Why one verb became a red flag

Invisible characters are the easy half of the problem. They are mechanical: a codepoint is either present or it is not. The harder half is that generated prose has a distinctive register, and readers have learned it fast.

The vocabulary fingerprint

A handful of words now function as a tell, not because they are bad words but because they appear at rates far above their background frequency in human writing. The usual suspects: to delve into a topic, a project that is a testament to something, a rich tapestry of ideas, a game-changing result, and the ever-present in today’s fast-paced world.

The individual choices are defensible. The pattern is not. When four of them land in the same three paragraphs, a reader who spends any time around model output recognises it immediately, and the credibility cost lands on the author rather than on the tool.

Structural tells

Vocabulary is only part of it. Generated text also tends toward:

• Openers that acknowledge the request before answering it. Certainly! Great question! Here is the polished version you asked for:
• Closers that offer more help. I hope this helps! Let me know if you need anything else!
• Tricolon everywhere. Three-item lists, three-clause sentences, three examples, regardless of whether the subject divides into three.
• The not only X but also Y construction, at a rate no human writer sustains.
• Typographic tidiness that no keyboard produces by default: curly quotes, an em dash with spaces around it, a single-character ellipsis.

That last group is a useful signal precisely because it is not about taste. A person typing on a normal keyboard produces " and - and .... Getting U+201C and U+2014 and U+2026 requires either a smart-quotes setting or a model that was trained on typeset text.

Substitution is not enough

The obvious move is find and replace, and it does help. Our engine ships substitutions for the clearest cases: leverage becomes use, utilize becomes use, meticulously becomes carefully, a plethora of becomes many.

But some of the list resists substitution. Landscape and realm are only tells in figurative use, and the same sentence pattern is fine in a literal one. In conclusion might be exactly right in a formal report. For those, the engine counts and flags without rewriting, and leaves the decision where it belongs.

That distinction is the whole design principle. A governance tool should be confident about mechanical facts, such as a zero-width space being present, and deferential about judgement calls, such as whether a particular metaphor earns its place. Conflating the two produces a tool that damages good writing while congratulating itself.

Where this goes

Cleaning after the fact is the free tier and always will be. The more interesting version enforces the rules during generation, with an industry profile that knows what a legal citation looks like or which disclaimer a financial summary requires. That is the governance API, and it is what the waitlist is for.