You can usually spot AI-written content before publishing by looking for five things: sentences that all run the same length, confident claims with no traceable source, headings that promise specifics but deliver generalities, a total absence of named examples, and phrasing that hedges everything ("can help," "may improve") while asserting nothing checkable.
The single most reliable check is the source test: pick the three most specific factual claims in the draft and try to trace each one to a named, findable source. If none of them trace, treat the draft as unverified regardless of how human it reads.
## Signals in the text
AI writing has a rhythm problem. It tends to produce sentences of similar length and similar structure, because it's generating the most statistically likely continuation rather than the sentence a person would actually write next. Read the draft aloud. Human writing varies — a three-word sentence, then a long tangled one, then a medium one. AI writing often marches along at a steady 15-to-20 words per sentence, paragraph after paragraph. That steadiness is a tell.
The second text signal is empty specificity. Look for headings that sound concrete but lead nowhere — "Key Benefits," "Important Considerations," "Best Practices" — followed by paragraphs that never name a tool, a number, a date, or a person. A human who actually knows a topic usually can't help dropping in a specific: a tool name, a version, a real constraint. AI often can't, because it's assembling plausible-sounding generalizations rather than reporting anything it knows.
## Signals in the sourcing
This is where detection gets reliable. Take the draft's most confident factual sentences and ask: where would this come from? According to our AI tool database, we maintain internally verified snapshots of 360 AI tools with pricing and capability details recorded at verification time, and the most recent verification date is 2026-09-18. A human writer working from that database can point to it. AI writing often can't point to anything — it states a number or a claim with no path back to a source, and the number may not exist anywhere outside the draft.
The practical test: highlight every sentence containing a number, a price, a date, a percentage, or a named study. Then try to find each one. If you can't find it in a source you trust, cut it or rewrite it as a general statement. This one pass catches more AI-fabricated content than any stylistic reading, because fabricated specifics are the failure mode that actually hurts you after publishing.
## A five-minute pre-publish check
Run these in order. First, read the first sentence of every paragraph out loud. If they all have the same shape, rewrite at least half. Second, circle every number and claim, and trace each one. Third, check whether the piece contains at least one named, concrete example — a real tool, a real scenario with real inputs. If it doesn't, it isn't finished, whether AI wrote it or not. Fourth, look for hedge stacking: "can potentially help improve" is a phrase that promises nothing. Replace hedges with the actual condition ("helps when X, fails when Y").
Here's a worked example. Suppose a draft says: "Using AI tools can significantly improve your content workflow and save time." That sentence has no named tool, no mechanism, no number, and no condition. Now compare: "If you're drafting product descriptions, generating a first pass and then rewriting the opening line yourself usually beats editing the whole thing, because the opening is where the generic phrasing concentrates." The second version names a task, a procedure, and a reason. The first is a signal; the second is writing.
## Where this check fails
Detection isn't proof. A careful human editor can produce uniform sentences and vague claims under deadline, and a heavily edited AI draft can pass every test above. The signals tell you a draft needs work, not who or what produced it. Treat them as a quality filter, not a forensic tool.
The check also costs time — tracing claims is slower than reading for flow, and on a short social post it may not be worth it. The rule of thumb: the more specific and checkable the claims, and the more a wrong number would embarrass you, the more the source pass earns its keep. On anything with prices, statistics, dates, or attributed findings, never skip it.
If you want a deeper method for catching AI-invented facts specifically, the guide on how to stop ChatGPT from making up fake facts in your content walks through the verification workflow. And if you're spotting robot-sounding phrasing rather than fabricated facts, how do I stop AI from writing content that sounds like a robot wrote it? covers the rewrite side.