In an unexpected paradox, some writers and students are deliberately making language errors and abandoning their usual styles, fearing that AI-detection tools will classify their texts as machine-written.
اضافة اعلان
These tools were designed to distinguish between human writing and AI-generated content, but they appear to be affecting the way humans write themselves.
Mastering a language or using polished rhetorical structures is no longer enough proof of a writer's skill. It can instead become a reason to question the originality of their work. As AI-detection tools spread, some authors are rewording their sentences, avoiding certain punctuation marks, and even inserting deliberate errors into their texts, simply to prove that they wrote them themselves.
A report published by Bloomberg on 30 October 2026 highlighted this phenomenon. It said detection tools, led by Pangram, are no longer limited to identifying machine-written text, but are beginning to influence the writing style of humans, including people who insist they did not use AI for their work.
A novel withdrawn after a test result
Novelist Mia Ballard is one of the most prominent cases that has sparked debate about the reliability of detection tools.
In March 2026, the publisher Hachette withdrew her novel The Shy Girl after Pangram classified around 78.3% of the text as AI-generated.
Ballard denied using the technology to write the novel, suggesting that her editor may have used it without her knowledge.
Doubts did not stop with that work. The Bloomberg report noted that the first chapter of her later novel Poly, which she says she wrote by hand, received a result indicating the text was entirely AI-generated, while her earlier novel Sugar scored 83%.
This case reveals a basic problem: the percentage a detection tool produces is not, on its own, conclusive proof of how a text was written, especially when the result is disputed between the writer and the body relying on it.
Why have punctuation marks become suspicious?
The problem is not limited to the content of texts. It extends to small details of writing style, such as word choice, punctuation and sentence construction.
According to data from Pangram, which is based on analysis of millions of documents, the em dash appears in AI-generated texts at a rate of 17 times per 10,000 words, compared with just twice in the human texts analysed.
An English term meaning "delving into a subject" appears 30 times in AI texts, compared with three times in human writing.
The difference also extends to rhetorical styles. The so-called "rule of three," in which ideas or qualities are organised in groups of three, appears 19 times, compared with five times in human texts.
Language models also use constructions that set two ideas against each other, such as "it is not just this, but that." According to the same data, these appear roughly three times more often than in human writing.
Although these figures reflect statistical differences between groups of texts, they do not mean that the use of a particular punctuation mark or linguistic structure proves that a text was machine-generated.
These patterns have nonetheless prompted some writers to change their language habits. Some avoid the em dash, rhetorical repetition and parallel structures, while others deliberately introduce spelling errors to make their writing appear more human.
As a result, some writers are abandoning styles they used for years before AI became widespread, fearing that these styles will become signals that cast doubt on their work.
Can detection results be trusted?
Pangram says its tool does not rely on individual words or specific punctuation marks to judge texts, but analyses thousands of linguistic signals and patterns.
The company says its latest model has a false positive rate of one case per 10,000, while acknowledging that errors remain possible.
A study published in June 2026 in the International Journal for Educational Integrity tested four detection tools on 160 academic papers.
The results showed that Pangram was the most consistent of the tools tested. It detected AI use in 37 of the 40 papers that combined human and machine writing.
It also detected the technology in 37 of 40 cases where humans had rewritten AI-generated text. However, these results measure the tool's ability to spot AI involvement in those cases. They do not prove that it can determine the source of all human-written texts with absolute accuracy.
Questions about the reliability of classification have grown with cases in which human writing has received high scores suggesting AI use.
Science writer Janelle Shane had one of her texts tested, and the tool classified it as AI-generated with a confidence level of 98.7%.
A specialist report noted that in some tests, changing a single punctuation mark, such as replacing an em dash with a colon, shifted the classification of the same passage from entirely AI-generated to entirely human.
If limited formal changes can affect detection results in this way, that raises questions about whether these results alone can be relied on to make decisions that could affect a writer's or a student's future.
Students and writers under pressure to prove their identity
The consequences of this phenomenon are not limited to the world of literary publishing. They extend to universities and educational institutions that have been using AI-detection tools since 2023.
Some students have reported being wrongly accused of using the technology, forcing them to defend the originality of work they say they completed themselves.
At the same time, new tools have emerged to help users identify words and structures that may trigger detection systems, reflecting a shift in how writing is approached.
Rather than focusing on the clarity of ideas and the quality of expression, some users are now concerned with whether their sentences will pass an automated check.
Nigerian writer Jerry Falwadi has warned that avoiding patterns that these tools classify as AI-linked could weaken literary styles shaped by different languages and cultures.
The danger is that detection tools could push writers to adhere to particular linguistic patterns, even when those patterns do not match their literary voices or cultural backgrounds.
Where does assistance end and AI authorship begin?
The issue raises another question: how to distinguish between text produced entirely by AI and text a human writes and then uses the technology to edit or improve.
According to Pangram's classifications, asking a chatbot to turn a list of points into a full article counts as using AI to produce the text. Using the tool to rephrase material a human has already written, however, falls under writing that has been improved with AI assistance.
The line between the two cases may nonetheless become unclear in practice, especially as proofreading, editing and rewording tools are used more widely.
This presents a challenge for educational institutions, publishers and media
organisations, which need to distinguish between permitted use of assistive technology and reliance on it to produce content that is supposed to be human-authored.
Is AI changing the way humans write?
Linguistics researchers argue that the influence of language models is no longer limited to producing text, but is beginning to extend to the language humans themselves use.
Some people adopt common expressions found in chatbot outputs, while others deliberately avoid these expressions so that their writing does not sound machine-like.
This interaction may give rise to a hybrid writing style, in which the characteristics of human writing blend with patterns produced by AI technologies.
As generation and detection tools continue to develop, writers may find themselves having to spend time proving they are the authors of their texts, rather than directing their efforts to developing their ideas and their work.
Writer Sally McGrow, whose work was wrongly classified, summed up this dilemma with a striking comparison, describing her experience as akin to trying to trick a machine into believing she is not a machine.
Has good writing become a burden for its authors?
These developments reveal a paradox that goes beyond the technical side. Tools designed to protect the originality of writing may push some of their users to abandon their natural styles and weaken the quality of their texts, in order to avoid being accused of using AI.
This does not mean detection tools are useless, but it does highlight the importance of treating their results as fallible indicators rather than final judgements on a writer's identity.
There is a clear need for evaluation mechanisms that do not rely on automated classification alone, but also consider drafts, the stages of development, the nature of any technical assistance used, and the quality of the ideas, sources and research.
The question this phenomenon raises is no longer limited to whether machines can imitate human writing. It is now about whether humans will have to imitate poor writing to prove they are human.
If polished writing can become grounds for suspicion, the problem may lie in the standards of the detection tools themselves, rather than in the writing styles of those who use language in their own way.
