AI detection startup in New York Pangram is on a mission to fight Infection with artificial intelligence It went viral and raised $9 million in a bet that demand for tools that distinguish human-generated content from AI-generated text would only grow.
Led by Menlo Ventures, with participation from Haystack, ScOp, Script Capital and Cadenza, Pangram’s fundraising comes as the startup also launched Pangram 4, a next-generation AI text recognition model, and Pangram Image, an AI image recognition model.
Pangram says the new text-recognition model is more than 99% accurate at finding AI-powered text and mixed human-AI content, plus it can more easily detect AI humanizers. The AI image detector is currently only available via research preview; Pangram plans to release it more widely in the coming weeks.
Stanford AI and machine learning grads Max Spero and Bradley Amy launched Pangram nearly two years ago after ChatGPT’s launch opened the floodgates to an internet filled with bots, AI-generated SEO slop content and what Spero calls “LLM-powered Russian disinformation campaigns and UAE influencer campaigns on Twitter.”
“I think it’s incredibly valuable to know if what you’re looking at was generated by AI,” Spero told TechCrunch. “Write the text you read, especially because it changes the way people approach the text. Is this something I should be hallucinating and going into with skepticism, or is it well-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 that replicated the subject, length, and tone of voice, but was written by Frontier LLM.
“Our model learns the stylistic differences and the choices the AI consistently makes and can use that to learn what the AI generated with high confidence,” Spero said, adding that the AI detector doesn’t rely on copy-pasted metadata or hidden watermarks.
AI detection for Pangram isn’t just about whether a piece of text was actually written by an AI. It’s also about differentiating between AI and levels of assistance – like a person writing something themselves, but then asking an AI to edit or clean it up. Spero believes that AI assistance can be acceptable as long as the writer discloses the use of artificial intelligence.
The emergence of Pangram comes at a time when the use of artificial intelligence is becoming more common. In some cases, as Canadian politician reading AI survey Mistakes in his loud speeches to MPs result in laughter. In other cases, there may be consequences, as some lawyers have argued using fake citations created by ChatGPT. sanctions and fines.
This backlash is not just individual shaming or sanctions, but is also beginning to manifest itself 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 product (such as hallucinatory references or meta comments such as “Would you like me to make any changes?”) may result in a one-year submission ban.
Pangram isn’t the only bet that AI detection will be more sought after. Competitors such as Winston AI, Originality.ai, Copyleaks and GPTZero follow suit, each building their own detector.
While Pangram’s technology isn’t perfect, it could help bolster resistance to the acceptance of AI-generated content that floods the Internet, courtrooms, and academic papers.
Users can access Pangram online for a $20 monthly subscription or download a Chrome extension that automatically tags posts on X, LinkedIn, Substack, Reddit and Medium in real-time. It also provides a nutrition health score with a percentage breakdown of human and AI content on your screen.
Pangram also offers its technology via an API. Note that Substack recently integrated Pangram Technology to its platform to show readers which of their favorite authors write newsletters using AI. Other API customers include Quora, schools and universities, publishers and agents, and recruiters for Spero.
Does pangram work?
Spero said that about one in 10,000 human documents are incorrectly labeled as AI by the Pangram model, so I decided to put it to the test. The text detection model was very impressive, but not perfect. It easily flagged completely 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 the sentences that I completely rewrote as if they were written by artificial intelligence. My attempts to encourage ChatGPT and Claude to avoid AI detectors when creating content didn’t fool Pangram at all.
I also gave ChatGPT and Claude one of my articles and asked them to polish it. Pangram gave it an AI-assisted score of 13%, which is probably close to accurate, but the model was able to detect subtle word choice changes in some sentences and ignore them in others. He also marked some sentences as AI-powered when written by a human. This was remarkable because when I submitted the entire article to Pangram as I wrote it, it scored 100% human.
Perhaps 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 more vocal, 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. In most cases, Pangram easily detected human-written text compared to AI-written text.
Limited tests of Pangram’s new image detection model have been 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 mostly detect their own output. It works on pixel-level distributions, studying subtle statistical differences between real photos and AI-generated images. Spero says the model can even detect an AI image appearing inside a real-world photograph.
In my test, the model easily detected AI-generated images, whether photorealistic or animated. I can also confirm that the model can detect an AI image appearing in a real-world photo – the heatmap Pangram provides clear illumination over the image – although in one case it incorrectly labeled an AI-generated image as human content.
Spero says he doesn’t want his technology to power a witch hunt against people who use AI for writing, but it needs to have some sort of fallback mechanism.
“The future I see is that AI content continues to grow,” 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, then we’re going to see more and more AI, drowning out any human signal we have.”
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.
