I received a piece of writing that feels polished but a little off, and now I’m trying to figure out if it was written by AI or a real person. I need help with practical ways to check for AI-generated text, red flags to look for, and any reliable tools or methods that actually work.
Best check is pattern, not vibes.
A few red flags:
- It sounds smooth but empty. Lots of clean sentences, low detail.
- It repeats the same point with slightly different wording.
- It avoids strong opinions, odd memories, or specific facts.
- It stays grammatically neat for too long.
- It gives balanced, generic takes on everything.
Practical tests:
- Ask the writer follow-up questions. Ask why they chose a phrase, source, or example. Human writers usually explain their choices fast. AI users often get fuzzy here.
- Look for fake citations, broken quotes, or sources that do not exist.
- Check for specificity. Real people tend to mention names, dates, places, mistakes, side notes.
- Paste a few paragraphs into 2 or 3 AI detectors, but do not trust one score. False positives happen a lot.
- Compare it to the person’s older writing. Sentence length, vocab, and tone usually give it away.
One big thing. AI detection tools are weak. Some studies found false positive rates high enough to flag human writing, espiecally if the writing is formal or if the writer is a non-native English speaker.
So, do this. Use detectors as a small clue. Use source checks and follow-up questions as your main test. That gets better results.
One thing I’d add to what @viaggiatoresolare said: check the editing history, not just the final text.
If it came from Google Docs, Word, Notion, etc., a human draft usually shows messy growth. Deleted lines, rewrites, weird mid-sentence fixes, paragraphs moved around. AI-pasted text often appears all at once, then gets a few cosmetic edits. That’s honestly one of the strongest clues imo.
Also, watch for context failure. AI can sound polished but miss the social reality of the piece. Example: it references events in the wrong order, uses examples that don’t fit the audience, or answers the topic in a slightly “adjacent” way. Humans drift too, sure, but AI drift has a very specific floaty feel.
I kinda disagree with the “grammatically neat for too long” thing as a major red flag. Some people just write clean. Esp if they’re trained to write for work or school.
A couple more checks:
- Ask for the draft notes or outline
- Ask them to revise one paragraph in a very specific new direction
- See whether the voice stays consistant across emails, texts, and older writing
- Check if the piece uses terminology correctly, not just confidently
If they can defend the logic, adapt it fast, and show the mess behind the draft, it’s probably human. If it’s all surface and no process, ehh, that’s where I’d get suspicious.
Biggest thing: don’t trust AI detectors as proof. They’re useful as smoke alarms, not judges. False positives hit non-native writers, formal writers, and anyone using plain, predictable structure.
A check I use that complements @viaggiatoresolare’s points is fact density. Human writing usually contains a few oddly specific details that connect naturally. AI often gives “specific-looking” details that are generic on inspection. Try verifying 3 to 5 concrete claims. If two start wobbling, that’s a bad sign.
Another one is compression pressure. Ask for the same idea rewritten in half the length without losing meaning. Humans who understand their own text usually compress cleanly. AI-assisted text often turns vague fast when forced to get tighter.
I slightly disagree with the idea that “surface with no process” always means AI. Some humans draft in their head and paste clean copy. So I’d focus more on whether the writer can explain choices than whether they have messy drafts.
Quick pros and cons of using ':
- Pros: can improve readability, consistency, and cleanup
- Cons: can flatten voice, introduce confident mistakes, and make authorship harder to judge
Best test is a bundle, not one clue:
- Verify facts
- Ask for a shorter rewrite
- Ask why certain examples were chosen
- Compare with older writing
- Use detectors only as a weak signal
That combo gets you closer than any single “AI tell.”