AI2 min read

The banned word list, and why a list is not enough

Every word here appears in real writing. They are banned because the density in generated text gives it away.

Noomachy DevTracker checks every generated post against a list of words and phrases. Not because any of them is wrong, but because their frequency in machine written text is high enough to be a signal.

Here is the list and the reasoning.

Marketing verbs

leverage, unlock, empower, elevate, supercharge, revolutionize, transform

All of these have plain equivalents. "Leverage" is "use". "Unlock" is "enable" or nothing. The plain version is shorter and does not sound like a product page.

Adjectives that assert rather than show

seamless, robust, cutting-edge, game-changing, revolutionary, best-in-class, state-of-the-art, next-level, powerful, innovative

The problem is that they are assertions the reader has to take on faith. "Seamless" is a claim. "It handles the reconnect case without dropping the queue" is evidence.

Set phrases

"In today's fast-paced world" "It's not just X, it's Y" "Let that sink in" "Here's the thing" "And that's when I realized" "The future is here" "I'm thrilled to announce"

These read as templates because they are templates. Each signals that a specific rhetorical move is coming, and readers have learned to skip ahead.

Structural bans

Em and en dashes. The strongest punctuation level tell. Models use them constantly, most writers rarely.

Stacked fragments for emphasis. "Faster. Cleaner. Better." A drum roll where an argument should be.

Engagement bait. "Thoughts?", "Am I the only one?", "Follow for more", "Drop a comment below". Banned outright by both editorial specs this product uses. A genuine question is allowed. "Has anyone hit this with large files?" passes because it is asking for information rather than for engagement.

Why a list is not enough

Take every banned word out and the text can still read as generated, because the real tells are structural: uniform sentence length, perfectly symmetric sections, a tidy summarising sentence at the end of every part, no dead ends, nothing committed to.

The list catches the surface. It does not catch the shape.

That is why these violations are warnings rather than blocks in Noomachy. They tell you the post is drifting toward a register you probably do not want. They do not tell you the post is good, and no list can.

And why it is still worth having

The list is cheap, deterministic, and catches the most common failure. A post that says "leverage this powerful, seamless integration" is broken in a way that is trivial to detect and would otherwise ship.

Detecting the easy failures reliably is worth more than detecting the hard ones occasionally.

The hard ones are handled upstream, by starting from real commits. A post with four specific file paths and a real line count has something to be about. That is a better defence against generic prose than any list of forbidden words.

aiwritingediting

Keep reading