You can spot it in a sentence. Here is what the tells are, why every model has the same ones, and what it takes to make them go away.
Most people can now tell when a paragraph was written by a model. They cannot always say how. The answer is that chat models share a small set of habits, and once you can name them you see them everywhere: in emails, in product descriptions, in the replies under a post.
This page lists the tells and explains where each comes from. None of them is a mystery. They are the predictable result of training a model to produce text that raters like.
"Certainly!", "Great question!", "I'd be happy to help." A model learned that answers which start warmly were preferred, so it starts every answer warmly. The opener carries no content. Cut it and nothing is lost, which is the test for filler.
Models use em dashes far more than most people do, often several to a paragraph, because the dash is a cheap way to bolt a clause onto a sentence and a lot of well-edited training text uses it. It became the most talked-about tell of AI writing in 2025. A human editor would turn most of them into a comma, a period, or parentheses.
Fast, cheap, and reliable. Not X but Y. The model favors tidy structures because they are common in polished prose and they are easy to complete. Once every list has three items and every contrast is perfectly balanced, the writing feels manufactured, because it is.
"It depends", "there are many factors to consider", "it is important to note that". Hedging avoids being wrong, and a model that is never wrong scores well. The cost is that the reader has to dig for the actual answer, if there is one.
A paragraph announcing the structure, headers for a four-line reply, and a closing recap of the points just made. Length reads as effort to a rater skimming two answers. To a reader who wanted the answer, it is a tax.
"Delve", "tapestry", "navigate the complexities", "in today's fast-paced world". These are words that are slightly formal and slightly rare in ordinary writing and very common in the kind of text the model was rewarded for producing. They sound written rather than said.
Think of a new employee who learned to write from a stack of corporate blog posts. Every habit of that genre is now their voice, including the ones nobody actually likes. They can write differently, but only if someone tells them to, and reminds them every time.
Berges AI strips these habits on every message: no warm-up, no em dashes, no recap, no headers on short answers, prose that gets to the point. It is enforced by rules that run on every response, not by asking the model nicely once. You can read what they do on the interceptors page.
Try Berges AIMostly, yes. Tell it what to avoid in concrete terms: no opener, no em dashes, no closing summary, no lists unless asked. Give it a sample of your own writing to match. Expect to repeat the instructions, since long conversations drift back to the defaults.
Poorly. They look for the same tells, which means they flag careful human writers and miss a model that was told to avoid the habits. Treat any detector score as a weak hint, not a verdict.
No. It is a fine punctuation mark. The problem is frequency: one in a page reads as style, five in a paragraph reads as a model. If your own writing uses them, keep using them.
They are trained on overlapping data with similar methods, and tuned on human ratings that reward the same qualities. The habits are the fingerprint of that process, not of any one lab.