Pangram is positioning itself for a future in which identifying AI-generated content becomes a routine part of reading, publishing, hiring, teaching, and moderating online spaces. The New York-based AI detection startup has raised $9 million while rolling out Pangram 4, its next-generation AI text detection model, and Pangram Image, a new image detection system now available in research preview.
The round was led by Menlo Ventures, with participation from Haystack, ScOp, Script Capital, and Cadenza. The timing reflects a broader bet: as AI-generated text and images spread across the internet, demand may grow for tools that help people understand whether content was written, edited, or produced by AI.
What Pangram Is Launching
Pangram says Pangram 4 is over 99% accurate at identifying AI-assisted writing and mixed human-AI content. The company also says the model is better at detecting AI humanizer programs, which are designed to make machine-generated text appear more human.
The company is also moving beyond text. Pangram Image is currently available through a research preview, with a wider release planned in the coming weeks. According to the source, the image detector is designed to identify AI-generated images across AI models, rather than relying only on watermark-based systems that mostly recognize output from a specific provider.
For text, Pangram’s system is described as a large machine learning model trained on tens of millions of known human documents. The company then created a synthetic version of each document that preserved the topic, length, and tone of voice, but was written by a frontier LLM. That gave the model a basis for learning recurring stylistic choices associated with AI-generated writing.
Max Spero said the detector does not depend on copy-paste metadata or hidden watermarks. Instead, it looks for patterns in writing style and word choice that can separate human work from AI-generated or AI-assisted text with high confidence.
Why AI Detection Is Becoming A Business
Pangram was launched about two years ago by Stanford AI and machine learning grads Max Spero and Bradley Emi. The company’s origin sits in the period after the launch of ChatGPT, when AI-generated posts, bot content, SEO-focused material, and political influence campaigns became a bigger concern for online platforms and readers.
The company’s pitch is not limited to catching fully automated writing. Pangram is also trying to identify degrees of AI assistance. That matters because a piece of writing may begin with a human author and later be edited, polished, or cleaned up with AI. Spero’s view, as described in the source, is that AI assistance can be acceptable when it is disclosed.
That distinction gives Pangram a broader role than a simple yes-or-no detector. Its product is aimed at helping readers, institutions, and platforms make judgments about how to treat content. A fully AI-generated article, a human draft lightly edited by AI, and a human-written article with no AI assistance may carry different levels of trust depending on the setting.
The source points to several examples of why this has become sensitive. A Canadian politician read an AI prompt aloud in a speech to lawmakers. Lawyers using fake citations created by ChatGPT have faced sanctions and fines. The open-access archive arXiv introduced a new enforcement policy this year that can trigger a one-year submission ban when submissions show evidence that authors failed to review LLM output, including hallucinated references or meta comments such as, "Would you like me to make any changes?"
Where Pangram Fits Into The Market
Pangram is not alone in chasing this demand. Winston AI, Originality.ai, Copyleaks, and GPTZero are also building AI detection tools. The market exists because many organizations now face a practical question: how should they evaluate content when AI can produce text and images at scale?
Pangram offers several routes into its system. Individual users can access it through a $20-per-month web subscription. The company also offers a Chrome extension that labels posts in real time on X, LinkedIn, Substack, Reddit, and Medium. The extension also provides a feed health score with a percentage breakdown of human versus AI content visible on the user’s screen.
For organizations, Pangram provides its technology through an API. Substack recently integrated Pangram’s technology into its platform to show readers which authors write newsletters using AI. Other API customers include Quora, schools and universities, publishers and agents, and recruiters, among others, according to Spero.
Those use cases show how AI detection could become part of different workflows. A publishing platform may use it to increase transparency with readers. A school or university may use it when reviewing student work. Recruiters may apply it to written material submitted during hiring. Each use case raises its own judgment calls, especially when content is only partly AI-assisted.
The Limits Of Detection
The source makes clear that Pangram’s technology is not perfect. Spero said roughly one in 10,000 human documents are incorrectly labeled as AI by Pangram’s model. In the source’s testing, Pangram performed strongly on entirely AI-generated articles from ChatGPT and Claude and was rarely fooled by edited AI-generated text. It also was not fooled by attempts to prompt ChatGPT and Claude to evade AI detectors.
At the same time, the model sometimes labeled fully rewritten human sentences as AI-written. In another test, when an article was polished by ChatGPT and Claude, Pangram returned a 13% AI assisted score. The model detected some subtle word-choice changes and missed others, while also flagging some human-written sentences as AI-assisted. When the original article was tested in full as written, Pangram scored it as 100% human.
The image detector also showed strengths and limits in the source’s testing. It detected AI-generated imagery, including photorealistic and cartoonish examples. It also detected an AI image appearing inside a real-world photo, with a heat map highlighting the area. In one case, however, it incorrectly labeled a photo of an AI-generated image as human content.
That mixed picture is important. AI detection can provide useful signals, but the source does not support treating it as infallible proof. The stronger practical role may be as a transparency and review tool, especially when the stakes involve publishing, academic submissions, legal work, or trust in online feeds.
The Bigger Question For Human Content
Spero said he does not want Pangram to drive a witch hunt against people who use AI for writing. His concern is the volume of low-quality AI-generated material and the difficulty of preserving a clear human signal online.
That is the central issue behind Pangram’s funding and product launch. As AI content becomes easier to produce, readers and institutions may need better ways to understand what they are seeing. Pangram’s answer is not to block every use of AI, but to make AI involvement more visible.
If that approach gains traction, AI detection could become less of a niche tool and more of a trust layer across platforms, schools, publishers, and professional workflows. Pangram’s $9 million raise suggests investors see that need expanding as AI-generated text and images continue to spread.