How can I tell if something was written by AI?
Read for substance, not style. AI text is fluent but low in specifics: no names, dates, numbers, or first-hand detail, and citations that don't resolve when you check them. Style tells are unreliable — seven detectors flagged 61% of non-native English essays as AI. Verifiable content is the honest signal.
Why — the first-principles explanation
A language model writes by predicting likely next words. It has no memory of a Tuesday, no source it can point to, and no way to check whether a fact it produced is true. What it has is an extremely good sense of what a sentence about this topic usually sounds like. So the characteristic failure of AI writing isn't bad grammar — it's fluent emptiness. The prose is smooth, the paragraph is the right shape, and when you ask what it actually told you, the answer is often very little.
This is why substance beats style as a test. Style is exactly what the model is good at, and it's also what innocent humans share with it — formal writing, second-language writing, and plain competent prose all look 'AI-ish' by any statistical measure. Substance is where the model is structurally weak. It can't invent a specific verified number it never saw. When it tries, it produces plausible-looking specifics that fall apart on contact: a citation to a paper that doesn't exist, a statistic with no source, a quote no one said.
So the checks that work are the ones a machine can't cheaply pass: Does this contain checkable claims? Do they check out? Click the citation. Search the quote. Look up the number. A human expert writing from knowledge produces details that survive verification. A model producing plausible text produces details that dissolve.
The catch worth naming: a person using AI well — generating a draft, then adding real detail and verifying it — produces text that passes every one of these checks. That's not a flaw in the method. At that point a human has done the work of making the claims true, which is most of what you actually cared about.
An example that makes it click
Imagine two people describing a restaurant they claim to have visited. The first says: "The atmosphere was warm and inviting, and the dishes showcased a thoughtful balance of traditional technique and modern flair." Beautiful. Also, they could have written that without leaving the house.
The second says: "We waited 40 minutes because the host lost our name, the branzino was $34 and came with the head on, and the guy at the next table was arguing about his fantasy draft." You believe the second one instantly — not because the writing is better, but because it contains things you could go check, and things a person only knows by being there. AI writes like the first diner. It's fluent about the vibe and silent on the branzino.
How to do it
- Ask what the text actually claims. Summarize it in one sentence. If you can't, because it's all connective tissue and no content, that's your strongest signal.
- Check every citation. Click the link, search the paper title, look up the DOI. Invented or non-resolving references are the single most reliable tell.
- Verify one specific number or quote. Models produce plausible statistics with no real source behind them.
- Look for first-hand detail — names, dates, prices, the small friction of real events. Its total absence across a long piece is suspicious.
- Notice hedged universality: text that applies equally well to any company, any student, any topic, with the nouns swapped.
- Consider the context and the person. A student who's written this way all semester didn't suddenly become a robot.
- Only then, optionally, run a detector — and treat the score as a hint that prompts a conversation, never as a finding.
Key facts
- Seven widely used GPT detectors flagged 61.22% of 91 TOEFL essays written by non-native English speakers as AI-generated, while performing near-perfectly on essays by U.S.-born eighth-graders (Liang et al., Patterns, 2023).
- All seven detectors unanimously misclassified 18 of the 91 non-native essays (19%) as AI-generated — meaning cross-checking multiple tools does not eliminate the error.
- Detectors score on 'perplexity,' which measures how surprising word choices are and correlates with writing sophistication rather than authorship (Prof. James Zou, Stanford).
- The same study found simple prompting strategies — asking a model to use more literary language — both mitigate the bias and let AI text bypass detectors entirely.
- OpenAI discontinued its own AI Text Classifier on July 20, 2023, citing a 'low rate of accuracy.'
▶ The 60-second explainer (script)
How can you tell if something was written by AI? Stop looking at the style. Start looking at the substance. Here's why. A language model predicts likely words. It has no memory of last Tuesday, no source it can point to, no ability to check if something is true. What it's brilliant at is sounding exactly like writing about this topic normally sounds. So the signature of AI text isn't bad grammar — it's fluent emptiness. Smooth prose that, when you ask what it actually told you, told you almost nothing. Style tests fail because style is what the model is best at, and it's what innocent people share with it. Stanford researchers found seven detectors flagged 61 percent of essays by non-native English speakers as AI. Those writers weren't cheating. They were writing plainly. But substance is where models are structurally weak. So: click the citation. Does the paper exist? Search the quote. Did anyone say it? Look up the number. Does it have a source? Real knowledge produces details that survive checking. Generated text produces details that dissolve. That's the test that works — and it's the one that actually matters anyway.
What authoritative sources say
People also ask
Isn't the em dash a giveaway?
No. Punctuation habits, the word 'delve,' and phrases like 'it's important to note' are style markers that plenty of humans use and that any model can be told to drop. Style tells are the least reliable category.
What's the most reliable single check?
Citations. Models routinely produce references that look correct and don't exist. A source that fails to resolve is far stronger evidence than any detector percentage.
What if a human edited the AI text heavily?
Then no method distinguishes it — including detectors. If the person added real, verified specifics, they did the substantive work, which is usually what the rule was protecting anyway.
Should I run it through a detector first?
Run it last, if at all. Leading with a score anchors you to a number with a known error rate that falls hardest on non-native and plain writers.
The same question, asked other ways
- How can I tell if AI wrote this text?6,600/mo
- Was this text written by AI?3,600/mo
- Is this AI generated text?2,400/mo
- Is this AI written?2,400/mo
- How to tell if something is written by AI?1,900/mo
- How to check if something was written by AI?1,600/mo
- Is this text AI generated?1,600/mo
- How can you tell if writing sounds like it was made by AI?1,300/mo
- How to detect AI writing?1,300/mo
- How to check for AI writing?1,000/mo
- How to detect AI generated text?1,000/mo
- How to check if text is AI generated?880/mo
- How to check if something is written by AI?720/mo
- How to tell if something was written by AI?720/mo
- How can you check if a text is AI-generated?720/mo