Pangram's $9M bet: AI detection without watermarks, and Substack and Quora are already buying it

Pangram's $9M bet: AI detection without watermarks, and Substack and Quora are already buying it

Substack has integrated it. Quora is an API customer. Universities and recruiters are paying. The product is a machine learning model that labels text as human, AI-assisted, or AI-generated, without needing AI labs to embed watermarks or cooperate in any way. The company behind it, Pangram, just closed $9 million to scale.

Founded by Stanford AI and machine learning grads Max Spero and Bradley Emi, Pangram raised $9 million led by Menlo Ventures, with Haystack, ScOp, Script Capital, and Cadenza participating 1. The round coincides with two product launches: Pangram 4, the company's latest text detector, and Pangram Image, an AI image detector currently available only in research preview 1.

For two years, the dominant approach to AI content detection assumed cooperation from the labs producing the content. Google's SynthID, the most visible watermarking system, embeds imperceptible markers into output from Google's own generative products 2. But SynthID's coverage is structurally narrow. It reliably detects Google's output and struggles with content from other providers like OpenAI or Meta, and Google itself acknowledges that detection confidence drops sharply when AI-generated text is thoroughly rewritten 3. Watermarking depends on every lab marking its own work. That coordination has not materialized.

Pangram does not ask labs for help. The company built its training dataset by pairing tens of millions of known human documents with synthetic copies of each one, generated by a frontier LLM to match the same topic, length, and tone. The model then identifies patterns in how AI consistently writes differently from humans, rather than searching for embedded markers 1. The advantage: it works on text from any source, not just one lab's products. Pangram 4 claims over 99% accuracy and can distinguish between fully AI-generated text and text where a human wrote it but AI edited or polished the prose 1.

Independent benchmarks partially back the accuracy claims. Pangram tied for first at 99.3% on the COLING 2025 academic benchmark 4. Spero cited a 1-in-10,000 false-positive rate in his interview with TechCrunch 1. The TechCrunch reporter's own testing confirmed the model flagged AI-generated articles from ChatGPT and Claude and resisted manual edits, but also incorrectly labeled some fully human-written sentences as AI-assisted 1.

That failure mode is where the story gets complicated. A 1-in-10,000 false-positive rate sounds negligible until a platform has already flagged a student's dissertation and the appeals process is thin.

Pricing is accessible at the individual level. A web subscription costs $20 per month, with a free tier of four checks per day 4. A Chrome extension auto-labels posts in real time on X, LinkedIn, Substack, Reddit, and Medium, with a feed health score showing the percentage breakdown of human versus AI content on screen 1. Enterprise API pricing for customers like Substack and Quora has not been publicly disclosed. Pangram holds SOC 2 Type II compliance 4.

The customer roster is the signal. Substack has integrated Pangram to show readers which authors write their newsletters with AI 1. Quora, schools, universities, publishers, agents, and recruiters are API customers, according to Spero 1. These are paid contracts at platforms that have decided AI content is an operational problem with a budget attached. The preprint server arXiv moved in the same direction this year, with a policy that can trigger a one-year submission ban for authors who submit unreviewed LLM output 1.

Pangram is not alone in betting this market is real. Winston AI, Originality.ai, Copyleaks, and GPTZero are building competing detectors 1. Capital is flowing to the category.

The question that $9 million does not answer is durability. Model-based detection does not depend on lab cooperation, which makes it broadly applicable today. But it is also a moving target. New LLMs produce text with different characteristics. AI humanizer tools that rewrite generated text to evade detection are already a product category, and Pangram 4 claims to catch their output 1. Each cycle of evasion and detection will compress the accuracy margin that makes the product worth buying. Platforms are betting real money on a tool whose effectiveness depends on winning an arms race that has barely started.