The hidden cost of letting the machine finish your sentence

Two recent opinion pieces argue that the debate over AI-assisted writing should focus less on whether text can be detected and more on what readers lose when prose becomes generic, imprecise, or difficult to trust.

Dima Bolkunov, writing on his personal blog, and Phil Eaton of The Consensus Blog approach the issue from different angles. Bolkunov examines how AI-generated text can weaken an author’s voice and reduce precision. Eaton describes how a technical publication is using a rule-based system to identify articles that strongly resemble common AI writing patterns, then limiting their promotion.

Both writers make clear that they are not presenting proof that every text with familiar AI patterns was produced by a model. Their concern is broader: readers may react negatively when writing feels standardized, overly long, or disconnected from an identifiable author.

Style carries context

Bolkunov defines AI writing as text in which AI generates at least half the words. He argues that a writer’s style conveys more than personality. It also gives readers clues about the writer’s perspective, context, and level of engagement with the subject.

His argument draws on his earlier academic work on authorship attribution. Even short passages, he writes, can reveal patterns in word choice, sentence structure, and storytelling. These patterns make it easier for readers to recognize a human voice.

Generative AI, by contrast, tends to produce an averaged style, Bolkunov argues. Prompting can alter the output, but he says it remains difficult to see the individual author behind the text. That can make it harder for readers to empathize with the writer or interpret ambiguous passages through knowledge of the writer’s usual voice.

He also argues that AI can reduce precision. A model must fill gaps in a prompt with assumptions. To avoid obvious contradictions, it may choose broader and less specific wording. A human editor may miss that ambiguity because they already know what they intended to say.

Writing itself also serves as a form of thinking, Bolkunov says. Drafting forces authors to confront gaps and inconsistencies in their reasoning. If a person moves straight from an idea to model-generated copy, that opportunity for scrutiny may shrink.

For readers, he says, vague language and verbose output increase the time needed to understand a text. Any time saved by the writer may be outweighed once many people need to read, question, revise, or verify the result.

Publishers respond to reader preferences

Eaton reports that The Consensus has built an “AI tells” detector for its feed of software infrastructure articles. The system does not use machine learning and does not claim to determine whether AI wrote an article. Instead, it applies static rules based on writing patterns Eaton associates with AI-style prose.

According to Eaton, the publication labels articles that fall in the 99th percentile for these signals. Those articles can still appear in the feed, but they cannot receive certain promotional placements, including the site’s daily recommendation section, newsletter, social media promotion, and an icon for selected articles.

Eaton says the tool can be wrong and has flagged some of his own writing. He also acknowledges that writers could adapt prompts or editing practices to avoid common AI patterns. In his view, that would not necessarily be a problem if it produced writing with more distinct character.

The two pieces point to a practical issue for content teams: AI use may be less controversial when it supports research, outlining, or editing than when it replaces the writer’s judgment and voice. For publishers, the emerging question is not only whether a model contributed to an article, but whether the finished text remains clear, specific, and recognizably accountable to a human author.

Sources

• On AI writing – Dima Bolkunov’s Blog
• Has AI tells – The Consensus Blog