There is no reliable way to prove text was written by AI, and any tool that claims otherwise is overselling. AI detectors produce false accusations and miss edited AI text often enough that no score should be treated as evidence. What you can do is read for human tells: specific personal detail, a real voice, small imperfections, and checkable facts. AI writing tends to be smooth, generic, evenly toned, and vague on specifics. Below are the honest signals, why detectors fail, and a practical checklist for teachers, editors, and anyone who just wants to judge for themselves.
The uncomfortable truth about AI detectors
Let’s start where most articles won’t: AI detectors don’t work well enough to trust. They routinely flag genuine human writing as AI-generated and pass lightly edited AI text as human. They’ve famously scored historical documents and human-written essays as machine-made. The reason is simple: as AI writing improves, the statistical “fingerprints” detectors look for get fainter, and human writing that happens to be clean and formal looks exactly like the thing detectors flag.
So if a detector says “98% AI,” that number is not proof. Treat it as a nudge to look closer, nothing more. Acting on a score alone can wrongly accuse a real person, which is a serious harm.
The real tells: what AI writing tends to look like
You’ll do better reading the text than running it through a tool. AI-written text often shares these traits:
- An even, over-polished tone with no rough edges, no asides, no personality.
- Generic filler: phrases like “in today’s fast-paced world,” “it’s important to note,” or “plays a crucial role” that add words but no information.
- Vague, unfalsifiable examples: “many experts agree,” “studies show,” with no specific study, name, or number.
- Perfect but flavorless grammar: correct, but reading like a well-behaved template.
- Repetitive structure: every paragraph the same length and shape, lists of exactly three, tidy summaries that restate the intro.
- Hedging and both-sidesing where a human would just take a position.
None of these is a smoking gun. Plenty of humans write in generic, polished prose, especially students trained to write for exams. That’s why any single tell is weak.
The stronger signal: specific, checkable detail
The best human tell is the hardest thing for AI to fake convincingly: real, specific, personal, verifiable detail. A first-hand anecdote with an odd concrete detail. A named source you can look up. A number that’s oddly precise because it’s real. An opinion with a reason behind it. A mistake that reveals a genuine thought process.
AI can produce specifics, but when it’s used lazily it defaults to the vague and safe. So when you read, ask: Could a machine have written this without knowing anything real? If the whole piece could be generated from the title alone, that’s your signal.
A practical checklist
| Signal | Leans human | Leans AI |
|---|---|---|
| Specific, checkable facts | Present, correct | Vague or absent |
| Personal voice | Distinct, uneven | Smooth, generic |
| Examples | Named, concrete | ”Studies show,” unnamed |
| Structure | Varied | Repetitive, tidy |
| Opinions | Takes a stance | Hedges both sides |
| Small imperfections | A few | Suspiciously none |
| Detector score | Weak signal only | Weak signal only |
Add these up. One AI-leaning trait means little; a piece that hits every column is worth a closer look, but still isn’t proof.
For teachers: the only approaches that actually hold up
If you grade student work, software won’t save you. What works:
- Compare to known writing. A sudden jump in polish and vocabulary is more telling than any detector.
- Ask them to explain it. A short conversation about their argument, sources, and choices reveals whether they understand what they “wrote.”
- Build in process. Outlines, drafts, and in-class writing create a paper trail that’s hard to fake.
An accusation based on a detector score alone is both unfair and easy to challenge. Base concern on evidence you can defend.
For editors and readers
If you’re vetting content you might publish or rely on, the checkable-detail test is your friend. Verify the facts, look up the sources, and if the piece dissolves into generic claims the moment you probe, it may be AI, low-effort, or both, and either way it’s not worth publishing.
The honest bottom line
You can get better at sensing AI writing, but you cannot prove it from the text, and detectors won’t prove it for you. Read for specifics, voice, and verifiable claims; treat detector scores as a whisper, not a verdict; and never accuse someone on a number a tool made up. If anything, the rise of AI text is a good reason to value writing that’s specific, personal, and checkable, the kind a machine can’t fake without doing the actual work.
Curious how the tools you’re trying to spot actually behave? See how to use ChatGPT and our ChatGPT vs Claude comparison to understand what today’s AI writing really reads like.