Pangram raises $9 million to catch AI-written text and images before they fool you
A New York startup says its detector is over 99% accurate at spotting AI-generated writing. Substack already uses it. Here's what that means for everyday readers.
Key points
- Pangram, a New York AI-detection startup, raised $9 million in funding led by Menlo Ventures in 2025.
- Its new text model, Pangram 4, claims over 99% accuracy at spotting AI-written or AI-assisted content.
- A new image-detection model, Pangram Image, is currently in research preview and will launch more widely in coming weeks.
- Substack has already built Pangram's technology into its platform so readers can see which newsletter authors use AI.
- A personal subscription costs $20 per month, and a free Chrome extension labels posts on X, LinkedIn and other sites in real time.
There's a lot of text on the internet that nobody actually wrote. AI tools have made it trivially easy to flood blogs and social feeds with content that looks human but was generated in seconds. Pangram, a two-year-old New York startup, thinks that's a problem worth solving, and investors just handed it $9 million to try.
The round was led by Menlo Ventures, with Haystack, ScOp, Script Capital, and Cadenza also taking part, according to TechCrunch AI.
What does Pangram actually do?
Pangram builds AI detectors: software that reads a piece of text or examines an image and decides whether a human or a machine made it. The company launched two products alongside the funding announcement: Pangram 4, its latest text-detection model, and Pangram Image, which looks for telltale signs that an AI generated a picture rather than a camera capturing a real scene.
To train the text detector, Pangram fed the system tens of millions of known human documents, then built a "synthetic mirror" for each one: a version covering the same subject and tone, written by a frontier large language model (the kind of AI that powers ChatGPT or Claude). It's learned the stylistic choices AI makes consistently, without relying on hidden watermarks or copy-paste metadata.
Pangram Image works differently. Rather than checking for watermarks, which mostly catch whichever company's own output OpenAI or Google DeepMind left behind, it analyses pixel-level statistics: fine-grained mathematical patterns that differ between a real photograph and an AI-generated one. That approach is meant to catch AI images regardless of which tool produced them.
Should you trust the accuracy claims?
Pangram says roughly one in 10,000 human documents gets incorrectly flagged as AI. Reassuring, but no detector is perfect.
TechCrunch AI's reviewer tested the tool and found it correctly caught entirely AI-written articles from ChatGPT and Claude, and wasn't easily fooled by light human editing of AI text. When the reviewer asked those models to polish one of their own human-written articles, Pangram scored it at 13% AI-assisted, which the reviewer considered roughly accurate. False positives did appear on individual sentences, particularly in dry, factual writing that can sound AI-like even when it isn't.
For image detection, the tool correctly identified AI-generated imagery whether photorealistic or cartoonish, and even spotted an AI image placed inside a real photograph. One misclassification did occur during testing.
What does it cost, and who already uses it?
| Plan | Price | What you get |
|---|---|---|
| Web subscription | $20 per month | Full text and image detection on uploaded content |
| Chrome extension | Free | Real-time labels on posts across X, LinkedIn, Substack, Reddit and Medium |
| API access | Contact Pangram | Integration into your own platform or workflow |
Substack has already integrated Pangram's API so readers can see which authors write their newsletters using AI. We covered the partnership on 22 July in "Substack Now Flags AI-Written Newsletters Before You Read Them". Quora, schools, universities, publishers and recruiters are among other API customers, according to co-founder Max Spero.
Why does this matter for ordinary readers?
The risks of unchecked AI content are real. A Canadian politician accidentally read an AI prompt aloud during a parliamentary speech (we reported the full story at "Canadian politician accidentally reads out his AI chatbot's instructions during a floor speech"). Lawyers have cited fake legal cases that ChatGPT invented, leading to sanctions. The open-access research archive arXiv now bans authors for up to a year if submissions show clear signs that AI output wasn't reviewed before submission.
Pangram isn't the only company chasing this problem. Winston AI, Originality.ai, Copyleaks and GPTZero are all building similar tools. The funding and the Substack partnership give Pangram a meaningful foothold, though that foothold only matters if the accuracy holds up outside a reviewer's controlled tests.
One privacy note: when you paste text or upload an image to any cloud-based detection tool, that content leaves your device. Check Pangram's privacy policy before submitting anything sensitive.
Common questions
Does the Chrome extension cost anything?
The Chrome extension is free. It automatically labels posts on X, LinkedIn, Substack, Reddit and Medium as human or AI in real time, without you needing to paste anything into a separate tool.
Can it tell if a human wrote something and then used AI to tidy it up?
Yes, that's one of Pangram's specific goals. The system tries to flag different levels of AI involvement, from fully AI-generated text to content that was mostly human-written but lightly polished by a chatbot.
Will it flag my writing if I'm just a clear, plain writer?
Possibly, in isolated sentences. The reviewer found that dry, factual writing occasionally triggered false positives. Running a full document through the tool is more reliable than checking individual paragraphs.
