Pangram raises $9M to detect AI-generated content as internet faces ‘slop’ crisis

New York-based AI detection startup Pangram has raised $9 million in a funding round led by Menlo Ventures, with participation from Haystack, ScOp, Script Capital, and Cadenza, as the company launches its next-generation text and image detection models. The investment underscores growing demand for tools that can distinguish human-generated content from AI-generated text, a market Pangram believes will expand as AI-generated content — often called ‘AI slop’ — continues to flood the internet.

Pangram’s new detection models target AI-assisted writing and images

Pangram announced the launch of Pangram 4, its latest AI text detection model, alongside Pangram Image, an AI image detection model currently available in research preview. The company claims Pangram 4 is over 99% accurate at identifying AI-assisted writing and mixed human-AI content, and can also detect text generated by AI humanizer programs designed to evade detection. The image detector, which Pangram plans to release more widely in the coming weeks, analyzes pixel-level distributions to spot AI-generated images across different AI models, unlike watermark-based systems that only detect output from specific platforms like OpenAI or Google DeepMind.

Pangram’s detection system is built on a large machine learning model trained on tens of millions of known human documents. The startup created a ‘synthetic mirror’ for each document — replicating the topic, length, and tone of voice using a frontier LLM — to teach the model the stylistic differences and consistent choices made by AI. According to co-founder Max Spero, the detector does not rely on copy-paste metadata or hidden watermarks.

Why AI detection matters for publishers, academics, and legal professionals

The need for reliable AI detection has grown as AI-generated content appears in more contexts, from academic papers to legal filings. The open-access archive arXiv introduced a new enforcement policy this year, stating that submissions containing evidence that authors failed to review LLM output — such as hallucinated references or meta-text like ‘Would you like me to make any changes?’ — can trigger a one-year submission ban. In legal settings, lawyers have faced sanctions and fines for using fake citations created by ChatGPT. Spero told Bitcoin World that the value of detection lies in helping readers approach text with appropriate skepticism: ‘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?’

How Pangram’s technology works and who is using it

Pangram offers its detection technology through a $20-per-month subscription on the web and a Chrome extension that automatically labels posts in real time on X, LinkedIn, Substack, Reddit, and Medium. The extension also provides a feed health score showing the percentage breakdown of human versus AI content on a user’s screen. The company also offers an API, which has been integrated by Substack to show readers which authors use AI to write their newsletters. Other API customers include Quora, schools, universities, publishers, and recruiters, per Spero.

In testing by Bitcoin World, Pangram’s text detection model easily flagged entirely AI-generated news articles written by ChatGPT and Claude, and was rarely fooled by attempts to edit the AI-generated text into sounding more human. However, the model occasionally flagged human-written sentences as AI-generated, particularly in dry or formulaic writing styles. The image detection model performed well in identifying AI-generated imagery, including AI images embedded in real-world photos, though it incorrectly labeled one photo of an AI-generated image as human content.

Competition and the future of AI detection

Pangram faces competition from other AI detection startups, including Winston AI, Originality.ai, Copyleaks, and GPTZero, each building its own detector. Spero emphasized that he does not want the technology to fuel a witch hunt against people using AI for writing, but sees it as a necessary mechanism to push back against the proliferation of AI content. ‘If we do not actively discriminate in favor of human content, then we’re just gonna get more and more AI, and it’s just gonna drown out any human signal that we have,’ he said.

Conclusion

Pangram’s $9 million raise and new detection models arrive at a time when AI-generated content is becoming increasingly common across the internet, from social media to academic publishing. While no detection system is perfect, the company’s technology — used by platforms like Substack and available to individual users — represents a growing effort to help readers, publishers, and institutions verify the authenticity of the content they encounter. As AI models continue to improve, the race to build reliable detection tools will likely intensify.

FAQs

Q1: How does Pangram’s AI detection work?
Pangram uses a large machine learning model trained on millions of human documents and their AI-generated ‘mirrors’ to learn stylistic differences between human and AI writing. It does not rely on watermarks or metadata.

Q2: Is Pangram’s AI detection accurate?
The company claims Pangram 4 is over 99% accurate at detecting AI-assisted writing. In testing, it performed well but occasionally flagged human-written text as AI-generated, particularly in dry or formulaic writing styles.

Q3: Who is using Pangram’s technology?
Substack has integrated Pangram’s API to show readers which authors use AI. Other customers include Quora, schools, universities, publishers, and recruiters. Individual users can access it via a $20-per-month subscription or a Chrome extension.