Working on a personal checklist for reading AI-generated text by eye, without running anything through a tool. Partly because tools have false positive problems I’ve mentioned before in other threads, partly because I want to understand the underlying signals rather than outsourcing the judgment.
What I have so far, based on reading a lot of AI output:
- Transitions that are too smooth. Human writing has more friction, more abrupt shifts. AI tends to connect everything with confident bridge phrases.
- Closing sentences that summarize and “resolve” the paragraph. Human writers often end on something unresolved, a question, a contradiction, something the reader has to sit with.
- Balanced “on the other hand” structures that present multiple perspectives without actually committing to any of them.
- Vague attribution. “Studies show,” “experts agree,” “research suggests” with no specific source.
- The word “delve.” I don’t know a single human writer who uses this word regularly.
What am I missing? Specifically interested in signals that are reliable across content types, not just student essays. Things that hold up in professional copywriting, journalism, long-form editorial content.
Good list. A few I’d add from editing manuscripts:
The almost-specific anecdote. AI will describe a scenario in enough detail to feel concrete but not enough to be verifiable. “A marketing manager at a mid-sized company” rather than something with genuine identifying texture. Real writers either use real examples or clearly fictional ones. AI lands in an uncanny valley between the two.
Opinions without stakes. Human writers with genuine views have something to lose by stating them. AI expresses opinions that feel weightless — technically a position, but not one that cost anything to take.
And the positive-to-negative ratio in any piece of analysis. AI rarely reaches a conclusion more negative than “there are some challenges.” Human analysis is more willing to say something is bad.
The “delve” observation is well documented at this point. There’s a small set of words that appear at statistically anomalous rates in AI output — “certainly,” “it’s worth noting,” “importantly,” “in today’s [landscape/world/environment].” These aren’t definitive but they’re useful flags.
The more reliable signal in professional content is specificity calibration. AI matches its level of detail to the genre it’s mimicking, not to actual knowledge. A piece that’s correctly structured for an expert audience but doesn’t actually say anything an expert couldn’t have said in 2019 is suspicious in ways that are hard to articulate but easy to feel once you’ve read enough of it.
From my experience reviewing a lot of content: the absence of a wrong turn is a reliable signal.
Human writers make choices, some of which are slightly off. They pursue a line of reasoning and then partially walk it back. They include an example that doesn’t quite land. They have an odd sentence that’s clearly where they were thinking out loud and didn’t fully clean it up.
AI is consistently smoother than that. Everything is at the same level of competence. No rough patches, no evidence of actual thinking happening in real time. Paradoxically, the cleanness is suspicious.
Interesting list. I’d add: the relationship between the claim and the evidence.
Human writers who know their subject well make claims that are slightly ahead of the evidence — they’re drawing on experience and intuition in addition to what they can directly cite. AI makes claims that are exactly as strong as the evidence supports, no more, no less. It’s very well-calibrated in a way that real subject matter expertise isn’t.
Also: humor. Not whether it’s funny, but the location and type. Human writers make jokes in unexpected places, sometimes at their own expense. AI places light humor at predictable structural moments and it’s always the inoffensive observational variety. Never dry, never weird, never targeted.
the “balanced perspectives without committing” thing on your original list is probably my most reliable signal tbh.
human writers have opinions. even when they’re trying to be fair, you can usually tell what they actually think by which side gets the better examples or the last word. AI genuinely doesn’t have a preference and it shows. everything gets equal weight even when the evidence doesn’t support equal weight.
the other one i’d add: callbacks. human writers return to earlier ideas, sometimes explicitly, sometimes not. they build on something they said three paragraphs ago. AI treats each paragraph as mostly self-contained. the internal coherence is local, not global.