AI Detector for News vs Academic Writing: Why the Same Tool Scores Them Differently

Run a hard news story and a journal article through the exact same AI detector and you’ll often get very different readings. Same tool, same underlying math, wildly different behavior. It’s not a malfunction. It’s the genres doing what genres do, and understanding the difference is the key to not being fooled by either score.

This is a look at why one detector reads news and academic writing so differently, what each style does to the numbers, and how an editor or a professor should actually interpret the result.

Key takeaways

• Detectors read texture, and news and academic writing have opposite textures. One is clipped and formulaic; the other is smooth, hedged, and uniform.
• Both score high, for different reasons. News trips on standardized phrasing; academic prose trips on its even, low-variation style.
• Neither high score proves AI wrote it. It proves the genre naturally produces the pattern detectors flag.
• The false-positive risk is real in both fields, and it lands hardest on non-native English writers, who are common in each.
• Read the score as a prompt to look closer, and verify the things a detector can’t see.

What a detector actually reacts to

Strip away the branding and every mainstream detector is measuring two things: how predictable your word choices are given the surrounding words, and how much your sentence lengths and structures vary. AI text tends to run smooth and statistically average; human writing tends to be lumpier. We wrote up the statistics behind how detectors score text in detail, but that’s the whole engine.

The crucial point: those two measurements are about style, not truth. A detector doesn’t know a fact from a fabrication. It knows whether the prose feels predictable and even. So anything that changes a genre’s style changes its score, and news and academic writing sit at opposite ends of the style spectrum.

Why news writing scores the way it does

News prose is engineered to be predictable, and that engineering is exactly what detectors mistake for a machine.

The inverted pyramid front-loads the who-what-when-where in a standardized order. AP style enforces uniform phrasing. Whole subgenres, sports scores, earnings briefs, weather, market recaps, use near-identical templates every single day. “The Fed raised rates by a quarter point Wednesday” is written a thousand ways that are all basically the same way. That low variation and high predictability is the signature detectors read as AI, so formulaic news scores high whether a reporter or a bot wrote it.

That’s why a newsroom’s real job isn’t chasing a clean score. It’s verification. A detector can’t tell whether the mayor actually said the quoted words or whether a 4.2% figure is real or hallucinated. We walked through a sane editorial approach in how newsrooms vet AI-assisted reporting, and the headline is simple: rank fact-checking far above any AI percentage.

Why academic writing scores the way it does

Academic prose trips detectors too, but through a different door.

Scholarly writing is deliberately smooth and uniform. It hedges (“these results suggest,” “it may be the case that”), it front-loads discipline-standard phrasing, and it’s built on conventional structures, especially in the methods and results sections, where a good author writes in the most neutral, reproducible way possible. That evenness is the opposite of news’s clipped punchiness, but it produces the same detector reaction: low variation reads as machine-made. A literature review packed with citation-dense, formally hedged sentences can score alarmingly high while being entirely, painstakingly human. We broke down how detectors read academic papers section by section if you want the granular view.

So both genres over-score, but the mechanism differs: news for being templated and terse, academia for being smooth and hedged. Same tool, opposite triggers.

A side-by-side scenario

Imagine feeding a detector two human-written passages.

First, a wire-style sports recap: “The Rangers beat the Kings 3-1 on Tuesday. Two goals came in the third period. The win moved them into second place.” Short, formulaic, predictable. The detector lights up, maybe 85%. A human wrote it; the genre just looks machine-made.

Second, a methods paragraph from a real study: “Samples were incubated at 37 degrees for 24 hours, then centrifuged at 3,000 rpm for ten minutes. The supernatant was collected and analyzed as described previously.” Smooth, standardized, hedged toward reproducibility. The detector lights up too, maybe 80%, for the opposite reason: the prose is deliberately even and conventional.

Two human passages, two high scores, two entirely different causes. If you didn’t understand the genres, you’d conclude both authors cheated. They didn’t. The tool measured style, and both styles happen to look like a machine.

How to read the scores without being fooled

The reliability caveats apply to both fields. OpenAI retired its own classifier in July 2023 because accuracy was too low to be useful, and a 2023 Stanford study found detectors were biased against non-native English writers, who fill both newsrooms and universities. So in either genre, a high number is a starting point, not a conclusion.

For a newsroom:

• Verify first. Confirm quotes, numbers, and names before anyone opens a detector. A hallucinated fact is the five-alarm fire; a high score on accurate copy is a curiosity.
• Expect formulaic sections to score high forever. Sports, weather, and earnings briefs are not scandals.

For a university:

• Weigh false positives heavily, especially for English learners and for smooth, methods-style writing.
• Ask about process. Drafts, version history, and a short conversation reveal what a percentage can’t.

In both cases, if you want to improve a passage, add real specifics rather than sanding your prose down to fool the tool. A confirmed number in a news graf or a concrete finding in a discussion section improves the writing and lowers the score at the same time. There’s more on the blog, and if you’re standardizing this across a desk or a department you can see the plans.

When AI is actually there, in each genre

None of this means the genres never hide real AI use. They do. It’s just that the tell is different from a high score, and the real damage differs too.

In news, the genuine problem is fabricated facts, not smooth prose. A model that drafts a market recap will happily invent a plausible-sounding figure, misattribute a quote, or hallucinate a study, and none of that changes the detector score much, because the phrasing was already formulaic. The cautionary tale is on the record: in early 2023 CNET was caught publishing AI-assisted finance explainers that contained real errors, and the outlet ended up correcting a large share of them. A detector caught none of it, because a detector reads style, and the style was fine. Verification caught it, or should have.

In academic work, the analog is fabricated or mismatched citations and a literature review that summarizes sources the author never actually read. A model can produce a fluent, hedged paragraph citing a paper that doesn’t exist or doesn’t say what’s claimed. Again, that fluent paragraph might score no higher than a genuinely researched one, because scholarly prose is smooth either way. What exposes it is checking the citations and asking the author to explain the argument, not the AI percentage.

So the practical lesson holds across both fields: the detector’s score is the least informative part of catching real AI misuse. In news, chase the facts. In academia, chase the sources. The number is a nudge to look; the looking is where the truth lives.

The bottom line

The same AI detector scores news and academic writing differently because it’s grading style, not authorship, and the two genres have opposite styles that both happen to trip the alarm. News reads formulaic; academic prose reads smooth and hedged; both produce the low-variation pattern detectors flag. Neither high score is proof of anything, both fields see documented false positives, and the responsible move in each is to keep a human in the loop and verify what the tool can’t.

Want to see how a draft scores and tighten the parts that read robotic, in either genre? Check how a draft reads and smooth the robotic parts in one place.