A reporter files a clean feature on deadline. Two days later a reader emails the editor claiming the piece was written by ChatGPT. The reporter wrote every word. But the story opened with "in an era of rapid change," described a source's career as "a testament to perseverance," and closed by noting the issue "remains a complex and evolving landscape." Case closed, as far as that reader was concerned.
This is the new occupational hazard of writing for a living. It is not that journalists are secretly outsourcing their work to chatbots, though a few high-profile cases have made everyone paranoid. It is that the shared vocabulary of professional writing, the stock phrases that filled trend pieces and profiles long before 2022, has been absorbed and amplified by AI models to the point where using it now looks like a confession.
Why journalism is uniquely exposed
Every profession has its clichés, but journalism's problem is structural. AI models were trained on enormous amounts of news copy. Wire stories, explainers, features, op-eds. When ChatGPT writes, it writes in a register that is heavily flavored by decades of published journalism. Which means when a journalist writes naturally, in the style their industry taught them, they are writing in the same register the machine learned to imitate.
The overlap is not hypothetical. Researchers tracking word frequency in published text found that certain words exploded in usage after AI writing tools went mainstream. "Delve" is the famous one, but the list includes "landscape," "navigate," "robust," "underscore," and "pivotal," all staples of news writing for generations. The machines did not invent these words. They learned them from journalists, then used them so relentlessly that they poisoned the well for the people who taught them.
Now the suspicion flows backward. Editors screen freelance submissions through detectors. Readers play amateur forensics in comment sections. And working reporters, especially freelancers whose next commission depends on trust, find themselves defending copy they sweated over.
The detectors making it worse
The obvious response, running copy through an AI detector to prove innocence, turns out to be a trap. Detectors measure statistical predictability, and professional news writing is deliberately predictable. Inverted pyramid structure, standardized attribution, house style rules, tight sentence length ranges. Everything that makes wire copy clean and scannable also makes it score as machine-like.
The tools are also demonstrably biased. A Stanford study found that AI detectors misclassified writing by non-native English speakers as AI generated more than half the time, a finding with obvious implications for international newsrooms and stringers filing in their second or third language. Stanford HAI's summary of the research is worth reading in full for anyone whose newsroom relies on detector scores. A tool that punishes standard, grammatically careful English is a tool that punishes exactly the people trained to produce it.
So the detector route is a coin flip, and the score comes with no explanation. A freelancer told their piece scored 74 percent AI has no idea what to change. This is where phrase-level analysis is more useful than a verdict. A tool like the AI phrase checker does a narrower and more practical job: it scans copy for the specific words and constructions that pattern-match to AI output and flags them individually. Instead of a mysterious percentage, you get a list. This sentence leans on "navigating the complexities of." That paragraph has two "underscores" and a "pivotal." You can fix a list.
The phrases quietly undermining your byline
Some of the biggest offenders in news copy are worth naming, because most reporters use them without noticing.
The scene-setting opener is the worst. "In an era of unprecedented change." "As artificial intelligence reshapes the modern workplace." These wide-angle openings were tired before AI existed. Now they are radioactive, because they are precisely how a chatbot begins every piece it is asked to write. An anecdote, a scene, a number, a quote. Almost anything beats the panoramic throat-clear.
The hedged verb cluster is next. "Could potentially," "may serve to," "seemingly appears to." Journalism has legitimate reasons to hedge, but AI hedges compulsively, stacking qualifiers until sentences say nothing. One hedge per claim. Pick your best one.
Then there is the empty elevation. Every source is "passionate." Every initiative is "innovative." Every career trajectory is "a testament to" something. Every challenge exists within a "landscape" that someone must "navigate." Human readers have seen these constructions ten thousand times in AI output and their eyes now slide off them, or worse, snag on them.
And the summary kicker. "Only time will tell." "The debate is far from over." "One thing is certain: the conversation is just beginning." A kicker should land on an image, a quote, or a fact. AI models cannot do that, because they have no reporting to land on. Which means a reported kicker is one of the strongest human signals available, and a vague one is one of the strongest machine signals.
Reporting is the moat
Here is the genuinely good news for journalists. The single most reliable marker of human authorship is the one thing AI cannot fake: original reporting.
A language model can produce a plausible 800 words on housing policy. It cannot produce the thing the councilwoman said off the record and then agreed to put on it. It cannot describe the smell of the flooded basement or the exact wording of the eviction notice taped to the door. Specificity earned through reporting is unfakeable, and readers can feel it even when they cannot articulate why.
This suggests a practical shift in how reporters spend their editing passes. The old instinct was to smooth copy, standardize it, sand off the odd edges. The new instinct should run the other way. Keep the odd detail. Keep the quote with the weird syntax instead of paraphrasing it into blandness. Let sentence lengths swing. A four-word sentence after a forty-word one is not sloppy anymore. It is a signature.
Voice matters more than it has in decades, and for a blunt economic reason. Generic competent prose now costs nothing to produce. If your copy is indistinguishable from what an editor could get from a chatbot in eleven seconds, your rate is negotiating against free. The freelancers commanding strong fees in this market are the ones whose ledes could not have been written by anyone else, because they are built from access, observation, and a recognizable way of seeing.
A pre-filing routine for the paranoid age
None of this requires abandoning craft. It requires a short extra pass before filing.
Read the lede and the kicker aloud. These are the two places AI patterns cluster and the two places editors and readers judge hardest. If either could sit on top of any story on the same topic, rewrite it around something you personally witnessed or were told.
Run a phrase check. Not for a score, for the list. Swap out the flagged constructions for plainer or more specific language. "Underscores the importance of" becomes "shows why." "Navigating a shifting landscape" becomes whatever is actually happening to whoever it is happening to.
Count your hedges and your triplets. AI loves three-item lists and stacked qualifiers. Humans, when they are saying something they know firsthand, tend to just say it.
Then file, and keep your notes. In a climate where accusations arrive by email, drafts, recordings, and notebooks are the ultimate detector-proof evidence.
The irony of the moment is that AI has not made good journalism less valuable. It has made the visible evidence of journalism, the texture of real reporting in the prose itself, the most valuable thing on the page. The delves and landscapes were always filler. Now they are liabilities. Cutting them is not a defensive chore. It is a return to what the job was supposed to sound like all along.
