The one AI detector people actually trust

When Max Spiro, CEO of Pangram, describes his job, he calls himself a "slop janitor." It is a deliberately unglamorous title for a company that has quietly become the most trusted name in AI text detection—a field that, until recently, was widely considered unreliable. The shift happened gradually. For two years, Pangram was just another startup publishing technical reports that nobody read. Then independent researchers started benchmarking it and publishing results. Suddenly, the conversation changed from "AI detectors don't work" to "well, Pangram says it's AI."

How Pangram actually works

The key innovation is not a clever algorithm but a training strategy called active learning. Most early AI detectors relied on a single metric called perplexity—a measure of how surprising a piece of text is to a language model. The logic was simple: AI models are trained to produce unsurprising, low-perplexity sentences, while humans write in more surprising ways. This approach broke down in predictable ways. The Declaration of Independence gets flagged as AI because the model has memorized it. English language learners get flagged because they write in simple, predictable prose. And as Spiro put it, "perplexity is not really a metric that can be improved upon. It's like measuring the density of a liquid."