When a colleague forwards something they “wrote,” I’ve learned to ask one question first: how do I tell if this text is AI generated? The pattern recognition comes quickly now — the prose is too even, transitions are too smooth, and not one example feels like it came from someone who actually lived through it. The most reliable method is to combine manual pattern-spotting with a free detection tool, because neither works well enough on its own.
AI writing has gotten better at mimicking human style, but consistent telltale patterns still appear. Free detectors have improved alongside it. Here’s what I look for, which tools I use, and what to do when results are inconclusive.
Quick Answer
Look for hedging openers (“It’s worth noting that”), flat uniform structure, and zero concrete personal detail. Run the passage through a free detector like GPTZero or Copyleaks — paste at least 150 words for a reliable score. When both the manual read and the tool point the same direction, you have a solid working signal.
What Are the Telltale Signs of AI-Generated Text?
Three patterns appear together in most AI-written content. Any one can appear in human writing too; the combination is what matters.
Hedging Language Everywhere
Watch for openers like “It’s worth noting,” “In today’s world,” or “It is important to understand.” AI models are trained to avoid overconfidence, so they qualify constantly. Human writers make direct claims and own them.
Uniform Sentence Rhythm
AI prose follows a predictable beat: intro sentence, three supporting points, summary. Paragraph lengths are similar, transitions are stock phrases (“Furthermore,” “In conclusion”), and sentence length barely varies. Real writers break that rhythm without thinking.
No Concrete Specific Detail
This is my most reliable single check. A human who used a tool for six months names the exact error message they saw or the week a feature changed. AI offers generic examples that apply to anyone — because they’re designed to. I look for one detail I couldn’t have invented; AI rarely delivers one.
Hedging phrases, uniform rhythm, and absent specifics are the three most consistent manual signals that text came from a language model rather than a person.
How Do AI Detection Tools Work?
Free detectors measure two statistical properties. Perplexity checks how predictable each word choice is — AI tends to pick the most expected next word, while humans make surprising choices. Burstiness measures sentence-length variation — human writing mixes short and long lines; AI clusters near an average.
Here’s how I run a check:
- Open GPTZero or Copyleaks — both are free and require no account for a paste-and-check.
- Paste a full passage. Under 100 words returns unreliable results; use at least a full paragraph.
- Read the confidence percentage, not just the overall label. Under 70% confidence means the result is inconclusive.
- Check sentence-level highlighting. Most tools flag individual high-probability sentences — that’s more useful than a document-level score.
Pro tip: Run the same passage through two tools. If both flag it at 80%+ AI probability, that’s a meaningful signal. A split result means you should treat the text as uncertain.
AI detectors measure word predictability and sentence-length variation — the statistical fingerprints that separate how humans and language models construct sentences.
Which Free AI Detector Should You Use?
All three tools below have free tiers that work for basic paste-and-check with no signup required.
| Tool | Free Tier | Sentence Highlights | Best For |
|---|---|---|---|
| GPTZero | 10,000 chars/check | Yes | General text, education |
| Copyleaks | Unlimited paste checks | Yes | Business and academic docs |
| Originality.ai | Trial credits only | Yes | Publishers and agencies |
I start with GPTZero for anything under 1,500 words. Copyleaks handles longer documents better on its free tier and tends to be slightly more sensitive to lightly edited AI text in my experience.
GPTZero and Copyleaks are the strongest free starting points — both highlight results at the sentence level, which is far more informative than a single document-level verdict.
Does Editing AI Text Fool the Detectors?
Yes, often completely. Adding personal anecdotes, varying sentence length, and replacing hedging phrases with direct claims all reduce scores noticeably. I’ve seen AI text drop from 92% AI probability to under 40% after about ten minutes of targeted edits — that’s the biggest limitation of any detector.
What that means practically: a low detection score does not confirm text is human-written. It may be edited AI output. When you can’t verify the author’s process, treat a “probably human” result on polished text as uncertain rather than cleared.
Troubleshooting tip: If a detector returns inconsistent scores across multiple pastes of the same text, check that you’re submitting the full passage. Most detectors need at least 150 words for a reliable reading; shorter excerpts produce noise.
Understanding why AI output is statistically predictable becomes clearer once you know how AI tokens and context windows shape every response a model generates.
Edited AI text reliably fools detectors — which is why reading for concrete personal detail and genuine voice remains an essential second check that no tool can replace.
What Common Mistakes Undermine AI Detection?
- Using only one tool. Detectors disagree. Run two and treat a split result as inconclusive — don’t assume one is right by default.
- Testing short snippets. Anything under 100 words returns unreliable scores. Always paste a full section or the complete document.
- Treating a “human” score as proof. Edited AI text passes clean. A low AI-probability score means no statistical signals were found, not that the author is human.
- Skipping the manual read. No tool catches the absence of a real anecdote or lived experience. Your judgment about whether an example is genuinely specific still matters more than any percentage.
- Flagging all polished writing as AI. Some people are just clean, structured writers. Look for the combination of signals, not just good prose.
Frequently Asked Questions
Is AI detection ever 100% accurate?
No. The best free tools reach roughly 85–90% accuracy on unedited AI text, and accuracy drops sharply once text is edited. I once pasted a colleague’s genuinely human-written technical report into GPTZero and it returned 78% AI probability — treat every result as a probability, not a verdict.
Can I detect AI writing in just one paragraph?
Yes, but with less confidence. Paste the suspect paragraph alone and check the per-sentence breakdown. When nearly every sentence shows high AI probability individually, that’s a meaningful signal even if the rest of the document is human. Pair it with a manual check for hedging phrases.
Do AI detectors work equally on ChatGPT, Claude, and Gemini output?
Mostly yes — major detectors train on output from multiple models. Newer model versions often score lower because detector training data lags behind releases. GPTZero and Copyleaks both update their detection models regularly, so it’s worth checking which model version your tool was last trained on.
What if my own writing gets flagged as AI?
Add two or three specific personal examples, vary your sentence length deliberately, and replace generic transitions with your own phrasing. I’ve dropped a score from 75% AI to under 30% just by adding a few concrete anecdotes to a dry section — and re-running confirmed it.
Conclusion
Knowing how to tell if text is AI generated is becoming a standard part of reading critically online. A manual check for real specifics combined with a free detector covers most cases reliably. For keeping your own AI-assisted work credible, I’d also recommend this guide on how to fact-check AI answers before you share them — the habits complement each other well.