Pangram, a New York-based AI detection startup on a mission to combat the spread of AI slop on the internet, just raised $9 million on a bet that the demand for tools to distinguish between human-generated content and AI-generated text will continue to grow.
Pangram’s funding, led by Menlo Ventures with participation from Haystack, ScOp, Script Capital, and Cadenza, comes as the startup also announced its next-generation AI text detection model, Pangram 4, and its AI image detection model, Pangram Image.
Pangram says its new text detection model is more than 99% accurate at detecting AI-assisted text and mixed human-AI content, and can more easily detect AI humanizer programs. AI Image Detector is currently only available via Research Preview. Pangram plans to release more widely in the coming weeks.
Stanford University AI and Machine Learning graduates Max Spero and Bradley Emi launched Pangram about two years ago after the launch of ChatGPT opened the floodgates of an internet filled with bots, AI-generated SEO-bad content, and what Spero calls “Russian disinformation campaigns from LLM and UAE-inspired campaigns on Twitter.”
“I think it’s extremely valuable to know whether what you’re looking at is generated by AI or not,” Spero told TechCrunch. “Especially because the text you’re reading changes how people approach the text. Is this something that you have to be wary of hallucinations and jump in with skepticism, or is this something that I believe has been well researched by real journalists?”
Pangram’s AI detection system is essentially a large-scale machine learning model 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, created by Frontier’s LLM.
“Our model consistently learns the differences in styles and the choices the AI makes, and we can use that to learn with a high degree of confidence what the AI has produced,” Spero said, adding that the AI detector doesn’t rely on copy-and-paste metadata or hidden watermarks.
For Pangram, AI detection is about more than just determining whether a piece of text was written entirely by AI. It is also important to differentiate between levels of AI assistance. For example, someone might write something themselves and then ask an AI to edit or clean it up. Spero believes that AI assistance is acceptable as long as the author is transparent about the use of AI.
Pangram’s arrival comes at a time when the use of AI is becoming more commonplace. In some cases, mistakes can lead to ridicule, such as the Canadian politician who read an AI prompt in a speech to MPs. In other cases, the consequences can be sanctions and fines, like certain lawyers using fake citations created by ChatGPT to make their claims.
This backlash is also beginning to appear in institutional rules.
Open access archive arXiv introduced a new enforcement policy this year, stating that posts containing evidence that the author did not review the LLM output (such as hallucinatory references or meta comments such as “Are you sure you want to make changes?”) can trigger a one-year posting ban.
Pangram is not alone in predicting even greater demand for AI detection. Competitors such as Winston AI, Originality.ai, Copyleaks, and GPTZero are chasing the same demand, each building their own detectors.
Pangram’s technology isn’t perfect, but it could help stoke resistance to the AI-generated content that floods the internet, courtrooms, and academic papers.
Users can access Pangram on the web via a $20/month subscription or download a Chrome extension that automatically labels posts on X, LinkedIn, Substack, Reddit, and Medium in real time. It also provides a feed health score that includes a breakdown of the percentage of human and AI content on the screen.
Pangram also offers its technology via API. Notably, Substack recently integrated Pangram’s technology into its platform to show readers that their favorite authors are writing newsletters using AI. Other API customers include Quora, schools and universities, publishers and agents, and recruiters, among others, according to Spero.
Is Pangram effective?
Spero said that Pangram’s model incorrectly labels approximately 1 in 10,000 human documents as AI, so I decided to put it to the test. Although the text detection model was very impressive, it wasn’t perfect. I easily flagged completely AI-generated news articles written by both ChatGPT and Claude, and I rarely fell for my attempts to edit the AI-generated text to sound more human. However, Pangram flagged a sentence I completely rewrote as written by an AI. Pangram also didn’t fall for my attempts to encourage ChatGPT and Claude to bypass AI detectors when generating content.
I also gave one of my articles to ChatGPT and Claude and asked them to polish it up. Pangram gave an AI-assisted score of 13%, which was probably about right, but the model was able to detect subtle word choice changes in some sentences and ignore them in others. Also, if some text was written by a human, it was flagged as AI-assisted. This is worth noting. Because when I gave Pangram the same full article I wrote, it got a 100% human score.
Perhaps the problem was that the news articles were a little dry and easy to sound like AI. So I tried a different tactic. I tested Pangram with my own more audio-rich personal Substack newsletter content, pasting the first half of the text into Pangram and asking ChatGPT and Claude to copy my style and write the second half. In most cases, Pangram easily detected human-written text and AI-written text.
In our limited testing of Pangram’s new image detection model, we saw similarly impressive results.
Pangram’s AI image detection system promises to identify AI-generated images across AI models, unlike OpenAI and Google DeepMind’s watermark-based checks, which primarily detect their own output. It operates based on pixel-level distributions and learns subtle statistical differences between real photos and AI-generated images. Spero said the model can also detect AI images in real-world photos.
In my tests, the model easily detected AI-generated images, whether photorealistic or cartoonish. We also confirmed that this model was able to detect AI images in real-world photos. The heat map provided by Pangram is clearly lit on the image. However, in one instance, a photo of an AI-generated image was incorrectly labeled as human content.
Spero says he doesn’t want his technology to fuel a witch hunt against people who use AI in their writing, but says there needs to be some mechanism to counter that trend.
“The future I see is that AI content will continue to proliferate,” Spero said. “We’re getting new GPUs faster than new people are being born. If we don’t actively discriminate against human content, more and more AI will just drown out the human signals.”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
