90+ phrases that scream "an AI wrote this."

Why these phrases give AI away

Language models are trained to produce the most statistically comfortable next word — which is why they reach for the same "safe" vocabulary millions of times a day. None of these words is wrong. The tell is density: three pompous adjectives in a paragraph, transitions opening every sentence, an em-dash every second line. Human writers are messier — and the mess reads as honest.

The blacklist below is organized by category, with a plain replacement for each entry. It reflects the current generation of models (GPT-5.x, Claude 4–5, Gemini 3.x); we update it as the fashions shift — see the changelog.

Why models write like this (the 30-second version)

A language model predicts the most probable next token given everything it has seen. Words like "delve", "moreover" and "crucial" are probability magnets: they fit almost any sentence, offend nobody, and appeared in exactly the kind of polished web prose the models were trained and fine-tuned on. Reinforcement tuning makes it worse — raters rewarded text that sounded confident and complete, which selected for exactly the padded, transition-heavy register this blacklist catalogs. The model isn't being lazy; it's being optimal against a target that wasn't "sound like a specific human."

That's also why the blacklist works: when you ban the probability magnets, the model is forced off the beaten path into choices a person might actually make. It's not about tricking anyone — text with varied rhythm and concrete words is measurably easier to read. The tells and the quality problem are the same problem.

How many tells is too many?

From our editing work, useful thresholds per 1,000 words: 0–4 tells — normal human range; professional writers land here without trying. 5–11 — noticeable; an attentive reader senses "something off" even without naming it. A single humanizing pass fixes it. 12+ — unmistakable machine register; most editors call it on the first paragraph. The in-browser checker above computes exactly this number. Two caveats: short texts distort the ratio (one "moreover" in a 60-word email is fine), and genre matters — corporate reports tolerate more "robust" than a personal newsletter does.

Model dialects: not all AI sounds the same

The families have accents. GPT-5.x output leans on em-dash chains, "it's less about X, more about Y" pivots and rule-of-three lists; its register is confident-consultant. Claude writes cleaner prose but over-hedges ("it's worth noting", "generally", "typically") and loves structured enumeration; its register is careful-professor. Gemini is the most adjective-dense of the three — "vibrant", "dynamic", "comprehensive" cluster heavily — with a marketing-brochure warmth. Knowing the dialects helps twice: you can predict what to fix before scanning, and when editing mixed-source documents you can even tell which sections came from which tool.

Using the blacklist without lobotomizing the text

The goal is de-roboting, not word laundering. Three rules keep you honest. Replace for meaning, not for camouflage: if "crucial" is truly the right word, keep it — one occurrence is a word, five are a pattern. Fix the sentence, not just the word: "moreover" usually signals two sentences glued together that should either merge properly or separate; swapping in "also" hides the seam without repairing it. Never substitute into synonyms you wouldn't say aloud: cheap paraphrasers turn "utilize" into "make use of" — a worse tell than the original. When in doubt, read the sentence out loud; your ear catches what the list can't.