Theboxology's Posts
Nairaland Forum › Theboxology's Profile › Theboxology's Posts
1 (of 1 pages)
A history essay and a chemistry lab report walk into the same AI detector. Same tool, same settings, same cutoff. One gets read correctly almost every time. The other is closer to a coin flip. That isn't a hypothetical. It's what a 2026 peer-reviewed study found, and it points to something most writers, editors, and teachers never consider: the format of a piece decides how well a detector can read it, often more than whether AI was used at all. This analysis ranks common content formats from most to least detectable, using published research and tool documentation, and explains why each one lands where it does. Detectable and Flagged Are Two Different Things Before ranking anything, one distinction matters. A format is truly detectable when a tool labels it correctly, catching the AI text and clearing the human text. A format that simply gets high scores may be getting flagged a lot, rightly or wrongly. Mixing these up leads to bad calls. LinkedIn posts, for example, score very high on AI detectors. That tells you detectors react strongly to them. It doesn't tell you the detector is right. The Most Detectable Formats Long-Form Humanities Prose Essays, opinion pieces, and arguments written in full paragraphs sit at the top. In a study of 192 academic texts, Originality scored 0.96 accuracy on humanities writing and Turnitin scored 0.86 (Hadra et al., 2026). Varied vocabulary and personal argument give detectors plenty of signal to separate a human voice from a model's. Length helps, too. Turnitin's own documentation says its model is built for long-form prose sentences, the kind found in essays, dissertations, and articles (Turnitin). Raw, Unedited AI Output Text pasted straight from a chatbot is the easiest case in any format. Its rhythm is even and its word choices are predictable, which is exactly what detectors look for. In Phrasly's side-by-side test, both detectors flagged the raw AI sample correctly before any editing happened ([url]Phrasly, 2026[/url] https://phrasly.ai/reports/common-ai-phrases ). The Formats That Score Highest (but Not Always Fairly) LinkedIn Posts and Social Copy Phrasly analyzed 50 real blogs, LinkedIn posts, theses, and web pages, and found LinkedIn posts scored nearly twice as high as marketing web pages on average. LinkedIn was the only format where the median score hit 100% (Phrasly, 2026). Across the whole sample, 17 of the 50 pieces scored a full 100%. Phrasly's breakdown of What Percentage of Content Is AI-Generated? found that about a third of the pieces showed only a light AI presence while another third showed a heavy one. Short, hook-driven posts follow templates, and templates read as predictable whether a person or a model wrote them. Blog Posts and Web Pages Blog posts ranked alongside LinkedIn among the higher-scoring formats in that study, often tripped up by stock constructions like "not just X, but also Y." Marketing web pages scored roughly half as high as LinkedIn posts on average. Polished sales copy usually goes through several rounds of editing, which may break up the even rhythm detectors are tuned to catch. The Least Detectable Formats Scientific and Technical Writing In the same 2026 study, accuracy collapsed on science texts: Originality fell to 0.58 and Turnitin to 0.51, against 0.96 and 0.86 on humanities texts (Hadra et al., 2026). Technical writing reuses fixed terms and standard phrasing, so human and AI versions look statistically alike. Format inside the format matters here too. A 2026 preprint on math-heavy scientific writing found that pulling text out of a PDF damages the math formatting and lowers detection accuracy, while reading rendered page images worked far better (Huang et al., 2026, preprint). How a document is fed to a detector can change the result. Short Text Under 300 Words Very short pieces give a detector too little to work with. Turnitin raised its minimum from 150 to 300 words of prose after finding accuracy improved with more text (Turnitin, 2023). Product blurbs, short answers, and captions often fall below that line. Non-Prose Formats Turnitin states plainly that its model doesn't reliably detect AI in poetry, scripts, or code, and doesn't detect short or unconventional writing such as bullet points, tables, or annotated bibliographies (Turnitin). A slide deck, a spec sheet, or a listicle built mostly from bullets sits largely outside what the tool was designed to read. Paraphrased and Edited AI Text Rewriting changes the surface a detector reads. The research shows a steep drop, though how steep depends on the rewriter and the detector. A strong research paraphraser cut DetectGPT's accuracy from 70.3% to 4.6% (Krishna et al., 2023), and simple evasion edits pulled six detectors from 39.5% to 17.4% (Perkins et al., 2024). Across 14 tools, text run through QuillBot was classified correctly only 26% of the time (Weber-Wulff et al., 2023). Newer tools have fought back, though. In Phrasly's October 2026 test, one QuillBot pass lowered one detector's score but GPTZero still rated the text 100% AI. The full results are in Can AI Detectors Really Spot Paraphrased AI Text ( https://phrasly.ai/blog/ai-detectors-paraphrased-text ), and the short version is that light rewording blurs the signal without reliably erasing it. A note on the writer behind the format One factor cuts across every row. A Stanford-led study found detectors flagged essays by non-native English writers as AI about 61% of the time on average (Liang et al., 2023). Simpler vocabulary looks low-perplexity in any format, so writer background can tilt scores as much as the format does. What This Means for Writers, Editors, and Educators For writers Know where your format sits. If you write technical or templated content, expect noisier scores and keep your drafts and version history. If a piece gets flagged, check whether the flagged lines are the formulaic ones, like openings, closings, and stock transitions. For editors and educators Weigh a score by format. A high score on a humanities essay deserves a closer look. A high score on a lab report or a LinkedIn post deserves more caution. And a low score on bullets, code, or edited text proves very little, because the tool may barely be reading it. Running a sentence-level check with Phrasly's AI Detector shows which lines triggered a flag, which tells you far more than the headline percentage. Conclusion: Read the Format Before You Read the Score AI detectors are most reliable on long, varied prose and raw chatbot output. They struggle with scientific writing, short text, non-prose layouts, and anything that has been rewritten. In between sit formats like LinkedIn posts, which get flagged constantly for reasons that have as much to do with template writing as with AI. So the format comes first. Know what kind of text you're looking at, how much a detector can actually read in it, and how the research says tools perform on it. Then the number on the screen starts to mean something. |
1 (of 1 pages)