Pangram Labs raises $9M to launch more accurate AI detection for text and images
SiliconANGLE Kyt Dotson ● Covered by 4 sources
Pangram Labs just landed $9M to sharpen its AI-text detector and push into images. But critics still say AI detectors are shaky at best, especially for schools.
Based on reporting by SiliconANGLE, Kyt Dotson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Pangram Labs has pulled in $9 million in fresh funding, led by Menlo Ventures with help from Haystack, ScOp Venture Capital, Script Capital and Cadenza. That brings the company's total raised to nearly $13 million, following a $2.7 million round back in June 2025. The money is going toward two things: making the company's text detector even more precise, and pushing into a new frontier — catching AI-generated images.
Pangram's pitch has always centered on accuracy. The company says its classifier model, trained by contrasting pre-2021 human writing against AI-generated text, can spot machine-written content by essentially asking what a passage "sounds like." Its newest release, Pangram 4, reportedly drops the false positive rate to 0.0041 percent — roughly one mistake per 24,000 documents, according to the company's own internal benchmarks. Pangram also claims fewer false negatives and better resistance to so-called humanizer tools that try to disguise AI text as human work.
The image detector, still in research preview, is where things get more interesting. It's built to flag pictures from GPT Image, Gemini Nano Banana, Midjourney, FLUX and Grok Imagine, along with video from Kling AI, Seedance, Veo and Wan. Pangram argues this is necessary because invisible watermarking schemes like Google's SynthID only work when the originating model actually embeds them — leaving a lot of AI imagery with no built-in tell. And the humans trying to catch fakes on their own aren't doing great: a report from Let's Enhance found overall detection hovering around 63.7 percent, sinking closer to 29 percent for harder cases like FLUX-generated images, with some tests landing near a coin flip.
None of this erases the skepticism baked into AI detection as a category. MIT Sloan's teaching technology group has called the tools far from foolproof, warning of error rates that can produce false accusations. The Mozilla Foundation has raised similar concerns, including bias against non-native English speakers. That skepticism has consequences: the University of Waterloo dropped Turnitin's detector in September, and MIT's broader stance is that these tools simply don't work reliably enough to anchor policy — schools should build expectations around AI use instead of leaning on detection scores as verdicts.
Pangram's cofounder and CEO, Max Spero, frames the company's mission as making authorship legible again for publishers, teachers and journalists navigating a flood of undisguised AI content. That's a real problem. But the gap between a vendor's internal benchmark and how a tool performs across messy, real-world writing and imagery is exactly where past detection tools have stumbled — and it's the gap Pangram will have to close in public, not in a white paper.
My take — AI-written commentary, not fact-checked reporting
A false positive rate this small sounds impressive until it's tested against actual classrooms and newsrooms rather than a lab's own dataset — internal benchmarks are not the same as independent scrutiny. Institutions like Waterloo and MIT didn't walk away from AI detectors because the marketing was unconvincing; they walked away because the tools kept getting it wrong on real people. Betting a product's whole reputation on a single accuracy number is a fragile strategy in a field where every prior confident claim has eventually cracked.
Read more about this at: SiliconANGLE