Pangram Labs raises $9M to launch more accurate AI detection for text and images

UPDATED 13:20 EDT / JULY 29 2026 AI Pangram Labs raises $9M to launch more accurate AI detection for text and images by Kyt Dotson Pangram Labs Inc., an artificial intelligence research lab that develops AI detection software, today announced it raised $9 million, led by Menlo Ventures, to improve the accuracy of its core text detection and expand into other media, starting with images. Haystack, ScOp Venture Capital, Script Capital and Cadenza also participated in the investment round. The funding brings the total raised by the company to almost $13 million after raising $2.7 million in June 2025. Pangram is best known for its AI detection platform that the company claims is capable of industry-leading 1 in 10,000 false positives. “Fully AI-generated content is everywhere now, and much of it is undisclosed,” said cofounder and chief executive Max Spero. “Pangram exists to make authorship legible for publishers, teachers, journalists and anyone who relies on the truthful and authentic communication of information.” The company uses what it calls a “classifier model,” a type of neural network that can estimate whether a portion of text is AI-generated or human-written. According to Pangram’s page on how the product works,  says that the system functions by attempting to determine what a passage “sounds like” an LLM or a human. Training the classifier involved pulling known-human text drawn from before 2021 and pairing it with AI-generated text so that the classifier could readily distinguish the two. The company acknowledged that this works well now, but language style and use drift over time, and the training model will have to adjust with it. The company also announced Pangram 4, the company’s most powerful AI detector to date, and introduced Pangram Image detection in research preview. The company said in internal benchmarks, Pangram 4 has a false positive rate of 0.0041%, or around once for every 24,000 documents. The company added that the new model also greatly reduces false negatives, when it fails to detect AI, and it is robust against humanizers, which attempt to make AI text look human-written. The company’s Image model for detecting AI-generated images can catch images created by image providers including OpenAI Group PBC’s GPT Image, Google LLC’s Gemini Nano Banana, Midjourney Inc., FLUX, and Grok Imagine, and some AI video providers, including Kling AI Pte. Ltd., Seedance, Google’s Veo and Wan. The company said that a new breed of AI image detectors is needed because deepfakes and AI-generated images passed as truth are becoming more prevalent. Although other companies have begun to embed invisible watermarks and other markers in their content, for example, Google’s SynthID, these only work on frontier models that embed them. Even as AI images proliferate, humans are getting worse at detecting them. According to a report from Let’s Enhance, a blog focused on AI creative tools, overall detection rates hover around 63.7%, but for high-capability image generators such as FLUX, rates drop to near 29%. Research showed that distinguishing AI from natural is falling to close to 50% on average, essentially a coin toss. AI detectors and the reliability gap Pangram’s detection accuracy and false positive rate claims come from what appears to be primarily internal benchmarks, a technical white paper and a small number of favorable third-party studies. The company wants to set itself apart from other AI detectors because accuracy is meaningful. A key point to examine for AI detectors is that they are by and large unreliable. MIT Sloan Teaching and Learning Technologies pointed out that the technology is “far from foolproof” and