I’m trying to figure out whether some writing I received was created by AI, but I’m not sure what signs to look for. It came up during a review process, and I need help identifying reliable ways to spot AI-generated text without making the wrong call.
Treat AI detection like fraud review, not like a vibe check.
A few signs help:
- Low variance. Sentences stay the same length. Paragraphs feel even. Human writing usually drifts a bit.
- Generic specificity. It names concepts, but skips lived detail. Example: “The team improved efficiency through collaboration.” Human text often adds one odd, concrete thing.
- Clean transitions everywhere. “Additionally,” “Furthermore,” “In today’s environment.” AI leans on glue phrases.
- No real risk. The text avoids strong claims, niche opinions, or messy nuance.
- Repetition. Same point, rephrased 2 or 3 times.
- Wrong confidence. It sounds sure even when the facts are thin.
What to do:
- Ask for sources. Fake or vague citations are a red flag.
- Ask follow-up questions about choices the writer made.
- Compare with their older writing.
- Check for factual slips and made-up references.
- Paste small sections into a few detectors, but do not trust one score.
Most detectors throw false positives. Some studies put accuracy all over the place, esp for edited AI text. Use them as one input, not proof. If this is for a review process, build a file with patterns, version history, and inconsistencies. One weirdly polished paragraph is not enough on its own.
Big thing: don’t confuse “polished” with “AI.” I kinda disagree with the low-variance test people use, including part of what @codecrafter mentioned, because plenty of humans write in a steady rhythm, especially in business, legal, or academic stuff.
What helps more is process evidence.
- Check revision history. Human drafts usually show detours, deletions, half-finished thoughts, wording swaps. AI-pasted text often appears in one big chunk.
- Ask for the draft before the final. Not “prove you wrote it,” more like “show your development.”
- Look for mismatched register. If one section sounds like a grant proposal and the next sounds like a reddit explainer, that’s interesting.
- Watch pronouns and perspective. AI often drifts from “I” to “we” to impersonal voice for no real reason.
- Test domain depth. Ask them to expand one specific sentence using examples from their own workflow or experience. This is where weak AI-assisted writing starts to wobble.
- Check whether quotes, page numbers, and references actually map to real material. Not just whether the source exists.
Honestly, the most reliable answer is usually: you can suspect it, but you rarely can “detect” it cleanly from text alone. Detectors are shaky, vibes are shakier. Build a case from metadata, drafts, and follow-up answers. That’s the less flashy but more defensible route, imo.
Text-only detection is weaker than people want. I partly disagree with @codecrafter on one common tell too: “too generic” is suggestive, not decisive. Plenty of rushed human writing is generic.
A few extra checks that help:
- Look for compression without consequence. AI often summarizes complex ideas smoothly but leaves out the one messy caveat a real practitioner would naturally include.
- Probe unusual certainty. Confident claims with no visible uncertainty, tradeoff, or edge case can be a flag.
- Compare against known samples from the same author. Not just tone. Check sentence length, punctuation habits, favorite transitions, and how they handle specifics.
- Ask for source-grounded edits. “Rewrite paragraph 3 using only these two documents.” AI-assisted writers often struggle to stay tightly bounded.
- Check factual density. AI text can be fluent while oddly thin, repeating abstractions instead of adding new information.
Pros of using ':
- Can improve readability fast
- May help standardize tone
- Useful for cleanup and structure
Cons of using ':
- Can flatten voice
- May introduce bland phrasing
- Can make authorship questions harder later
Best rule: don’t try to “catch AI” from vibes. Try to test authorship, familiarity, and control over the material. That holds up better.