AI detectors work best on long, continuous prose like essays and articles, and least reliably on short, formulaic or highly structured text such as bullet lists, code, poems, forms and translated passages. The reason is simple: detectors estimate how "predictable" writing is, and some kinds of writing are predictable by design no matter who wrote them. Knowing which category your text falls into tells you how much weight a score deserves.
Why content type changes the result
Most AI text detectors compare a passage against patterns learned from large amounts of human and machine writing. They look at signals like how expected each word is, how much sentence length and structure vary, and how phrasing is distributed across the text.
Those signals behave differently depending on genre. A personal essay naturally wanders and varies. A safety procedure should not. A detector that does not account for that difference may read "consistent and clear" as "machine-made." For a deeper look at the mechanics, see our post on false positives.
Reliability at a glance
| Content type | How well detectors usually cope | Main caution |
|---|---|---|
| Long essays and articles | Best case | Heavily edited mixed drafts still blur results |
| Short answers and discussion posts | Weak | Too few sentences to judge |
| Emails and business messages | Mixed | Templates and polite formulas look generic |
| Technical and scientific writing | Mixed to weak | Standard phrasing and passive voice are common in human work |
| Lists, outlines, slide text | Weak | Fragments lack natural variation |
| Creative writing and poetry | Unpredictable | Unusual structure confuses models trained on prose |
| Code | Poor fit | Prose detectors are not built for programming languages |
| Translated or second-language text | Weak | Translation tools and careful learners both produce plain phrasing |
These are general patterns, not measured rates. Individual tools differ, so check what a detector says about its own supported formats and minimum length.
Essays and long-form articles
This is the home turf of most detectors. A few hundred words of continuous writing gives the model plenty of sentences to compare, and essays usually have a single author voice.
Where it gets harder: essays built from a strict template (five-paragraph structure, stock transitions, formula conclusions) and drafts where someone wrote half and AI helped polish the other half. Sentence-level highlighting helps here, because you can see which parts drove the score instead of reacting to one number.
Tip: if you are checking your own essay, scan the whole thing rather than a single paragraph. A paragraph alone is closer to a short answer.
Short answers, discussion posts and captions
Anything under a few sentences is hard to judge. "Photosynthesis converts light energy into chemical energy stored in glucose" is a correct answer that thousands of humans and every chatbot would phrase almost identically.
Some tools refuse to score text below a minimum length for exactly this reason. If yours does not, treat short-text scores as close to meaningless.
Emails, cover letters and business writing
Professional writing relies on conventions: polite openers, standard closings, familiar phrases like "please find attached." Those conventions are predictable for everyone. A templated email from a human can score as AI, and a lightly edited AI draft can score as human.
In workplaces, a detector is rarely the right tool anyway. What matters is whether the message is accurate, appropriate and reviewed by the person sending it.
Technical, legal and scientific writing
Lab reports, methods sections, contracts and documentation are deliberately standardized. Writers are taught to use precise, repeated terminology and neutral structure. That is good writing for the genre, and it also happens to resemble model output.
If you teach or review this kind of work, rely more on process evidence: lab notebooks, data files, drafts and the writer's ability to explain choices.
AI DetectorSentence-level AI scoreLists, outlines and slides
Bullet points strip out connective tissue. There are few full sentences, little variation and lots of noun phrases. Detectors trained on paragraphs struggle to interpret them. If you need to check slide content or notes, convert key points into a paragraph first or, better, evaluate the paragraph-based work that came from them.
Creative writing and poetry
Creative text can swing either way. Experimental structure, fragments and invented words may look unlike anything in the training data, which some detectors read as human and others as anomalous. Rhyming, highly regular verse can read as predictable. Results here are hard to interpret, and many tools do not claim reliability for poetry at all.
Code
General AI text detectors are built for natural language. Programming languages have strict syntax, repeated patterns and conventional names, so a prose detector has little useful to say. In programming courses, better signals include commit history, incremental progress, code reviews and asking the student to walk through their logic.
Translated and second-language writing
Two separate issues overlap here:
- Machine-translated text is generated by a model, so it can carry machine-like patterns even when a human wrote the original.
- Writers working in a second language often choose safe, common phrasing and simpler sentences. That is a sensible strategy for clarity, but it can raise AI scores.
This is one of the main fairness concerns with AI detection. If you are a multilingual writer, keep your drafts and notes, and if you used translation tools, check whether your course allows them and disclose them if required.
How to get the most reliable result from any detector
- Use enough text. Several paragraphs of continuous prose beats a single paragraph.
- Remove non-prose parts. Strip out reference lists, tables, block quotes and code before scanning, since they distort the analysis.
- Check one author's work at a time. Group projects with several writers produce mixed signals.
- Read the sentence-level view. Look at which parts drove the score and whether they are simply formulaic.
- Compare with process evidence. Drafts, version history and notes often settle the question faster than any score.
- Never treat the number as a verdict. It is a probability estimate for a conversation, not a conclusion.
A worked example: one student, three kinds of text
Imagine a biology student submits three things in the same week: a 1,500-word research essay, a 120-word discussion post, and a lab report written from the department template.
The essay is the only one where a detector has a fair chance. It is long, continuous and written in the student's own voice, with examples from lectures and specific references. If it comes back with a few highlighted sentences, those are worth a look, but they are likely the most generic lines, such as a textbook definition.
The discussion post is too short to say anything meaningful. Even if it scores high, the honest reading is "not enough text."
The lab report follows a fixed structure with standard phrases in the methods section, because that is what the template asks for. A high score there says more about the template than about the student. The lab notebook and raw data are far better evidence of who did the work.
Same student, same week, three very different levels of reliability. That is why the type of text should shape how much attention any score receives.
Choosing the right check for the job
Sometimes the question is not really "is this AI?" at all.
- Worried about copied material? That is a plagiarism question. See are plagiarism checkers accurate.
- Worried about wrong facts? That is a fact-checking question, and it applies to human and AI writing alike. See AI hallucination.
- Worried your own writing sounds stiff? That is a style question. Reading it aloud and revising is often more useful than any score.
Key takeaways
- Detectors are most informative on long, continuous prose by a single author.
- Short text, lists, templates, technical writing, poetry, code and translated text all reduce reliability.
- Prepare text properly before scanning and always read results in context.
- When a score matters, process evidence such as drafts and version history should carry more weight than the score itself.