Practical
Cleaning AI text before you publish it
Text from a model is usually structurally sound and full of small artefacts. Some are invisible, some are stylistic, and only the first kind can be fixed automatically.
01The mechanical pass
Strip invisible characters and normalise exotic spaces first. These break search, corrupt data imports and produce phantom differences in version control, and they are entirely automatic to remove.
Watch for smart quotes and em dashes converted by the rendering layer. They are usually fine in prose and actively harmful in code blocks, filenames and configuration values.
If the text is going into a system that parses it — a CSV, a JSON field, a database — run the cleaned version through that parser before publishing rather than after.
02The parts a tool cannot do
Verify every factual claim, especially numbers, dates, names and citations. Models produce confident, well-formed statements that are wrong, and fluency makes errors harder to notice, not easier.
Cut the hedging. Phrases like it is important to note, in today's landscape and when it comes to are filler that survives because it is grammatical. Removing them usually shortens a draft by a fifth without losing meaning.
Break the rhythm. Model prose tends toward uniform sentence length, which reads as flat. Varying it is the single change that most improves how the text lands.
03What cleaning does not accomplish
It does not make the text undetectable, and it does not remove a statistical watermark, which lives in word choice rather than in characters. If disclosure is required by your publisher, institution or jurisdiction, cleaning the text does not discharge that obligation.
The honest framing is that this is hygiene rather than concealment: you are removing artefacts that should not have been there, not hiding the origin of the work.
Questions
Should I disclose that I used AI?
+
That depends on your publisher, employer, institution and jurisdiction, and increasingly on regulation such as the EU AI Act. Cleaning a file changes nothing about that obligation — it removes artefacts, not responsibility.
Does editing text by hand remove the watermark?
+
Light editing does not. The signal is measured across the whole sequence, so fixing punctuation or swapping occasional words leaves most of it intact. Substantial rewriting weakens it, though nobody outside the key holder can confirm by how much.
What is the single most useful cleanup step?
+
Reading it aloud. Every hedge, repetition and unnaturally even sentence becomes obvious the moment you hear it, and none of those are things an automated cleaner can detect.