How to Detect AI Writing in a Single Paragraph: Read It, Mark It, Then Weigh Any Score

Someone hands you a paragraph, or you wrote one with help from a chatbot, and the question is the same: was this machine-written? Learning how to detect AI writing at this size means accepting a limit early. A single paragraph carries little evidence, so the reading you do yourself matters more than any percentage. Below: the reading checks to mark first, the minimum lengths vendors publish, what studies found about detector errors, and a worked example.

How to detect AI writing by reading the paragraph first

Read the paragraph before any tool reads it for you.

A reader has one advantage no detector has: context. You know the assignment, the writer and the topic. Read the paragraph once, then again with a highlighter, looking for the habits below. The fuller list, with examples, is in the guide to highlighting sentences that read as machine-written, and the stock vocabulary has its own page on words and phrases worth marking in a draft.

• Claims with no owner. Sentences that could sit in any essay on the topic: "This plays a crucial role in many areas." Ask what, where and according to whom.
• Even rhythm. Three sentences of the same length and shape in a row, each ending on a tidy summary.
• Missing specifics. No names, dates, numbers, places or quotations, where a person who did the reading would have them.
• A closing moral. A last sentence that restates the paragraph as a lesson.
• Sources that do not check out. A citation, title or quotation you cannot find. This is the strongest sign on the list, because it is a fact, not a style.

Count the marks, but do not score them. Two highlights in eight sentences mean little. Highlights on nearly every line, plus a source that does not exist, mean it is time to ask how the paragraph was written.

Why one paragraph is hard for any detector

Detectors look for statistical patterns across many words, and a paragraph does not give them many. The vendors say so in their own documentation. In a May 2023 update, Turnitin's chief product officer wrote that the company had raised its minimum from 150 to 300 words for a document to be evaluated, because accuracy increases with a little more text. The same update reported a higher incidence of false positives in the first few and last few sentences of a document, which is exactly where a lone paragraph sits.

Turnitin also gave two rates in that update, and they measure different things. For documents with over 20% AI writing, the document false positive rate was less than 1%, checked against 800,000 academic papers written before ChatGPT's release. At the sentence level, the false positive rate was about 4%: roughly 4 of every 100 highlighted sentences may be human-written. When you check a single paragraph, the sentence figure is the one that applies.

GPTZero sets a lower floor. Its support guide for the Microsoft Word add-in says a selection needs to be at least 250 characters, about two average sentences, before it returns a result. A floor is where a tool will answer, not where its answer becomes reliable.

What independent studies found about detector errors

Two studies are worth knowing before you trust a score on any length of text. In Patterns (2023), Liang and colleagues at Stanford ran seven widely used detectors on 91 TOEFL essays by non-native English writers and 88 essays by US eighth graders. The detectors classified the eighth-grade essays accurately but labeled more than half of the TOEFL essays as AI-generated, an average false positive rate of 61.3%. The cause was low perplexity: plainer vocabulary is more predictable, and many detectors measure predictability.

In the International Journal for Educational Integrity (2023), Weber-Wulff and colleagues tested 14 tools, including Turnitin, on 54 documents with known origins: human-written, machine-translated, AI-generated, and AI-generated then edited or paraphrased. Every tool scored below 80% accuracy and only five scored above 70%. For AI text run through a paraphrasing tool, overall accuracy fell to 26%. Their conclusion was that the tools are neither accurate nor reliable. More findings, and how detectors work, are on the page on what AI text detectors can and cannot tell you.

A worked example: one paragraph, marked

Here is an example paragraph written for this page, the opening of a short response on urban trees:

Reading it with a highlighter gives these marks:

• Generic claims: "play a crucial role," "multifaceted benefits" and "our shared future" could open any essay on parks, rivers or libraries.
• A list of three: shade, air quality, wildlife, with no city, species or figure attached.
• Even rhythm: five sentences of similar length, two opened with a connector (Moreover, Ultimately).
• A closing moral: the last sentence restates the paragraph as a lesson.
• Missing specifics: no source, place or number anywhere.

The paragraph is 62 words long, far below Turnitin's 300-word floor, so a Turnitin score is not available and a score from any tool would rest on very little text. The marks do not show who wrote it either; a tired student at midnight could produce every one. They show where to ask: which city, which study, which trees? A writer who did the reading can answer at once.

Steps to check a paragraph for AI writing

1. Read it once, clean

   Read it through without marking anything, so your first impression comes from meaning.

2. Mark the reading checks

   On the second pass, highlight generic claims, stock phrases, runs of even sentences, lists of three and missing specifics. Use one color per habit so the pattern is visible at a glance.

3. Check every source

   Open each citation, quotation and figure. A reference that does not exist is the most concrete finding you can make. Tracing a quotation back to its source shows the method.

4. Check the length before any tool

   Count the words. Below 300, Turnitin will not score it; below 250 characters, GPTZero will not either. Above those floors, a short text still gives a detector little to work with.

5. Run a detector only as a second opinion

   Note the tool, the date, the score and the length. Agreement between your marks and the score is a reason to talk, not a finding.

6. Ask the writer

   Ask about the specifics the paragraph lacks, and for notes or drafts. Answers settle more than scores do.

If you want a second opinion after the reading, one option is the AI detector built into Walter Writes AI, an AI humanizer that supports over 80 languages and includes a detector on every plan. Like the tools in the studies above, it returns a probability, not a finding about who typed the words, and two detectors can disagree on the same paragraph.

Pitfalls when checking a single paragraph

• Treating one tell as proof. A single stock phrase or connector shows a habit, and humans have the same habits.
• Forgetting the writer. Liang and colleagues found plain, predictable prose from non-native writers flagged most. Simple English is not evidence.

Tips from editors who check drafts

• Compare like with like. Read the paragraph beside earlier work by the same writer. A sudden change in voice is more telling than any single phrase.
• Keep your marks. A marked copy shows what you saw and why, which matters if anyone asks later. The routine for fair markup when you suspect AI writing builds on it.

If the paragraph is for a class, the score that matters is usually the one your school's tool shows: how Turnitin's AI indicator works explains what instructors see, what an AI percentage actually measures covers the number itself, and the ranking of AI detectors for teachers compares the tools on documented limits.

Checking a paragraph for AI writing is mostly reading: mark the habits, open the sources, compare the voice and ask the writer what the text leaves out. Detectors can add a second opinion on longer passages, but most paragraphs miss the floors their own vendors set. Start at the AI writing guides for the full set of methods, and keep your marked copy, because the marks are the part of the check anyone can verify. And if the paragraph turns out to be yours but reads flat, the humanizer ranking and the rewrite method beside it are the next stop.