Skip to content
replicapong

I write

How do I stop AI from changing my text without asking?

Facts as of October 17, 2026

Say it explicitly: flag, don't change. The model names the passage and what speaks against it – you decide. And when it copies over someone else's text, the rule is: Don't correct anything when transcribing, not even typos. In verbatim quotes, even the typo stays.

In detail

A model smooths out whatever seems odd to it, and it is usually right. The case in which it isn't right is the expensive one – and it looks exactly like the others. A silent correction goes unnoticed precisely because it usually does the right thing. The rule takes away the model's decision about which case it is facing. All it has to do is flag.

Flagging also pays off: it finds things that changing would have covered up. And a wrong flag does no harm, because it stands next to your text, not in its place.

The first prompt below belongs in any work on a text you still consider yours: manuscript, minutes, talk. Alongside it, you can explicitly allow small things like quotation marks, dashes and obvious spelling mistakes – except in verbatim quotes.

The second applies when the model copies over a text it didn't write itself: transcribing, pasting, reformatting. It only takes effect where the model recognizes the text as someone else's. So write it into every single transfer task, not just once into the general instructions, and pass it on to every helper instance that does a transfer.

The protection is aimed at the AI, not at you. What you want to change yourself, you change.

To copy

Flag factual errors and internal contradictions; don't change them. Name the
passage, say what it says and what speaks against it, and leave the text
as it is. The author decides what gets changed.
Don't correct anything when transcribing — not even the author's typos.

In practice

In October 2026, a model fact-checked the statements of a dictated talk by Thomas J. Weiss and marked four of them as a “correction.” One of these corrections was itself wrong: he had said “about 20 euros” („etwa 20 Euro“) and meant net, while the model calculated gross prices for him. Because the rule applied, the wrong correction stood as a flag next to the dictation. He read both and decided.

In August 2026, a verbatim quote of his read Schauspier. The stored version read Schauspieler (“actor”). Nobody had ordered that, and nobody noticed it while reading; it was found by a machine comparison. Since then, the second sentence has been in every transfer task. He himself doesn't see the core of it in the typo:

Overstepping boundaries means breaking agreed rules, even if in the end this breach wasn't a particularly bad one.

(Original: „Grenzüberschreitungen sind das Brechen von vereinbarten Regeln, auch wenn dieser Bruch am Ende kein besonders schlimmer war.“)

Sources

  • Prompt „Melde Sachfehler, ändere sie nicht“, 10.10.2026
  • Prompt „Korrigiere beim Abtippen nichts“, 16.09.2026
  • Rules of the theater project, review of 02.09.2026

Cite: Thomas J. Weiss: “How do I stop AI from changing my text without asking?” replicapong.de, as of October 17, 2026. https://replicapong.de/en/question/dont-change-my-text/ – CC BY 4.0

Related