A few weeks ago, I read through a stack of fifty draft cover letters for a writing workshop I run. Somewhere around the twentieth one, I started losing the ability to tell them apart. Not because the writers were bad. Most of them were sharp, several were genuinely talented, and a few had stories I would have read a whole book about. The prose itself had this strange sameness to it though, like fifty different people had agreed to write in the same voice without ever talking to each other.
You probably know the voice I mean. It is the prose that sounds professionally adequate and personally absent. Every paragraph opens with a transition word. Every conclusion gestures at growth. Every sentence is grammatically clean and emotionally inert. I started calling it cubicle voice, because it reminded me of the way office emails sound, but the more accurate term is the one most students already use: it sounds robotic.
Robotic prose is older than generative AI. We were producing it for decades before ChatGPT existed, mostly because high school and college taught us to. Five-paragraph essay templates, transition word lists taped to the wall, rubrics that rewarded "varied sentence structure" without ever defining what counted, the steady drift toward whatever phrasing felt safest to a grader. AI just industrialized the habit. Now anyone can produce a thousand words that hit every rubric category and still leave the reader with the sense that nobody, exactly, was speaking.
So how do you get rid of it?
Start by hearing it
The first move is diagnostic, and it is harder than it sounds. Most writers cannot tell when their own prose has gone flat, because they wrote it, and they remember thinking each sentence through. Reading it aloud helps. Reading it aloud to another person helps more. A piece of writing that feels merely correct on the page often sounds embalmed when you have to perform it.
There are specific markers worth listening for. Sentences that all run about the same length. Phrases that could appear in anyone's essay on anything, like plays a crucial role in or highlights the importance of. Openings that announce what the essay will do instead of doing it. Conclusions that restate the thesis in slightly different words and then thank the reader for their time without quite saying so. One thing researchers studying AI-generated text often point to is low burstiness, which is a fancy way of saying the sentences are too uniform in rhythm. The human ear catches this even when it cannot name it.
George Orwell, writing in 1946, identified essentially the same disease in Politics and the English Language. He blamed it on prefabricated phrases. Prose, he wrote, consists less and less of words chosen for the sake of their meaning, and more and more of phrases tacked together like the sections of a prefabricated henhouse. That was eighty years before anyone trained a language model. The instinct to reach for a phrase someone else has already used is a very old human shortcut. AI just made it instant.
Write toward something only you would have noticed
The fastest way to make prose sound less mechanical is to anchor it in something specific. If you are writing about leadership, do not write about leadership. Write about the Tuesday afternoon when you had to tell a sophomore on your robotics team that her wiring was the reason the arm kept failing, and how you did it, and what she said back. If you are writing about resilience, skip the word resilience entirely and tell me what the kitchen smelled like the week your mother was in the hospital.
Specificity is harder to fake than abstraction. A model can produce I learned the importance of perseverance all day long. It cannot produce the exact wrong thing you said to your father at the airport, because it was not there.
This is the principle that good writing teachers have been pushing for a long time. Helen Sword, who directs the writing program at the University of Auckland and wrote Stylish Academic Writing for Harvard University Press, argues that the cure for institutional prose is concrete language. Real nouns. Active verbs. Sensory detail. The advice sounds almost too simple to work. It works anyway.
Let your sentences disagree about how long to be
A robotic paragraph tends to have sentences of similar length, all built from similar grammar. A human paragraph wanders. A short sentence. Then a longer one that doubles back on the short one and complicates it slightly, perhaps adding a parenthetical the writer was not sure about. Then a medium one. Then a very short one, because the writer has run out of breath.
This is not really a trick. It is what actually happens when a person thinks on the page. If you go back through a paragraph you wrote and find that every sentence is between fifteen and twenty words long and starts with a subject followed by a verb, you have a problem. Break some of them. Combine others. Let one run long enough that it almost loses its footing. The page should look uneven.
Cut the words that are not doing anything
Robotic prose is bloated prose. It loves in order to when to would have done. It says due to the fact that when because exists. It uses utilize when nobody outside a software documentation team has ever utilized anything. Steven Pinker calls this category of writing zombie nouns, the long abstract phrases that should have been verbs.
Go through a paragraph and ask whether each multi-word phrase could collapse into one word, and whether each abstract noun could be a verb instead. The implementation of the policy resulted in a reduction of complaints becomes the policy reduced complaints. The second version is shorter, clearer, and more obviously written by a person who knew what they were trying to say. You can do this kind of editing on almost any paragraph and watch about fifteen percent of the words disappear without the meaning changing.
Admit something
The single most reliable signal of human writing, in my experience, is the presence of an admission. A sentence where the writer says I am not sure this is the right way to think about it, but, or I tried this and it did not work, or I keep coming back to this even though I should have moved on. Models can mimic confidence indefinitely. They can mimic enthusiasm. They have a harder time mimicking the quiet, slightly embarrassed sound of someone telling you what they actually think.
You do not have to overdo this. One real admission per essay is usually enough to change the texture of everything around it. It signals to the reader that there is a person on the other side of the page, working something out in real time, rather than performing a finished version of themselves.
A note on the work itself
The editing I described above is genuinely difficult, and most students do not have the time to do it well between classes, jobs, and the rest of their lives. That is part of why we built EssayTone the way we did, as a way to take prose that has gone stiff and loosen it back into something readable. But no tool replaces the underlying instinct, which is to listen to your own writing as if a stranger handed it to you and asked what kind of person had written it. If the answer is I cannot tell, there is work to do.
The good news is that the work is mostly subtractive. You are not trying to add personality to an essay. The personality is already there, underneath the rubric language and the prefab phrases and the transitions that came pre-installed. You are mostly just trying to clear enough away that a reader can hear you.
