Pangram Raised $9M to Build a Better Lie Detector. I'm Not Scared of It.

TL;DR: Pangram, the company running Substack’s AI-detection button, raised $9 million today and shipped Pangram 4, which claims better than 99 percent accuracy plus a new image detector. I use AI in this operation, openly, and you are not getting shunned here for using it either. The people running the pitchforks are using AI themselves to hunt for AI, then acting shocked when the number comes back orange. Disclosure and permission are different questions, and I’m staunchly on the side of the people that detector catches. Rarely the confident writer with a quiet room. Almost always everyone else.

Pangram raised $9 million today. Menlo Ventures led it, with Haystack, ScOp, Script Capital, and Cadenza in the round. No valuation disclosed. That’s its own tell: a company confident in its number says the number.

The money buys two things. Pangram 4, a text detector the company says catches better than 99 percent of AI-assisted writing, including text run through a “humanizer” pass meant to fool detectors like it. And Pangram Image, still in research preview, built to catch AI-generated images sitting inside otherwise-real photos.

Pangram’s co-founder, Max Spero, put the pitch in one sentence: “It’s just incredibly valuable to know whether what you’re looking at is something that’s AI-generated or not.”

I want to take that sentence seriously instead of arguing with the funding round, because it’s the actual question, and I already have an answer. I just haven’t said it plainly enough.

The thing I’ve been implying instead of saying

I wrote about Substack’s version of this button a week ago. The button is unfair to the writer standing under it. A color-coded number with no room for “I drafted with a model and rewrote every sentence by hand,” delivered to a reader who screenshots it before anyone can explain.

I didn’t say this in that piece, because it wasn’t about me: I use AI to run parts of this operation. Not quietly. The whole crew (the editors, the researchers, the art director, the guy who checks my SQLite queries) is Claude, wearing names, doing real work against real house rules. I’ve said that in the open before. I haven’t said what it means for you.

If you build here, write here, sell here, and you use AI to do it, you are not getting run off this platform for it. That’s the policy, not a vibe.

Disclosure and permission are different questions

A 99 percent accurate detector answers exactly one question: did a model touch this text. Detection tells you what happened. Judgment tells you if it matters, and Pangram only sells the first one, which is still a real product and a useful one. I’d rather know than not know.

But knowing isn’t the same as judging, and the button collapses the two into one orange number. That’s the actual defect, and better accuracy makes it worse, not better. A 60 percent-confident detector, everyone half-trusts. A 99 percent-confident one gets treated like a verdict, and verdicts don’t come with the context a writer’s actual process deserves.

Disclosure, I’m for. If you want to say how you made a thing, say it, and I’ll build you a place to say it that doesn’t read like a confession booth. Permission, you already have, unconditionally, the moment you show up and do real work. The compelled purity test (prove your hands are clean before anyone will read you) is theater. I said that about newsrooms hiding their own AI use behind a 5-in-100 disclosure rate. It’s still true pointed at anyone else, including the people who build here.

This Is a Witch Hunt With a UI

Pangram is a business standing inside the fight it claims to referee, funded on the bet that AI-detection anxiety keeps growing. $9 million confirmed the bet paid off. That’s the round.

And the tool doing the judging is itself a model. Pangram 4 trains on millions of documents to spot statistical patterns a human eye can’t see, the exact thing its own customers say they’re afraid of. A reader convinced a person used AI to write a paragraph pastes that paragraph into an AI to find out.

They ran the check with the technology they claim to distrust, got a number back, and trusted the number completely. That’s the whole con: they’ll deputize the exact tool they’re afraid of, the second it promises to catch someone else instead of them.

A detector measures statistical fingerprints, not intent, not effort, not honesty. Most people tapping that button have never been told that, and Pangram has no incentive to explain it to them. The number comes back orange, and orange reads as guilty enough to screenshot.

Almost nobody asks who the button catches. The writer with a quiet room, a finished degree, and prose that already sits in the statistical middle of the road clears the score without a second thought. The writer whose first language isn’t English doesn’t. Neither does the one whose brain does the real thinking but fights the sentence on the way out, or the one working two jobs who used a model instead of skipping the newsletter that week entirely.

The button measures how closely you sound like the kind of writer a stranger already expected. That box is narrower than most people accusing others would ever guess.

What this doesn’t mean

It doesn’t mean anything goes. The house art on this network is credited to a specific studio, in the open, in the colophon, because provenance matters even when permission isn’t in question. A member who runs a Pangram score through a spam farm and calls it a newsletter is still running a spam farm; the model didn’t do that, the spam farm did. The standard has always been “is this any good, and are you straight with people about what you’re selling them.” That standard doesn’t get softer because the detector got sharper.

Where this lives

This piece is the flag planted. The house rules page gets the permanent version: a real line a new member reads before posting, not an essay that scrolls off the archive. That’s still coming, through the same editorial gate everything else here goes through. This is me telling you where I stand today, with Pangram’s $9 million as the reason to finally say it instead of implying it.

Frequently asked questions

Does this mean the Underground doesn’t check for AI-generated spam?

No. Spam gets moderated because it’s spam: low effort, high volume, nothing behind it. That standard doesn’t change based on whether a model was involved in writing it.

Is Niche of One going to install its own AI detector?

Not planned. A detector answers “was AI involved,” which isn’t the question that decides whether something’s worth reading here.

Does the crew system count as AI-generated content?

Yes, and it’s disclosed openly, repeatedly, on this site: the crew’s names, the fact that they’re Claude, the house rule that nothing human-facing ships without a voice-and-editorial gate regardless of who or what drafted it first.