Max Spero calls himself the internet’s slop janitor.
He is the CEO of Pangram, an artificial intelligence (AI) detection company that has become a leading tool for deciphering whether an article, a social media post or another piece of text was written by a human or a chatbot. In other words, is it AI slop or not?
As generative AI continues to permeate more and more forms of communication, New York-based Pangram is one of a new group of startups vying to be the best at distinguishing AI from human-generated content. The company says its AI text-detection model can accurately identify 9,999 out of 10,000 times whether text was AI-generated. And it has recently signed up some high-profile online publishers as customers.
AI writing and AI-detection tools raise a host of big, thorny questions: How should society treat AI-generated text? And how should AI-detection tools like Pangram be used given that they can – even if rarely – be wrong?
Pangram announced a US$9mil (RM36.7mil) fundraising round led by Menlo Ventures with participation from ScOp Venture Capital, Haystack Ventures and others. The fast-growing company is also announcing the release of its latest AI detection model – Pangram 4.0 – and its first-ever AI-image detection model.
Spero and his co-founder, Bradley Emi, both 30, met as undergraduates at Stanford University, studying in the university’s artificial intelligence lab and playing poker late into the night. Upon graduation, they followed well-trodden paths for Stanford computer science engineers: Spero went to Google and Emi to Tesla, where he worked under Andrej Karpathy, the prominent AI researcher and OpenAI co-founder.
After the release of ChatGPT in 2022, the two friends felt called to build something “for the world that this technology will bring forth,” Spero said. An inevitable feature of the AI future, they felt, would be scores of AI-generated content flooding the Internet and other channels of communication.
Spero and Emi founded Pangram in 2023, and their prediction turned out to be prescient. Today, roughly 50% of news articles and blogs on the Internet are AI-generated, according to a recent analysis by the digital marketing firm Graphite.
Honing accuracy
Pangram’s interface resembles a chatbot. The website prompts users to insert a piece of text, and Pangram spits out a verdict – AI, human or AI-assisted. Behind that simple interface is a powerful model. Pangram trained it by feeding its AI system millions of examples of text to analyse, and asking which were written by humans and which by AI. Over time, the model honed its accuracy.
Pangram has proved remarkably effective at detecting whether a full piece of text was written by AI or a human. A 2025 University of Chicago study found that Pangram maintained near-perfect accuracy in assessing texts that were roughly 500 words or more, outcompeting Originality.ai and GPTZero.
The company’s tool is not as deft at detecting AI in shorter pieces of text where the model has less context. It’s also less capable of identifying text that was partly AI-generated, though the model released today is aimed at improving those capabilities.
The University of Chicago study contradicted a belief that had been gaining traction in tech circles and beyond about AI-detection tools: that they don’t work. The belief solidified after OpenAI shut down its own AI-detection tool in 2023 because of its poor accuracy.
But as word of Pangram’s effectiveness spread, its user base began to grow.
A majority of Pangram’s revenue comes from individual subscriptions, which cost US$20 (RM81.75) per month. The company has also signed contracts with a number of universities and publishers looking to crack down on AI writing.
Publishing platforms, interested in weeding out AI posts or bots, are also using Pangram. Quora, the Q&A site, was one early customer. Substack started a partnership with Pangram to allow readers to opt out of seeing AI-generated content.
That customer traction has helped Pangram’s annual revenue rate grow to 35 times what it was a year ago. The company declined to share its exact revenue but said its number of monthly users grew to 120,000 in June, up from 2,700 a year earlier.
‘Cognitive security’
The addition of its AI-image-detection tool – which Pangram says is 99.5% accurate in detecting images fully generated by AI – will augment Pangram’s goal of providing users with “cognitive security.” Pangram also plans to release an AI-video-detection tool this year.
Pangram’s lead investor believes Pangram is meeting a broad and growing demand. “I see AI detection as a kind of universal need that everyone has,” said Deedy Das, a partner at Menlo Ventures.
There’s one big caveat, though: Pangram doesn’t always get it right. And people might not apply it correctly, either. That creates the potential for accusations about AI-generated writing that might be wrong.
Pangram can struggle to accurately assess text that was partly AI-generated, as Substack author Freddie deBoer recently pointed out. DeBoer found that when he appended a few sentences of AI writing to a paragraph otherwise written by him, Pangram reported the entire thing as AI-generated. In a post on social media, Spero said the new model was able to more accurately identify text partly generated by AI, including in the example shared by deBoer.
Then there is the question of how to treat text that Pangram labels as partly written with AI, perhaps through editing or rewriting. Spero said he thought it was “completely OK” for people to use AI as an assistant, but noted that he personally chose not to.
How schools, publishers and individuals want to treat that messy middle remains to be worked out. – ©2026 The New York Times Company
This article originally appeared in The New York Times.
