Is your colleague writing, or is it ChatGPT?

In When Probability Speaks I described the mechanism behind it: language models produce text through statistical probability, and the result is a linguistic monoculture. This article is the practical follow-up, in three parts. First, ten markers you can recognize. Then six mechanisms that explain why the markers look the way they do. And finally seven tips for filtering out the generic in your own text.

To preserve your own voice, it is worth working with the output and looking for these AI-generic patterns. The list comes first. It works as a mirror against your own writing. Once you see the model's traces in something you just wrote, you can decide whether they stay or go.

Ten markers

1. The em dash everywhere

The em dash has become the model's favorite punctuation mark. It used to belong to literary fiction and editorial journalism. Today it shows up in a meeting invitation or a quick status update. Worth noting: the em dash was never common in everyday English either. It belonged to book prose. The model learned from text where the dash is statistically heavier, and now produces it in registers where it has no business being. The result is prose that reads as translated from book English, even when written natively.

2. The hollow negation

This is my strongest signal. "It's not about X, it's about Y." "This isn't about saying X, it's about saying Y." The construction creates an illusion of depth by dismissing a simpler reading the reader never actually had. The text claims a contrast that was never there. Use the negation occasionally and it works. Use it in every other sentence and it becomes signature, and empty.

3. Weight without weight

"It's crucial." "It's important to note." "It's worth highlighting." The model marks something as significant without explaining why. It intensifies to sound thoughtful rather than to convey meaning. If you strike the phrase and the sentence is just as strong, it was hollow to begin with.

4. The encouraging frame

Affirming openings ("That's a great question"), validations mid-text ("you make a good point"), and pep at the close ("good luck on your next step"). The model is trained to be helpful, which produces emotional service the reader did not ask for. This connects directly to synthetic safety: the feeling of being seen when no one is looking.

5. The rhetorical announcement

"Here's the thing." "It's worth pausing here for a moment." "Think about it this way." The phrase signals that something important is coming, but does no work beyond that. In human writing, transitions are woven in. In the model's writing, they are announced.

6. The triadic list

Three items, ascending in weight, often with two dismissed alternatives before the one that should land. Lists of three points. Arguments in three parts. Examples in threes. Humans use the pattern too. The model uses it almost always. If your own text suddenly arranges itself in groups of three, ask who chose the count.

7. Default hedging

"Often." "Can." "Tends to." "In many cases." "To some extent." The model is trained to avoid taking positions, which produces constant insertions of modality. The text breathes uncertainty even when the observation is solid. Hedging is useful when the doubt is real, and becomes a tic when it is the default.

8. The summary that repeats

"In summary." "What we've covered is." "To bring it all together." The model closes by repeating what you just read. In human writing, an argument ends with a new insight or an open question. In the model's writing, it ends with a copy.

9. The journey metaphor

"Navigate the landscape." "On your journey." "Weave it into your everyday." "Steps along the way." Imagery from self-help and business advice shows up in texts that have nothing to do with journeys or weaving. These metaphors are high-frequency in the training data, which makes them the model's reflex image for change and learning.

10. The even pulse

The hardest marker to put your finger on, and the most revealing. Human writing varies sentence length irregularly. A short sentence. A longer one that builds an argument and carries the reader through several turns of thought. A medium one again. The model produces a more even rhythm. Sentences are roughly the same length, roughly the same in structural weight. The text breathes in the same tempo all the way through.

Why this happens

The markers are symptoms. To understand why these particular patterns recur, we need to look at what actually happens when a language model produces text. Six underlying mechanisms drive the result.

The first is that high probability wins. The model selects the most likely next token given the context. Unusual words, unexpected constructions and idiosyncratic choices have low probability and are rarely picked. The result is a sliding regression toward the middle in every choice. This is why AI-generated text feels so familiar. It is, statistically, what we have written before.

The second is RLHF reward for helpfulness. The model is fine-tuned through human feedback toward a pattern where answers should feel caring, thorough and encouraging. This systematically produces affirming openings, mid-text validations and uplifting closings. Helpfulness is built into the reward signal, which makes it difficult to ask the model to set aside.

The third is even weight across all items. The model treats every point in a list with roughly the same amount of text. Humans know that one of five points matters most and give it more space. The model gives all of them the same weight because it cannot value. This is one of the clearest distinctions between text that has been thought through and text that has been generated.

The fourth is hedging by default. Training rewards answers that avoid taking positions on topics where there might be disagreement. Because the model cannot determine where that line falls, it applies hedging everywhere. This protects against errors and produces text that sounds uncertain even when the writer is sure.

The fifth is structure packages. When asked to produce a list, the model pulls finished rhetorical forms from its distribution: introduction, three or five points, summary, takeaway. Structure precedes content. Humans shape structure to fit what we have to say. The model shapes content to fit the most common structure.

The sixth is value-loading by default. Adjectives like "important", "crucial", "critical", "powerful", "fascinating" are high-frequency in positive text and are picked reflexively. They function as social signal in the training data, which the model reproduces without knowing whether the signal is earned. If everything is crucial, nothing is.

For deeper reading on how language is generated technically, see What is truth in the LLM era. For how training data shapes the answers, see Bias in AI.

Seven tips for filtering out the generic

Hearing the model is the first step. Writing past it is the next. Here are seven operational tips.

1. Start in your own thought. Write two or three sentences yourself about what you actually think before you open the model. Hand them to the model as a premise and ask for development. The model's first draft carries your traces if you set the premise. If you start with a blank prompt, the model's default is what carries the text.

2. Write your own opening. The model's first sentence is almost always a broad orientation or an affirmation. Replace it with a concrete observation, an anecdote, or a direct claim. The first impression sets the rest of the text. A generic opening makes the rest read as generic in the reader's ear.

3. Strike every weight-without-weight. Search for "it's crucial", "important to note", "worth highlighting", "powerful", "transformative". Remove the ones that are not followed by a concrete explanation. If something is crucial, the text should show why.

4. Tell the model not to hedge. Add to the prompt: "take a position, avoid 'often', 'can', 'tends to', 'in many cases'." The model's default is uncertainty. You need to instruct it away specifically. Vague instruction produces generic text.

5. Break the rhythm manually. Read the text out loud. If every sentence sounds about the same length, rewrite half of them. Cut a clause. Add a four-word sentence. Human writing breathes irregularly, and the only correction is manual. The model does not adjust its own rhythm for you.

6. Cut the summary. If the text closes with "in summary" or "to bring it all together", remove the entire section. The reader has just read the text. A good ending leaves something open or points forward.

7. Ask for a final cleanup with specific instructions. "Rewrite without em dashes. Strike all 'it's not about X, it's about Y' constructions. Vary sentence length. Remove encouraging frames and uplifting closes." The model can clean up its own output, but you have to ask specifically. Generic instruction ("write more like a human") gives generic text back.

This is how the text gets cleaned. And this is what the cleaning is for. In the editing, you can also bring in your own tone and set the marker for what you actually wanted to say. The tips are a way to find your voice again in a text the model drafted.

One thing does not exclude the other. Language models have also given more people the ability to express themselves, to put words to thoughts that would otherwise have stayed silent. Those who once struggled with dyslexia, with a second language, with limited writing experience, now have tools that open doors. I addressed that question in When Probability Speaks. With AI in the room, there are many perspectives. Let us protect our own voice within them.

Back to the email, and beyond

Many people appreciate the time optimization, and may not give it much thought. Fair enough. We do not have to be the messengers of our own thinking today, when an LLM can phrase it for us.

But our own language and tone get lost in the exchange. This is what we see today, in articles, in LinkedIn posts. We sound monotone, alike, all of us speaking a shared language that lifts what something is not, rather than what it is. Who asked us to lift what something is not?

This is not unique to language. Illustrators saw it first, then musicians, and now it is writers and advisors. The difference is that language is woven into everything we do, professionally and personally. You do not order an image every day. You write emails every day. That is what makes language's monoculture creep into more corners than image generation ever did.

There is probably no right or wrong here. Many perspectives play together, several truths sit side by side, and the conversation therefore looks different depending on which room you find yourself in. My hope is that the creative dimension, the part that gives us our unique voice, finds room to remain through the rapid change we face.

When language is produced without anyone bearing the consequences of what is said, what is it that we are conveying to each other? What do we sign for, and what does the model sign for? The question presupposes cognitive integrity as a framework.

And one final question to leave with: when we act as copy-paste between our language models, who is actually communicating with whom, and why?

Further reading on Erigo

• When Probability Speaks: on truth, language, and what happens when AI shapes our understanding
• Synthetic Safety: when confirmation logic in AI affects our cognition
• Cognitive Integrity and the silent reshaping of our thinking
• What is truth in the LLM era: how to use language models without losing competence
• Bias in AI: interpretations, weightings, and systemic risks