The Witch Trial

An encyclical and a number

In May 2026, Pope Leo XIV publishes the encyclical Magnifica Humanitas — the year’s weightiest religious document and a defense of the human in the age of language models. Days after its release, someone runs the text through a substrate detector — a tool that scans a text for machine mediation. The tool reports: a substantial portion is machine-mediated.

The report travels through the forums, the feuilletons, the comment threads. The number is discussed. The thought is not. About the document itself, scarcely a sentence falls.

That is the incident. It is worth examining its form.

The trial

A witch trial does not ask what someone has done. It asks what someone is. The stool, the water, the mark on the skin — none of these tested a deed. They tested a hidden being. That was their convenience. The woman did not have to be understood. She had only to be identified.

The substrate detector is a witch trial.

It does not read the text. It scans it for a stigma. It does not ask whether a sentence holds, whether it carries something, whether it says anything. It asks: human or machine, yes or no. And like the witch trial, it delivers its verdict as a probability barely contestable. One in ten thousand, says the detector. Evidence visible only to the accuser — in Salem, this was called spectral evidence. The accused cannot rebut it. She waits to see whether the mark sticks.

The confusion

The worth of a text is one thing. Its origin is another.

Whether a sentence holds depends on its relation to the world — not on the hand that wrote it. A true sentence does not become false because a machine typed it. A dull sentence does not become wise because a human typed it. Logic calls this the genetic fallacy: a judgment about origin, passed off as a judgment about the matter.

The detector automates this fallacy. It answers the question where does this come from and lets the answer stand for the question is it any good. Meaning vanishes. Not because anyone refutes it. Because no one tests it anymore.

What is lost

Reading is work. A reader weighs words. She follows a thought. She holds it against the world. It is slow, and it cannot be delegated.

The detector offers the exit. A verdict without understanding. That is its seduction: reading is hard, a yes-or-no is cheap. Whoever scans no longer needs to understand. It is labor-saving of the worst kind — it permits a text to be dismissed or consecrated without having been grasped.

In the case of the encyclical, the form was visible in its purest shape. A document about the relation between human and language model became the occasion for a debate about a percentage. The meaning evaporated. The origin question remained. That is the witch trial’s victory.

A better detector does not help

It will be objected that the tools will improve. They err today; soon they will not.

This is the false hope, because it solves the wrong problem. Even a perfect detector would answer the wrong question. Machine-mediated: yes says nothing about whether the text deserves trust, whether one can learn from it, whether one can grow against it. The fault is not technical. It is categorical. No tool that measures origin will ever measure sense. The solution is not the sharper test. The solution is to stop mistaking the detector’s question for the reader’s.

The structural side compounds the case. Every detection layer forces a countermeasure implementable over a weekend: retyping substrate output with a randomized rhythm, voice input instead of keyboard, distribution across multiple sessions, word-by-word insertion. These techniques have been documented since January 2023. Whoever detects does not win: the tool catches the unwary and lets through the ones who wash their tracks deliberately. That is the worst selection. It harms the wrong people.

The opposite of the trial

Here lies the distinction that matters.

The detector is accusation. It comes from outside, against the author’s will, and claims to see what she has concealed. It asks about being, not about deed.

Self-declaration is the opposite. The author says herself what role the Sprachsubstrat — the language substrate — has played: how much of the Wortlaut, how much of the Geist, whether the work stands in the bibliographic world. The disclosure is given, not extracted. It does not replace reading. It accompanies it. It says: This is how it was made. Now read.

The witch trial lives on the hidden being. Where the hidden has been spoken, the trial has no ground left. Whoever declares disarms the detector — not by concealment, but by answer. That is the turn: not silence protects against the hunt, but speech.

The only trial

Maschinenschrift does not detect. It declares. A text wants to be read, not scanned. That is no sentiment — it is the only trial that has ever said anything about a text: to read it and test whether it carries.

The project’s diagnostic formula is Geist aus der Maschine. The seal answers it not in secret, but in the open. It does not count the worth of the work; it counts what the language substrate has supplied. What is to be said beyond that, reading says.

More on the form and the lineage: Idea. Why the project says Sprachsubstrat rather than machine or LLM is derived in a note of its own: Substrate, not Machine. Why it runs no detection — in another.