Sometimes you do not have a detector handy. You are reading a colleague's email at 9pm, a vendor's draft on your phone, a student's homework on a printout. You need a fast read on whether a real person typed this. This guide is the manual version. Six specific patterns that show up in ChatGPT, Claude, and Gemini output, a five-step workflow that pairs the eye with a quick TextSight scan, the ESL caveat that catches honest readers off guard, and the cases where manual spotting collapses. A reader who learns the six patterns gets to roughly 70 to 80 percent confidence on long-form text in under a minute, and the tool step closes the gap to a defensible verdict.
AI detectors are easy to reach for, but they are not always where you need them. A manager scanning Slack cannot paste 40 short replies into a scanner every morning. A teacher with a stack of printed assignments wants a first-pass read before deciding which papers to type into a tool. The eye is the screen; the tool is the verdict.
The trained eye runs in seconds and works offline. On a printed page, on a phone with no signal, or on the third Slack reply of a busy morning, you do not have time to copy text into a detector. A reader who knows the six patterns can flag a suspicious passage in under a minute, then decide whether the case is worth a full tool verification. Reaching for a scanner on every paragraph is a workflow that nobody actually keeps up.
A 78 percent score tells you the model is suspicious. The six patterns tell you the third paragraph hedges three times in a row and ends with In summary, which is what the tool was actually reacting to. Knowing the reason behind the signal makes you a better reader the next time and gives you specific evidence to cite in a conversation. The conversation lands harder when you can point to a sentence than when you wave a percentage.
Roughly 70 to 80 percent accuracy on long-form text of 600 words or more, dropping fast on short passages and on text that has been edited by a human after the AI draft. That is not enough to fail a student or fire a writer on. It is plenty for noticing, asking, and deciding whether to confirm with a tool. The five-step workflow below is built around that honest ceiling rather than against it.
Any single pattern can show up in genuine writing. The signal is when two or three cluster in the same paragraph. Three or more alongside a tool flag is usually conclusive, and that is the workflow the next section walks through.
"A robust, comprehensive, multifaceted approach" reads AI. "A robust approach" reads human, and so does any sentence where the noun does the work without the stack. Two or three tripled-adjective constructions per page is normal on a ChatGPT-assisted draft and almost never appears in unassisted writing. Watch for the stack at the start of important sentences, where AI tends to load up the modifier before getting to the point.
Frontier models have favourite words. The reliable tells in 2026 are delve, tapestry, navigate used metaphorically, robust, leverage as a verb, underscore, showcase, myriad, multifaceted, and foster. Two or three of these in a 500-word passage is statistically unusual for natural writing. Five or more is a near-certainty. Most undergraduate writers and most working journalists use zero or one of these in a typical piece. The cluster, not any single word, is the signal.
Watch for stacked transitions across paragraph boundaries: Furthermore, Moreover, In addition, Additionally, Notably. ChatGPT defaults to these at the start of body paragraphs the way human writers rarely do; humans usually trust the paragraph break itself to do the work. Five paragraphs in a row opening with a generic transition is templated structure, not a stylistic choice.
Count the words in five consecutive sentences. If every sentence lands between 16 and 22 words, the burstiness signal is low and the passage reads AI even when the vocabulary looks clean. Human writers vary length deliberately. A 30-word subordinate clause followed by a five-word punchline is the cleanest signal a piece was written rather than generated. The eye picks this up as a flatness in cadence; the detector measures it as low burstiness.
"In conclusion," "In summary," "To wrap up," "Overall." AI almost always signposts the ending and restates the thesis on the way out. Most experienced writers either drop a punchline, ask a question, or simply stop. The signposted summary is a tell, especially in informal contexts like Slack messages, dating-app replies, and casual emails where no human writes "In conclusion, please let me know if you have any questions."
Frontier models inherit a politeness layer from chatbot tuning. Watch for openings like "It is worth noting that," "It is important to consider," "While X, it is also true that Y," and balanced "on one hand, on the other hand" constructions. Humans hedge too, but they vary the phrasing. AI uses the same three or four hedges across an entire piece. Uniform hedging plus a polite-assistant opener on a Slack reply or email body is one of the cleanest signals when present.
Roughly four minutes per text once you know the workflow. The eye does steps 1 to 3, the tool does step 4, and step 5 brings the two together. Running steps out of order leads back to single-signal verdicts and the false positives that come with them.
Read the first two paragraphs aloud in your head. Listen for sentence-length variance and rhythm. AI prose tends to lock into a uniform 16-to-22 word band, while human writers mix short punchy sentences with longer extended ones. The cadence is the cheapest signal and the hardest one to fake, because a writer rewriting an AI draft usually edits the vocabulary first and forgets about rhythm. A flat cadence on the first read is the first finger on the scale.
Scan for the AI vocabulary cluster. Mark any instance of delve, tapestry, navigate, robust, leverage as a verb, underscore, showcase, multifaceted, or foster. Two or three of these in a 500-word passage is statistically unusual. Five or more is a near-certainty. Most natural writers use one or zero of these words in a typical piece. The cluster is what flags the passage, not any single word.
Count paragraph openers. Furthermore, Moreover, In addition, Additionally, Notably, In conclusion. AI stacks these at paragraph boundaries the way human writers rarely do. Five paragraphs in a row opening with a generic transition is templated structure, not stylistic choice. If three of the six patterns from the previous section have lit up by this point, you have enough to take the case to the tool.
Paste the passage into TextSight. Free for three scans a day with no signup. Read the overall 0 to 100 score, then read the per-sentence highlight map. Clustered red sentences in one paragraph mean something different from scattered red sentences across the piece. The headline percentage is the summary; the highlights are the case. The tool also returns perplexity and burstiness numbers, which give the cadence read from step 1 a numerical backing.
Go back to the text and check it against all six telltale patterns from the previous section. Mark each one that appears. If three or more patterns appear alongside a tool flag above 60 percent, the verdict is defensible and you have specific evidence to cite. If the tool flagged but only one pattern appears, the score is real but the case is weaker; consider a second independent classifier or a conversation with the author before acting. Methodology means running the eye and the tool against each other rather than relying on either alone.
Free includes 3 detector scans a day and a 1,500-word AI rewriter quota. Paid tiers raise the quotas and add the Chrome extension, file upload, and REST API. Yearly billing saves 25%.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing
The two methods are complementary, not redundant. The eye is fast and free; the tool is precise and defensible. Knowing which one to lead with on a given piece of text saves time and produces better verdicts.
Short Slack threads, casual emails, dating-app messages, social-media replies, and printed assignments where copy-pasting into a scanner is friction-heavy. On these the eye runs in seconds and gets to a usable confidence number fast. The polite-assistant opener and the In conclusion closer are especially loud in informal contexts where no human writer would use them. A pattern like "Furthermore, I wanted to follow up" in a casual reply is more obvious by eye than by tool because the tool weighs structural signals that are too thin to measure on three sentences.
Long-form essays, polished cover letters, vendor deliverables of 600 words or more, and any piece where the writer may have edited an AI draft after generating. The tool measures perplexity and burstiness the eye cannot estimate precisely, returns sentence-level highlights that survive the conversation that follows, and produces a defensible number you can quote in a grade review or a contract dispute. For high-stakes calls, never act on the eye alone; the tool is the verdict.
The hardest case is when the eye says AI and the tool says human, or vice versa. The eye-says-AI, tool-says-human case usually happens on heavily polished AI drafts where the writer swapped the signature vocabulary but kept the rhythm. The tool-says-AI, eye-says-human case usually happens on ESL writing or on highly structured genres like legal memos. When the two methods disagree, treat the result as inconclusive, run a second independent classifier if you can, and lean on a conversation about process before drawing a conclusion.
This is the single most important caveat on manual AI spotting. Get it wrong and the workflow produces unjust outcomes regardless of how confident the eye feels. Read this section before applying the six patterns to any classroom, hiring queue, or editorial pipeline with non-native English writers.
English-as-a-second-language writers often produce more uniform sentence shapes, a narrower active vocabulary, more formal register, and standard hedging because they learned English in formal classrooms. Those features overlap structurally with patterns 4 and 6 from the six telltale patterns, and sometimes with pattern 3 as well. The writer is not doing anything wrong; the eye is correctly measuring something that happens to mean a different thing in ESL prose than in native prose.
If you know the writer is ESL, weight your manual flags down by roughly half. Skip patterns 4 and 6 entirely on first read and rely on patterns 1, 2, 5, and the tool step. The vocabulary cluster from pattern 2 is the most language-neutral signal on the list because the AI words are statistically rare across all English varieties, while uniform rhythm and polite-assistant register can come from a confident ESL writer just as easily as from ChatGPT. The tool step matters more here than on native prose.
For graded or hiring decisions, never act on manual spotting alone with an ESL writer. Run the text through a calibrated detector like TextSight, which tunes its threshold roughly 40 percent lower for ESL prose than open-source baselines. Look for clustered residual highlights rather than the headline number. Bring the per-sentence evidence into a conversation about process before drawing a conclusion. The methodology is the same as for native writers; the threshold is what shifts.
Sometimes you are the writer rather than the checker. A polished first draft that looks too clean is a problem in academic, freelance, and hiring contexts. The TextSight AI rewriter rewrites AI-flagged passages in three modes designed for different reading audiences.
The default. Rewrites sentence rhythm, removes the signature vocabulary cluster, breaks up tripled adjectives, and varies transitions across paragraph boundaries. The output reads as a careful human draft and lands in the 60 to 80 band on most detectors. Use it for blog posts, marketing pieces, and any deliverable where the audience is general and the priority is detection-clean prose without losing the original argument.
More aggressive rewrites with deliberate burstiness and more colloquial vocabulary. The output reads as a confident human writer with a clear voice and lands above 80 on most detectors. Use it for academic essays, hiring materials, and any context where the audience may run a second tool. The trade-off is a slight loss of the original phrasing; the output preserves the argument and the structure but feels written rather than tightened.
Minimal edits that keep the original prose largely intact while removing the loudest AI tells. The output reads as a polished human draft and lands in the 40 to 60 band on most detectors. Use it when the original text is mostly human with AI-assisted polish and you want to remove the residual AI fingerprint without rewriting the piece. Light mode is the right pick for editing collaborative drafts or salvaging your own AI-assisted notes.
The procedural sister guide. A 5-step verification method built around two free detectors and the highlight map.
Read the procedural guideThe methodology-led companion. Classifier families, signals, workflow, and three confidence tiers.
Read the methodology guideRun the methodology on a real scan. Sentence-level highlights, calibrated overall score, bundled Plagiarism Risk.
Open the detectorFree check for the signature vocabulary cluster: delve, tapestry, navigate, robust, leverage as a verb.
Open the toolFree to try, no card. 3 detector scans a day, sentence-level highlights, perplexity and burstiness signals on every result.