Humanums vs Pangram: you cannot argue with a score, but you can attach evidence.

Substack added an AI detection feature built on Pangram. Readers can choose Scan for AI text and get back Pangram's breakdown of how much of a piece it reads as AI-generated, AI-assisted, or human. It applies to posts and Notes published on or after 21 July 2026.

The reaction from writers was immediate and mostly about misclassification. Writers disputed results on pieces they had written themselves. Others raised the detection bias that has been documented against second-language writers and against people whose prose is unusually regular for any reason. Concerns were also raised about the age of the training data behind the model.

Set aside whether any individual complaint is correct. The structural fact is that a scan is now something that happens to a piece of writing, initiated by someone other than the writer, with a result the writer sees at the same time as everybody else.

Pangram publishes a false positive rate on the order of 1 in 10,000 and points to third-party evaluation rather than resting on its own benchmarks. That is a genuinely strong claim, and it is stronger than most of this category can support. Critics note the figures come from controlled conditions, which is a fair caveat, but it is not the interesting problem.

The interesting problem is what happens when you multiply. Arvind Narayanan worked through the arithmetic on the published rate: a student produces something like 500 to 1,000 pieces of written work across four years, so if every one of them were scanned, 5 to 10 percent of a student body would be falsely accused at some point. The rate is excellent. The volume is what turns it into a policy problem.

The same shape applies to a publishing platform. Across a large enough archive, even a very accurate classifier generates a steady trickle of writers who are wrong about nothing and flagged anyway. For the platform that trickle is a rounding error. For the writer inside it, it is the whole event.

The practical difficulty with any detector result is that it comes with no reasoning. You are handed a number. You cannot point to the sentence that caused it, you cannot show the model was wrong about a specific thing, and you cannot rerun it under conditions you choose. Disagreeing gets you the only response available, which is to assert that you wrote it, which is exactly what someone who did not write it would also say.

Rewriting the piece to score better does not work either. It reads as tampering, it usually makes the writing worse, and the next model retrain moves the target anyway. Nothing available to you inside the finished text resolves a question about where the text came from.

That is the gap. Not accuracy, which Pangram takes more seriously than most, but the fact that the writer has no move.

Humanums works from the other end. While you write, it records keystroke cadence, pause structure, revision depth, paste behaviour, session timing, and how content accumulated over time. It never stores the characters you type. When the draft is done you certify it, and the result is a signed certificate, a badge, and a verification page at a public URL.

That evidence exists before anybody questions the piece, which is the part that matters. A reader who wants to check does not have to trust your account or trust a platform's classifier. They open the verification page and look at the record.

It also stays attached. Republish the piece on your own site or in a newsletter and the badge and link travel with it. A scan result cannot do that, because a scan is a moment rather than an artefact.

Pangram is doing a job that certification cannot do. If you have received text from a stranger and want an estimate of where it came from, a classifier is the only option, and a well-evaluated one is better than a badly evaluated one. Platforms moderating a large archive of writing they did not commission have essentially nothing else to reach for.

Humanums cannot help there at all. It has no opinion on text it did not watch being written. If your archive is already published, there is no session left to record, and no certificate can be issued after the fact. That is a real limit, not a soft one.

So the honest framing is not that one tool is better. It is that a probability estimate and a session record are different kinds of evidence, and a writer who wants to be able to answer the question should be holding the second kind.

A certificate establishes that the certified text was typed by a human, with human pacing, pauses, and revisions, inside a monitored session, and that it has not been altered since. It does not identify which human was at the keyboard. It does not prove no AI was consulted while you were thinking about the piece. It does not say anything about whether the ideas are original.

Those boundaries are worth stating plainly, because a proof mechanism that quietly overclaims is just a detector with better marketing. What the certificate is good for is answering the specific question a scan raises, which is whether a person actually composed the words.