# Pangram raised $9 million to expand AI detection into images > Source: > Published: 2026-08-11 01:50:50+00:00 On July 29, Stanford classmates [Max Spero (@max_spero_)](https://x.com/max_spero_?ref=runtimewire) and [Bradley Emi (@bradley_emi)](https://x.com/bradley_emi?ref=runtimewire) expanded Pangram from text classification into image detection. The founders released a new text model, Pangram 4, and opened their image detector for public testing as they pushed [Pangram](https://www.pangram.com/about-us?ref=runtimewire), the New York-based AI detection startup, into visual media. Spero and Emi met during their freshman year in a Stanford dorm, then followed separate paths through applied machine learning before starting Pangram. Spero, Pangram's CEO, led active-learning work at autonomous-vehicle developer Nuro after roles at Google, Two Sigma and Yelp. Emi, the CTO, worked on Tesla Autopilot's computer-vision team and later led deep-learning research at generative drug-discovery developer Absci. Both earned master's degrees in artificial intelligence from Stanford, where Emi also conducted research with the Stanford Vision Lab, according to [Pangram's company biography](https://www.pangram.com/about-us?ref=runtimewire). Their timing followed ChatGPT's release and the rapid spread of machine-written articles, reviews, schoolwork and social posts. [Spero told TechCrunch, as republished by Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/ai-content-floods-internet-pangram-110000725.html?ref=runtimewire), that knowing the origin of a piece of writing changes how a reader evaluates it, particularly when generated text may contain hallucinations or reflect little original research. Pangram now sells browser tools, document integrations and an API to publishers, schools, recruiters and content platforms. Substack and Quora are among the platforms identified as users of its technology. Pangram's [pricing page](https://www.pangram.com/pricing?ref=runtimewire) lists a free web tier alongside individual, professional, developer and enterprise plans. ### Financing the expansion [Menlo Ventures](https://menlovc.com/perspective/investing-in-pangram-to-stop-ai-slop-on-the-internet/?ref=runtimewire) led the $9 million round, with Haystack, ScOp Venture Capital, Script Capital and Cadenza participating, according to [TechCrunch's report republished by Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/ai-content-floods-internet-pangram-110000725.html?ref=runtimewire). In [Menlo's investment post](https://menlovc.com/perspective/investing-in-pangram-to-stop-ai-slop-on-the-internet/?ref=runtimewire), partner Deedy Das wrote that he contacted Spero on X after seeing Pangram identify accounts posting AI-generated material and that Menlo led the round through its Anthology Fund. Das also wrote that Spero and Emi had refused requests to build a humanizer that would help users evade detection. The round follows $3.98 million in earlier pre-seed and seed financing. Haystack led the pre-seed, while ScOp led the seed with participation from Script Capital and Cadenza. Adding the latest round brings Pangram's reported financing to roughly $13 million. Pangram has not published a valuation for the new investment. Its prior funding was detailed in a [June 2025 financing announcement](https://www.businesswire.com/news/home/20250623805035/en/Pangram-Closes-%244-Million-in-Seed-Funding-for-AI-Detection-Technology?ref=runtimewire). The GPTZero acquisition gives Pangram a recent example of consolidation in the category. Superhuman [acquired GPTZero](https://techcrunch.com/2026/06/23/superhuman-acquires-ai-detection-startup-gptzero/?ref=runtimewire) on June 23 after the startup had raised $13.5 million. Terms were not disclosed. Pangram was not part of that transaction and remains a separate company backed by Menlo Ventures and its other investors. ### How Pangram 4 works Pangram's text-classification method starts with tens of millions of known human-written documents, according to [Spero's explanation in TechCrunch's report republished by Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/ai-content-floods-internet-pangram-110000725.html?ref=runtimewire). Pangram creates what Spero calls a "synthetic mirror" for each document by asking a frontier language model to reproduce its subject, length and tone. The classifier then learns stylistic patterns that distinguish the human documents from the generated versions. The technique analyzes the text itself without relying on watermarks or metadata supplied by a model provider. Pangram also attempts to identify writing that began with a human draft and was subsequently edited by an AI system. [Pangram says Pangram 4 is more than 99% accurate at detecting AI-assisted and mixed human-AI writing and is better at identifying text processed by humanizer tools](https://www.pangram.com/blog/introducing-pangram-4?ref=runtimewire). Pangram says the model is over six times larger than Pangram 3 and identified AI involvement in 98.83% of output processed by 13 humanizer tools. These are company-reported results, rather than independent proof of authorship. The separately published [Pangram 4 technical report](https://arxiv.org/abs/2607.27183?ref=runtimewire) reports an AUROC of 0.9916, a 0.0041% false-positive rate and a 0.3396% false-negative rate on the authors' test data. It also reports improved performance on mixed human-AI text, fine-grained edits, out-of-distribution material and adversarial attacks relative to Pangram 3. The paper was written by Pangram researchers, including Spero and Emi, and its headline metrics depend on the composition of the test sets and evaluation conditions. ### Testing image detection as a second product Pangram Image applies the founders' authorship thesis to visual media. Pangram says the model can detect wholly generated pictures and synthetic material embedded inside real-world photographs. It also produces heat maps intended to locate the portion of an image that appears generated. [Pangram described the image detector as being in research preview in its July 29 release](https://www.businesswire.com/news/home/20260729222515/en/Pangram-Launches-New-Worlds-Best-AI-Detector-Expands-to-Images-Backed-by-%249M-From-Menlo-Ventures?ref=runtimewire), which said the public could test the model and provide feedback. The supplied sources do not establish an independently validated accuracy figure. Publishers and platforms increasingly handle combinations of human writing, generated copy, edited photographs and wholly synthetic visuals. Pangram still has to show that its image model can maintain its claimed performance across customers, image generators and editing workflows. [TechCrunch's limited test, republished by Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/ai-content-floods-internet-pangram-110000725.html?ref=runtimewire), found that the detector identified several generated images and synthetic content inside a photograph, while misclassifying one photographed AI-generated image as human content. The same validation problem applies to text. False classifications can carry serious consequences when a detector is used to question a student's work, a journalist's reporting or an author's manuscript. [TechCrunch's limited text testing, republished by Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/ai-content-floods-internet-pangram-110000725.html?ref=runtimewire), found that Pangram correctly scored one fully human article while flagging some human-rewritten sentences as AI-assisted. [The Atlantic](https://www.theatlantic.com/technology/2026/05/pangram-ai-detection-accuracy/687381/?ref=runtimewire) has separately raised questions about Pangram's reported error rates. Neither publication's testing amounts to a broad independent benchmark. Pangram 4 produces a statistical judgment rather than proof of authorship. That distinction becomes more consequential when publishers, schools and platforms use a score to decide whether work should be trusted, reviewed or rejected. The financing gives Spero and Emi room to improve the text classifier while testing image detection as a second product, but customers will determine how much authority those probabilistic scores receive.