a $20 per month web subscription or download the Chrome extension, which automatically labels posts in real time on X, LinkedIn, Substack, Reddit and Medium. It also offers a feed health score with a percentage breakdown of human versus AI content on your screen.

Pangram also offers its technology via API. Notably, Substack recently integrated Pangram’s technology into its platform to show readers which of their favorite authors write their newsletters using AI. Other API customers include Quora, schools and universities, publishers and agents, and recruiters per Spero.

Does Pangram work?

Spero said that about one in 10,000 human documents is 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 fully AI-generated news articles written by both ChatGPT and Claude, and was rarely fooled by my attempts to edit the AI-generated text to sound more human. At the same time, Pangram marked sentences that I completely rewrote as AI-written. Pangram also wasn’t fooled at all by my attempts to get ChatGPT and Claude to avoid 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 an AI-assisted score of 13%, which was probably close to accurate, but the model was able to detect subtle changes in word choice in some sentences and ignore them in others. It also flagged some sentences as AI-powered when they were written by humans. That was remarkable, because when I gave Pangram that same article in its entirety, as I wrote it, it got a 100% human score.

Maybe the problem was that news articles can be a bit dry and can easily sound like AI. So I tried a different tactic. I tested Pangram on my own, more high-profile, 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 was able to easily detect human-written text versus AI-written text.

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

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

During my testing, the model easily detected AI-generated images, whether photorealistic or cartoonish. I can also confirm that the model was able to detect an AI image appearing on a real photo – the Pangram heatmap provides clear illumination of the image – although in one case it incorrectly labeled a photo of an AI-generated image as human content.

Spero says he doesn’t want his technology to spark a witch hunt against people using AI to write, but that there needs to be some kind of mechanism to combat this waste.

“The future I see is that AI content continues to spread,” Spero said. “We’re getting new GPUs faster than new humans are being born. If we don’t actively discriminate in favor of human content, we’ll see more and more AI, and it will drown out any human signal we have.”