How to detect AI-written text
last updated 7 june 2026
To detect AI-written text, look for low burstiness (sentences of uniform length and rhythm), hedging filler like "it's important to note," an absence of specific detail or lived experience, and tell-tale vocabulary such as "delve," "tapestry" and "testament to." Then confirm with a detector.
Large language models like ChatGPT, Claude and Gemini write fluent, grammatically perfect prose — which is itself the giveaway. Human writing is bursty: it mixes long and short sentences, takes detours, and includes concrete, sometimes messy specifics. AI writing tends toward an even, hedged, slightly generic middle.
verifai scans the meaningful text blocks on a page and scores each one, tinting suspected AI passages in place so you can see exactly which paragraphs triggered the signal — without copying anything into a separate tool.
Read for AI in four passes
- 1
Check the rhythm (burstiness)
Read a few sentences aloud. Human writing varies sentence length sharply; AI writing often holds a steady 15–25 words per sentence with similar structure. Low variance is a strong signal.
- 2
Hunt for hedging and filler
Phrases like "it's important to note," "in today's fast-paced world," "plays a crucial role," and "a testament to" are statistically over-represented in AI text. So are the words "delve," "tapestry," "navigating," and "underscores."
- 3
Look for missing specificity
AI text is often confident but vague — no names, dates, numbers, or first-hand detail. Genuine expertise leaves fingerprints: a specific tool version, a real anecdote, an opinionated aside.
- 4
Confirm with a detector
Run verifai on the page. Each text block gets a 0–100 score and a confidence level; suspected passages are tinted so you can review them in context rather than as a wall of pasted text.
The linguistic tells, ranked
These signals are strongest when several appear together:
- uniform sentence length and structure (low burstiness)
- hedging and filler transitions: moreover, furthermore, in conclusion
- over-used vocabulary: delve, tapestry, testament, underscores, realm
- balanced "on one hand / on the other" framing with no real stance
- no concrete detail — names, numbers, dates, or lived experience
Why false positives happen
Careful non-native writing, heavily edited corporate copy, and formal academic prose can all read as "AI-like" and trip a detector. This is why responsible detection reports a confidence level rather than a verdict, and why a human should always make the final call — especially in high-stakes contexts like academic integrity.
frequently asked
Can teachers rely on AI text detectors?
Detectors are a signal, not proof. False positives are real, particularly for non-native English writers, so detectors should inform a conversation rather than decide an outcome on their own.
What words give away ChatGPT?
No single word is conclusive, but "delve," "tapestry," "testament to," "navigating the," and stock transitions like "it's important to note" are statistically over-represented in AI-written text.