AI-generated writing has distinct patterns—and Pangram Labs says it can detect them. The company says its AI-powered models can discern with over 99.5% accuracy whether a piece of writing was produced entirely by humans or AIs, according to its internal tests. Those tests also show its tools are able to differentiate between human- and AI-written segments within a given text, estimating the degree of AI involvement. As co-founder and CEO Max Spero explains, because most AI writing comes from just a handful of model families—ChatGPT, Claude, Gemini, and the like—identifying it is easier than many believed just a few years ago. “Ultimately, what we’re doing is an authorship detection problem,” he says.
Spero says his company, founded in late 2023, was “close to shutting down” before Pangram spiked in popularity last fall as AI slop proliferated across platforms. By July of this year, it had a $9 million fundraising round, a tool for detecting AI-generated images, and a partnership with Substack to integrate its technology into the blogging platform.
Some AI text detectors also offer “humanizers”—tools intended to reshape AI writing so it can slip past detection. Spero says Pangram Labs will never release one. It’s not just a matter of principle—Pangram’s long-term aspiration is for every layer of online communication to use its tools to verify authorship, a function for which he expects there to be a far larger market in the age of slop and AI impersonation.
The deeper challenge of tracing the extent to which ideas themselves are influenced by AI may be intractable. And frontier AI models are getting better all the time, while some people may be writing more like AIs. But Spero believes Pangram’s is deterring people from passing off AI writing as their own. “The LLMs aren’t static,” he says, “but neither is Pangram.”