New York-based AI detection startup Pangram has raised $9 million in funding — led by Menlo Ventures with participation from Haystack, ScOp, Script Capital, and Cadenza — and announced two new products: the next-generation text detection model Pangram 4 and an AI image detection model called Pangram Image.
What the new models do
Pangram says Pangram 4 is over 99% accurate at identifying AI-assisted writing and mixed human–AI content, and that it can more readily spot tools that ‘‘humanize’’ AI-generated text. Pangram Image is currently available as a research preview, with a wider release planned in the coming weeks.
Technically, Pangram’s detection system is a large machine-learning model trained on tens of millions of known human documents. For each human document the startup created a ‘‘synthetic mirror’’ — content that matches the topic, length, and tone but is written by a frontier large language model (LLM). According to the company, the model learns consistent stylistic choices made by AI and uses those patterns to classify content without relying on copy-paste metadata or hidden watermarks.
Degrees of AI assistance
Pangram treats detection as more than a binary question of whether a piece was entirely AI-written. The company aims to differentiate levels of AI assistance — for example, when a human author drafts a text and then asks AI to edit or polish it. Co-founder Max Spero has said he considers AI assistance acceptable if the author discloses the use of AI.
Availability and customers
Users can access Pangram through a $20-per-month web subscription or via a Chrome extension that labels posts in real time across X, LinkedIn, Substack, Reddit, and Medium. The extension also provides a feed health score showing the percentage breakdown of human versus AI content on the screen. Pangram offers an API for integration; customers mentioned include Substack (which integrated Pangram to indicate which authors use AI in their newsletters), Quora, schools and universities, publishers and agents, and recruiters.
Market context and regulation
Pangram enters a market with several competing detection tools, such as Winston AI, Originality.ai, Copyleaks, and GPTZero. Demand for detection tools is being driven by high-profile errors and institutional responses: for example, arXiv introduced an enforcement policy stating that submissions containing evidence authors did not review LLM output (such as hallucinated references or meta comments) can trigger a one-year submission ban.
Independent testing and limits
In limited testing reported by TechCrunch, Pangram’s model performed strongly but was not flawless. Pangram says roughly one in 10,000 human documents are incorrectly labeled as AI — about a 0.01% false-positive rate. In the tests, Pangram correctly flagged entirely AI-generated news articles produced by ChatGPT and Claude and resisted attempts to evade detection via edited output or prompting. However, there were instances where the detector flagged sentences as AI-written that had been completely rewritten by the human tester, and it sometimes marked human-written sentences as AI-assisted. When a full article written by the tester was supplied intact, Pangram gave it a 100% human score, while a polished version produced by AI received a roughly 13% AI-assisted score.
Pangram Image also showed promising results: the image detector analyzes pixel-level distributions to learn subtle statistical differences between real photos and AI-generated images and can detect AI-generated images even when they appear inside a real-world photo. In testing the model detected both photorealistic and stylistic AI images, and highlighted AI-origin regions with a heat map. There was at least one instance where the model misclassified a photo of an AI-generated image as human.
Company position
Max Spero has said Pangram does not want to promote a ‘‘witch hunt’’ against people using AI for writing, but he argues there should be mechanisms to push back against low-quality, indiscriminate AI output. Spero warned that AI-generated content could proliferate faster than human-produced content as compute resources (GPUs) scale, risking drowning out the human signal online.
Conclusion
With fresh funding and new models, Pangram seeks to expand its role in distinguishing human and AI content across news, social media, education, and publishing. Independent tests suggest the technology is useful but imperfect, highlighting both the potential and the limits of current AI-detection approaches.
