Pangram, a New York AI detection startup, raised $9 million led by Menlo Ventures to scale a system it says identifies AI-assisted writing with over 99% accuracy. The round, which also drew Haystack, ScOp, Script Capital, and Cadenza, lands alongside the launch of Pangram 4, the company's next-generation text detector, and Pangram Image, a research-preview image detector due for wider release in the coming weeks. Pangram says roughly one in 10,000 human documents are incorrectly flagged as AI.
The startup was founded about two years ago by Stanford AI and machine learning graduates Max Spero and Bradley Emi, shortly after ChatGPT opened the tap on AI-generated writing across the web. Pangram sells access through a $20-per-month subscription and a Chrome extension that labels posts in real time on X, LinkedIn, Substack, Reddit, and Medium, complete with a feed health score showing the human-versus-AI split on screen. It also offers an API.
Distribution is expanding. Substack integrated Pangram to show readers which of their favorite writers use AI to draft newsletters. API customers include Quora, schools and universities, publishers, agents, and recruiters, according to Spero. The company is targeting institutions that need to grade, moderate, or verify content at scale, not consumers running one-off checks.
Key facts
- 01Pangram raised $9M led by Menlo Ventures, with Haystack, ScOp, Script Capital, and Cadenza participating.
- 02Pangram 4 claims over 99% accuracy detecting AI-assisted writing, with roughly 1 in 10,000 human documents misflagged.
- 03The startup sells a $20-per-month subscription and a Chrome extension that labels posts on X, LinkedIn, Substack, Reddit, and Medium.
- 04Substack integrated Pangram to show which authors write newsletters with AI; Quora, schools, and recruiters are API customers.
- 05arXiv adopted a policy this year triggering one-year submission bans for papers showing unreviewed LLM output.
The detector is a machine learning model trained on tens of millions of known human documents. For each human document, Pangram generated a "synthetic mirror" — matching topic, length, and tone but written by a frontier LLM — and trained the model to spot the stylistic tells. The approach does not rely on watermarks or metadata, meaning it works on output from any model, not just those that ship provenance signals.
“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.”— Max Spero, Pangram co-founder
That last point matters because the watermark strategy pushed by OpenAI and Google DeepMind only catches their own output. Pangram Image, by contrast, works on pixel-level statistical differences and claims to detect AI-generated imagery regardless of source model, including an AI image embedded inside a real-world photograph. Competitors chasing the same demand include Winston AI, Originality.ai, Copyleaks, and GPTZero.
Demand is being pulled forward by institutions writing new rules. The open-access archive arXiv introduced an enforcement policy this year under which submissions showing evidence of unreviewed LLM output — hallucinated citations, or telltale meta phrases like "Would you like me to make any changes?" — can trigger a one-year submission ban. Courts have sanctioned lawyers filing briefs with fabricated ChatGPT citations. A Canadian politician was ridiculed for reading an AI prompt aloud during a speech.
TechCrunch's own testing found the model impressive but not flawless. It easily flagged full articles written by ChatGPT and Claude, resisted evasion prompts, and correctly scored an original human-written article at 100% human. When the same article was polished by ChatGPT and Claude, Pangram returned a 13% AI-assisted score, catching some edited sentences while missing others and occasionally flagging untouched human sentences as AI.
Spero frames the product as a defense against a signal-to-noise collapse, not a witch hunt. He says AI-assisted writing is acceptable as long as writers disclose it, and that the goal is to give readers the information to calibrate trust — whether a piece was researched by a journalist or generated by a model prone to hallucination.
The counterweight is that no detector is perfect, and the arms race cuts both ways. AI humanizer tools are explicitly designed to defeat detectors, and every false positive against a human writer carries real reputational cost. Pangram's own 1-in-10,000 error rate sounds low until it is applied to a university grading tens of thousands of essays a semester, or a recruiter screening a million resumes a year. Competing detectors have been publicly criticized for misfiring on non-native English writers.
For the AI market, Pangram's raise is a bet that provenance becomes a paid layer of the stack rather than a feature model providers ship themselves. Watermarking has stalled because it only works when every lab cooperates and no user runs open-weights models — neither condition holds. That leaves a durable opening for third-party detectors selling to platforms, publishers, schools, and courts who need to answer the question the labs cannot answer for them: was this written by a person, and if not, by which model, edited how much.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
