Pangram, an AI detection startup based in New York, is on a mission to tackle the growing deluge of AI-generated content flooding the internet. The company has just secured $9 million in funding, betting that the demand for tools capable of distinguishing human-written text from machine-made content will keep rising. The investment round was led by Menlo Ventures, with backing from Haystack, ScOp, Script Capital, and Cadenza. This funding coincides with the launch of Pangram’s latest AI text detection model, Pangram 4, alongside a new AI image detection model, Pangram Image. According to the startup, the updated text detector boasts over 99% accuracy in identifying AI-assisted writing and mixed human-AI content, and it can more effectively spot AI humanizer programs. The image detector is currently available only as a research preview, with a wider release planned in the coming weeks.

Pangram was founded about two years ago by Stanford AI and machine learning graduates Max Spero and Bradley Emi, following ChatGPT’s debut, which unleashed a wave of bots, AI-generated SEO slop, and what Spero describes as “LLM-powered Russian disinformation campaigns and UAE-influenced campaigns on Twitter.” He emphasizes the importance of discerning the source of text: “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 detection system is built on a large machine learning model trained on tens of millions of known human documents. For each document, the startup created a “synthetic mirror”—a version covering the same topic, length, and tone, but generated 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 explained, noting that the detector does not rely on metadata or hidden watermarks. For Pangram, AI detection goes beyond simply determining if text is entirely AI-written; it also assesses varying degrees of AI assistance, such as when someone writes a draft and then asks AI to edit or polish it. Spero believes AI assistance can be acceptable as long as the writer discloses its use.

Pangram’s launch arrives as AI usage becomes increasingly common, sometimes with embarrassing results—like the Canadian politician who read an AI prompt aloud during a speech. In more serious cases, lawyers have faced sanctions for using fake citations generated by ChatGPT. The backlash is now influencing institutional policies. For instance, the open-access archive arXiv introduced a new enforcement policy this year, stating that submissions showing evidence authors failed to review LLM output—such as hallucinated references or meta comments like “Would you like me to make any changes?”—can result in a one-year submission ban. Pangram is not alone in betting on the growing need for AI detection. Competitors like Winston AI, Originality.ai, Copyleaks, and GPTZero are also developing their own detectors. While not flawless, Pangram’s technology could help counter the flood of AI-generated content in online spaces, courtrooms, and academic papers.

Users can access Pangram through a $20-per-month subscription on the web or via a Chrome extension that automatically labels posts in real time on platforms like X, LinkedIn, Substack, Reddit, and Medium. It also provides a feed health score showing the percentage breakdown of human versus AI content on a user’s screen. Additionally, Pangram offers its technology through an API. Notably, Substack has integrated Pangram’s system to let readers see which authors use AI to write their newsletters. Other API customers include Quora, schools and universities, publishers, agents, and recruiters, according to Spero.

Spero claims that roughly one in 10,000 human documents are incorrectly flagged as AI by Pangram’s model. To test this, I ran several experiments. The text detection model proved highly impressive, though not perfect. It easily flagged entirely AI-generated news articles from both ChatGPT and Claude, and it was rarely fooled by my attempts to edit the AI text to sound more human. However, Pangram did flag some sentences I completely rewrote as AI-written. It also resisted my efforts to prompt ChatGPT and Claude to evade detection. When I gave ChatGPT and Claude one of my own articles and asked them to polish it, Pangram assigned a 13% AI-assisted score, which seemed close to accurate. Yet, the model detected subtle word-choice changes in some sentences while ignoring others, and it flagged some human-written sentences as AI-assisted. Interestingly, when I fed Pangram the same article as I originally wrote it, the model gave it a 100% human score. This might be because news articles can sound dry and resemble AI output. To explore further, I tested Pangram on my more personal Substack newsletter content, pasting the first half and asking ChatGPT and Claude to mimic my style for the second half. For the most part, Pangram successfully distinguished human from AI text. My limited testing of Pangram’s new image detection model was equally impressive.

Pangram’s AI image detection system promises to identify AI-generated images across different models, unlike watermark-based checks from OpenAI or Google DeepMind, which primarily detect their own output. It works by analyzing pixel-level distributions, learning subtle statistical differences between real photos and AI-generated images. Spero says the model can even detect an AI image embedded within a real-world photo. In my tests, the model easily identified AI-generated imagery, whether photorealistic or cartoonish. I also confirmed its ability to detect an AI image within a real photo—the heat map Pangram provides clearly highlighted the image—though in one case, it incorrectly labeled a photo of an AI-generated image as human content. Spero emphasizes that he doesn’t want his technology to spark a witch hunt against AI users, but believes there must be a way 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.”