You spot AI writing by reading for two things at once: the surface tells (flat rhythm, hedging phrases, suspiciously tidy structure) and the substance tells (vague or slightly-off facts, no lived detail, nothing that could only come from the writer's own experience). No single sign proves it, and a detector tool backs up your read — it shouldn't replace it.
Short answer: Read for repetitive sentence rhythm, stock transition phrases ("moreover," "in today's world"), overly balanced lists, and a lack of specific, checkable detail. Run the text through one AI detector like GPTZero or Originality.ai as a second opinion, not a verdict — in my testing, every detector I've used produces false positives, especially on writing by non-native English speakers.
I run AI-generated and human-written paragraphs through the same detectors most weeks as part of testing writing tools for this site, and the honest finding is that no tool gets it right every time. What actually works is a two-step read: scan the text yourself for the patterns AI models fall back on by default, then use a detector as a tiebreaker rather than a final answer. Below is the exact process, including where I've watched detectors get it wrong.
What you'll need
You don't need special software to start — your own reading is the first and most reliable tool. Open the text somewhere you can read it slowly, ideally printed or in a distraction-free view, since skimming is how AI tells slip past you. If you want a second opinion, have access to at least one free AI detector; GPTZero and QuillBot's detector both work without an account for short samples. If the writing is a student essay or something with a known human draft history, having the writer's earlier drafts or a Google Docs version history on hand is the single best evidence you can get — better than any detector score.
Step-by-step: how to spot AI writing
1. Read for sentence rhythm first
AI-generated text tends to default to similar sentence lengths in a row — three medium sentences, then a short one for "punch." Real human writing is messier: a long, winding sentence next to a three-word fragment, on purpose or not. In my testing, this rhythm tell is often more reliable than word choice, because paraphrasing tools change vocabulary but rarely fix the underlying cadence.
2. Flag the stock phrases
Certain phrasing shows up disproportionately in AI output: generic urgency framing about how fast a field is moving, a hedge inserted right before a claim ("worth noting that…"), false-choice openers ("no matter your experience level"), and inflated verbs used for ordinary actions — a tool doesn't just help, it supposedly transforms or supercharges everything. One or two of these isn't proof by itself, but a cluster of them packed into a short piece is a strong signal.
3. Check for real, checkable detail
Human writing usually contains something a model couldn't invent: a specific date, a named person, an exact number tied to a real event, an opinion with a reason attached. AI writing tends to stay generically correct and confidently vague. When I tested this against ten student essays with known authorship, the AI-written ones scored high on "sounds right" and low on "checks out."
4. Look at structure for over-tidiness
A five-paragraph essay with exactly three evenly-weighted points per section, each intro restating the thesis, is a pattern models default to unless told otherwise. Real writers usually let one point run long because it mattered more to them.
5. Run it through a detector as a second opinion
Paste the text into GPTZero, Originality.ai, or QuillBot's detector. Treat the score as one data point, not a verdict — in my testing, all three occasionally flagged human-written text, and the false-positive rate climbs sharply for non-native English writers. A Stanford study (Liang et al., 2023) found these detectors misclassified more than 60% of essays by non-native English speakers as AI-written, against roughly 5% of essays by native speakers — a bias gap wide enough that no single score should end the conversation.
6. Weigh evidence, don't chase certainty
If you're evaluating this for something with real stakes — grading, hiring, publishing — combine two or three signals (your own read, a detector score, and where possible a draft history) before acting. No single signal, including a detector score in the high 90s, should be the sole basis for an accusation.
Example prompts you can copy
Use these to get a second AI's read on a passage, or to stress-test your own writing before you publish it.
- Ask a model to critique for AI tells: "Read this paragraph and list any phrases, sentence patterns, or structural habits that sound like typical AI-generated text: [paste]."
- Check for generic vs. specific detail: "Which sentences in this text contain a specific, checkable fact or detail, and which are generically true of almost any topic? [paste]"
- Rhythm check: "Describe the sentence-length pattern in this paragraph — is it varied or repetitive? [paste]"
- Self-check before publishing your own draft: "Rewrite any sentence in this that sounds templated or AI-generated, and tell me which ones you changed: [paste]"
Common mistakes to avoid
The biggest mistake is treating a detector score as a verdict instead of a data point — I've seen a 98% "AI-generated" score turn out to be a human writer with a plain, repetitive style, and the Stanford research above backs up why that happens. The second is judging on a single paragraph; short samples give every detector less to work with and push false positive and false negative rates up. The third is ignoring context — a rushed, low-stakes email is naturally more generic than a personal essay, and that's not evidence of anything. The fourth is skipping your own read entirely and outsourcing the whole judgment to software; in my testing, my own eyes caught tells that GPTZero and Originality.ai both missed, and vice versa, which is exactly why you want both.
Tools that make this easier
Once you've read the text yourself, a detector is useful for a quick second check. Here's how the three I use most often compare on price and fit.
| Tool | Best for | Entry price | Free tier |
|---|---|---|---|
| GPTZero | Teachers and editors checking essays or articles | $14.99/month (Essential) | Yes — 10,000 words |
| Originality.ai | Agencies and publishers screening content at scale | $14.95/month (Pro, monthly) | No permanent free plan — pay-per-scan credits |
| QuillBot AI Detector | Quick one-off checks alongside paraphrasing | $8.33/month (Premium, billed annually) | Yes — 1,200 words/scan, 6 scans/day |
I checked Originality.ai’s pricing page and QuillBot’s premium pricing page this week before writing this — confirm the current number before you buy, since these move. If you're regularly rewriting your own AI-assisted drafts rather than just checking other people's, my best AI tool for paraphrasing guide covers the tools that clean up AI phrasing before you publish. And if you want the fuller picture on how any AI writing tool actually performs before you trust its output, see how I test tools in AI tool reviews and how to read vendor claims in AI tool ratings.
My take
Reading for the pattern yourself catches more than any single detector, and combining your read with a detector score catches more than either alone. I wouldn't act on a detector score by itself for anything with real consequences — the false-positive rate for non-native English writers alone is reason enough to treat these tools as a second opinion, not a judge. If you're on the other side of this — using an AI writing assistant to draft and want your final piece to read like you, not like a template — the fix is the same rhythm and specificity checks in steps 1 through 3 above, run on your own work before you hit publish. And if this question is coming up because of a student essay, my notes on using ChatGPT to write an essay without plagiarizing and on a professor's invisible prompt trap that caught cheating students are worth reading alongside this one — they cover the academic-integrity side that a detector score alone won't settle.
Frequently Asked Questions
Is it free to check if writing is AI-generated?
Yes, for casual checks. GPTZero's free tier covers 10,000 words and QuillBot's detector allows 1,200 words per scan, six scans a day, with no account needed for short samples. Paid plans mainly add higher word limits, batch uploads, and team features.
How long does it take to spot AI writing?
A few minutes for a careful read of a short piece — you're scanning for sentence rhythm, stock phrases, and specific detail, not analyzing every word. Running it through a detector adds under a minute. The slower part is deciding what to do with a borderline result, which shouldn't be rushed.
What is the easiest way to check for AI writing?
Read it once for rhythm and specificity, then paste it into a free detector like GPTZero as a second opinion. Neither step alone is reliable enough on its own — in my testing, the combination catches far more than either one used by itself.
Can AI detectors be wrong?
Yes, regularly. A Stanford study found detectors misclassified over 60% of essays by non-native English speakers as AI-written, and OpenAI discontinued its own AI text classifier in 2023 over accuracy concerns. Treat any single score as a data point, not proof.
Can I tell if my own writing sounds too AI-generated?
Yes — run it through the same checks: look for repetitive sentence length, stock transition phrases, and paragraphs that stay generic instead of specific. If you drafted with an AI assistant, editing in a few specific, checkable details usually does more to fix this than rewording sentences.