Common AI Words and Phrases That Get You Flagged

After a model is asked to write, it reaches for the same safe vocabulary every single time. None of these words is wrong on its own. The problem is density and predictability: when a paragraph stacks five of them together, a reader and a detector both feel the machine behind it. Here are the worst offenders I see, with a quick note on each.

Detectors do not keep a banned-word list. They score two statistical signals: perplexity and burstiness. Perplexity measures how surprised a language model is by your next word. When a model writes, it picks the most probable word at each step, so the output has low perplexity by design. Words like delve, moreover, and underscore are exactly the high-probability tokens a model defaults to, which drags the whole passage toward the predictable, machine-like end of the scale.

Burstiness measures variation in sentence length and rhythm. Human writing swings between short punchy lines and long winding ones. AI writing settles into a uniform register where every sentence is roughly the same length and weight. The stock vocabulary reinforces that flatness, because it carries the same formal, even tone everywhere. So it is not the single word that flags you. It is the cluster of low-surprise words inside a low-variation rhythm. That combination is what every detector is tuned to catch.

You do not need to memorize the list. You need to break the patterns that produce it. Five moves do most of the work.

Not directly, and this is the part people get wrong. There is no list inside GPTZero or Originality.ai that says flag any text containing delve. If there were, you could beat detection by find-and-replace, and detection would have collapsed years ago. Instead these tools train on millions of human and AI samples and learn the statistical fingerprint of generated text.

The catch is that these words correlate hard with that fingerprint. A draft that overuses delve, moreover, and underscore almost always also has low perplexity and flat burstiness, because the same model behavior produces all three at once. So removing the words helps, but only because it forces you to rewrite the structure underneath them. Swapping synonyms while keeping the robotic rhythm will not move your score. Rewriting the sentence will.

Different models have slightly different verbal tics, but ChatGPT is the loudest. Its signature openers are delve, navigate, and embark. Its signature connectives are moreover, furthermore, and additionally. Its signature adjectives are comprehensive, robust, seamless, and crucial. Ask it to wrap up and you will almost always get in conclusion or ultimately.

The reason is the training and the reinforcement tuning that rewards safe, fluent, agreeable prose. Safe and fluent means high-probability words, and high-probability words are exactly the ones on this page. If your draft reads like a polished essay that somehow says nothing surprising, these words are usually why.

Most people catch the obvious tells and miss the quiet ones. Hedges like it is worth noting and it is important to remember slip through because they sound careful. Empty intensifiers like truly, incredibly, and remarkably survive because they feel like emphasis. Structural crutches like firstly, secondly, and lastly read as organized but mark the text as generated.

These are the ones to hunt on a second pass. They do not jump out, but stacked together they keep your perplexity low and your detection score high. Strip them and the prose immediately reads more like a person and less like a model trying to sound thorough.