Pangram's Max Spero on AI Detection

Pangram Labs closed a $9 million Series A in July 2026 — led by Menlo Ventures, with participation from Haystack, ScOp, Script Capital, and Cadenza — and on the same day announced both a next-generation text detection model and the company's first AI image detector. The Brooklyn-based startup, co-founded by Stanford AI graduates Max Spero and Bradley Emi, has spent the past two years pushing a specific argument: asking whether content is "real or fake" is the wrong question, and the harder, more useful one is where a piece of writing falls on the spectrum between fully human and fully AI.

Background You Need

When ChatGPT launched in late 2022, it didn't just give people a new writing tool — it handed organizations a verification problem they had no infrastructure to solve. Educators, publishers, and hiring managers suddenly needed ways to tell whether the text in front of them had been written by a person. A cluster of AI detection tools rushed to fill that gap, but most treated the problem as binary: a piece of text was either AI-generated or it wasn't. That framing turned out to be both technically imprecise and practically useless in a world where most professional writers now use AI in some form.

Spero and Emi, both Stanford AI and machine learning graduates who later worked at Google and Tesla, founded what would become Pangram in 2023 — initially under the name Checkfor.ai, before rebranding in 2024. Their origin story, recounted in a Pulse 2.0 interview, traces back to graduate work on robotic feedback loops: collect operational data, use it to improve performance, iterate. That same philosophy, Spero says, informed how they train the detection model — using continuous feedback to keep pace with new LLM outputs rather than training once and shipping. The company raised $4 million in seed funding in 2025, led by ScOp Venture Capital, which gave it enough runway to reach meaningful scale before the Series A.

By June 2026, Pangram had grown its monthly active user base from roughly 2,700 to 120,000 in a single year — a figure documented in the company's Wikipedia entry — with annual revenue climbing by a factor of 35 over the same period. Those aren't metrics that suggest a niche academic tool; they suggest demand for AI-content verification has moved from early adopters into something approaching mainstream.

What's New

The $9 million round closes at the same moment Pangram is expanding in two directions at once. On the text side, the company launched Pangram 4, its fourth-generation detection model, which it claims achieves over 99% accuracy at identifying both fully AI-generated text and the harder mixed-authorship case — content that started with a human draft and was then substantially rewritten by an AI, or vice versa. TechCrunch's coverage of the fundraise notes that the new model is specifically designed to catch "AI humanizer" programs — tools that run AI output through a second model to make it read more naturally. That's a meaningful technical distinction: Pangram is now explicitly in an arms race with software built to defeat it.

On the image side, Pangram Image launched simultaneously, though it's currently available only as a research preview. A broader release is expected in the weeks following the July 2026 announcement. The expansion beyond text reflects where Spero sees the problem heading — AI-generated images are flooding product listings, social media, and journalism, and the same trust questions that apply to written content apply equally there.

The practical deployment drawing the most attention is Pangram's integration with Substack, the newsletter platform. Rather than blocking AI-assisted content or requiring authors to self-disclose, the integration surfaces a disclosure badge that tells readers what level of AI involvement, if any, was present in a given piece. Multiple reports confirm the partnership is live, not merely announced. Our read is that this is the most significant part of the Pangram story right now: it's the first production example of a major content platform embedding probabilistic AI disclosure at the platform level. The Substack model doesn't punish writers — it informs readers. That design philosophy choice is one other platforms will likely have to respond to.

Spero has been consistent across interviews in pointing to high-stakes, real-world use cases beyond media: résumés submitted through AI writing tools, product reviews generated at scale to manipulate search rankings, insurance claims drafted with AI assistance to obscure inconsistencies. In the TechCrunch interview, he frames Pangram not as a policing tool but as a "trust layer for the internet" — infrastructure that lets platforms and readers make informed decisions. The distinction matters in how a product gets designed, and in how it gets received.

The Pushback

Wikipedia's entry on Pangram is notably candid: even relative to other AI detection tools, Pangram is described as more accurate, but it still "struggles on certain types of text." More pointedly, the entry records that the company has been criticized for contributing to what critics call "witch hunts" — situations where authors have faced accusations of AI use based on detector output that was wrong, or technically right but lacking context. That's a real-world consequence of marketing any probabilistic system as a definitive verdict. A 99% accuracy claim sounds strong until you apply it to the volume of content produced daily and start counting false positives at scale.

There is also the structural problem that detection tools sit permanently on the back foot. Each new generation of LLMs and AI humanizers changes the statistical signatures detectors are trained to catch. Pangram's feedback-loop training philosophy is designed to address this, but it requires continuous model updates — which means there will always be a window between a new generation of AI writing tools and a corresponding detector response. Separately, some competitors are pursuing provenance-based approaches — cryptographic signing of content at the point of creation — that sidestep the accuracy problem entirely by shifting verification upstream rather than after the fact. Whether post-hoc statistical detection or upstream provenance becomes the dominant trust infrastructure is an open question that no single fundraise settles.

Sources