Via phrasly.ai
The Brooklyn startup claims a 99.98% detection rate for AI-generated text, but academic scrutiny suggests the reality is more complicated.
A Brooklyn-based startup called Pangram Labs says it can tell whether a human or a machine wrote something with near-perfect precision. The company claims its detection tool catches AI-generated text at a rate exceeding 99.98%, with a false positive rate of just 1 in 10,000 on human-written content.
Founded in 2023 by former Stanford researchers Max Spero and Bradley Emi, Pangram has positioned itself at the center of one of the most consequential questions in digital media: can you reliably prove that a piece of content was made by a person?
What Pangram actually does #
You feed it a piece of writing, anywhere from 75 words to 75,000 characters, and it spits back a verdict: Human-Written, Lightly AI-Assisted, Moderately AI-Assisted, or Fully AI-Generated. It also assigns a numerical AI assistance score for more granular analysis.
The tool supports more than 20 languages and integrates with learning management systems like Canvas and Google Classroom.
Independent testing has been encouraging. Evaluations by the University of Chicago Booth School of Business and the University of Maryland found accuracy rates between 99.8% and 100% on content from major language models, including ChatGPT and Claude. On the RAID benchmark, a widely used test suite for detection tools, Pangram hit 99.44% accuracy with a 0.05% false positive rate.
The company released Pangram 3.1 on January 16, 2026, adding capabilities for mixed-text scenarios where human writing and AI output are blended together. That update also brought multi-language support and model-specific detection tailored to individual AI systems.
On the funding side, Pangram closed a seed round of approximately $4 million in June 2025, led by ScOp with participation from Script Capital, Cadenza, and Haystack in the pre-seed.
The accuracy debate is where things get interesting #
Academic discussions as recent as March 2026 have highlighted a gap between Pangram’s marketed false positive rate and what some third-party evaluations actually found. While the company advertises the 1-in-10,000 figure, certain external studies recorded mean false positive rates ranging from 0.48% to 2%.
Pangram’s independent test results across academic, news, and creative writing datasets showed false positive rates between 0% and 0.17%. The tension lies in whether those controlled benchmarks translate cleanly to the messy, diverse real world.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our