The Paranoia of Sounding Like AI

For most of my life, I judged a sentence by whether it said what I meant and belonged with those around it. I considered the words, the rhythm, and how much explanation the reader might need. Recently, another question has begun to interfere with all of these: could somebody decide that the sentence was written by artificial intelligence?

This is not an imaginary concern. A passage placed into one detection program may be declared human, while another identifies supposed signs of AI. Neither program knows who wrote it or watched the paragraph taking shape. It can only examine the finished language and compare its patterns with those it has learnt to recognise.

The writer knows rather more about what happened. I can remember struggling with a sentence, returning to it the next morning and eventually finding the wording I wanted. None of this appears in the completed manuscript. If a detection program produces a confident percentage, the number may seem more reliable to somebody else than my account of having written the passage.

Knowing this has altered the way I work. I sometimes change sentences that I would previously have left alone, although the revision does not always improve them. I try to remove anything that might be mistaken for evidence.

The effect became particularly noticeable with groups of three. I used this construction long before generative AI appeared. Like many writers, I learnt that three related words or phrases could give a sentence a natural shape. The first introduces the movement, the second carries it forward, and the third completes it. I did not deliberately count out each part whenever I wrote. After enough years, the pattern had become one of how I organised a thought.

Groups of three are now regularly included among the signs said to reveal AI writing. Once I became aware of this, I began noticing them throughout my manuscript. A sentence that had previously seemed natural started to look dangerous. I might remove one element, separate the thought or replace the whole construction with something less likely to attract attention.

The sentence remained acceptable, but it was no longer quite the sentence I would have written.

Dashes have caused a similar problem. I never used them constantly. They appeared occasionally when a pause seemed helpful or when a qualification belonged within the sentence without requiring a separate one. There were times when a comma felt too slight and a full stop interrupted the thought too heavily. The dash occupied the space between them—which was precisely why I used it.

I now use dashes so rarely that I sometimes notice their complete absence from several pages. I replace them with other punctuation before anybody has the opportunity to question them. The meaning usually survives, although the pace of the sentence may have changed.

Writers can manage without one group of three or find an alternative to a dash, but these decisions begin to accumulate. Eventually, they reduce the number of choices and begin to interfere with the style developed through years of reading and writing.

Other features of prose have also become suspect. A sentence may appear too balanced, an explanation too orderly, or a conclusion too carefully connected with what came before it. Particular words are circulated online as evidence of AI, even though many were in ordinary use before language models began producing them. Writers who encounter these warnings may start avoiding the words, even if they are the most accurate ones.

Some concern about generated writing is understandable. AI frequently produces prose with repeated structures and explanations that continue after the point has already been made. The same can be true of human writing, which is why a familiar pattern cannot establish who produced a passage. At most, it provides a reason to look more carefully.

The problem will likely become harder because AI did not invent most of the features it imitates. It learnt them from human language. Groups of three, parallel clauses and occasional dashes entered its writing because they were already present in ours. As language models become more capable, they will reproduce a wider range of human styles, including some that currently appear individual.

If every successfully copied feature then becomes another warning sign, writers will be expected to abandon more of what they once used naturally. Avoiding a few overworked phrases may be easy, but there is no clear point at which the process must stop. Almost any recognisable habit might eventually be reproduced often enough to become suspicious.

There is something particularly strange about deliberately making prose less fluent to prove that a human being wrote it. An awkward sentence may satisfy a detection program, though awkwardness tells us nothing about authorship. The writer is left performing what a machine has decided human imperfection should look like.

For an independent author, an accusation can be difficult to answer. Drafts, notes and earlier versions may exist, but few readers will ever examine them. A doubtful remark can be made in a moment, while any attempt to explain how the work was written risks sounding defensive. The safest course may be to remove whatever might arouse suspicion before the question is asked.

The developments of recent years, or perhaps I should call them a revolution, have made me increasingly nervous about my own writing. I find myself waiting on tenterhooks for reviews and comments, wondering whether somebody will accuse me of using AI. Nothing may have been said, yet the possibility is already present while I work. I have become defensive in advance of an accusation that may never come.

Even individual words are now marked as AI-recognisable and treated with suspicion, regardless of being used as part of ordinary written English for generations. I find the exact word needed for a particular sentence, then hesitate because its appearance could be misinterpreted. Leaving it in place can feel like taking an unnecessary risk.

This begins to resemble paranoia, and it has affected much of my creative work. I doubt that I am alone. Other writers must also be looking back over sentences they once considered entirely natural and wondering how those sentences might now be judged by a program or a reader alert for certain patterns. Attention that once belonged to the argument, the language and the reader is diverted towards avoiding evidence of something that may not have happened.

That safety carries a cost. A writing style develops slowly, often without the writer noticing every part of it. Certain sentence lengths become comfortable, particular forms of punctuation suit the way a thought develops, and familiar structures help a difficult idea reach the page. They are among the tools of the trade, and a writer works with fewer tools each time one is set aside through fear.

I prefer to judge a sentence by asking whether it is clear, whether it is accurate and whether it sounds like me. Even writing those three questions now makes me hesitate. Should one be removed simply because three once provided the form that came naturally?

What I fear most is that this interference with established writing practices has only begun. AI will keep learning from human prose, including the habits that distinguish one writer from another. If each successfully imitated feature is then treated as evidence against us, the accepted forms of writing will be invaded still further. More words will become suspect, more structures will be avoided, and more writers will begin altering their work in anticipation of an accusation.

I do not know where this will end. Groups of three have already become something I watch for, and the dash has almost disappeared from my writing. Other familiar practices will probably attract the same suspicion, and I may remove those as well, although each removal leaves me with fewer ways to express what I intended.

Eventually, the argument may still be mine while rather less of the writer remains on the page.