How to Tell If Something Is Written by AI: Signals and Limits

You can't tell for certain whether text was written by AI from the words alone, but several signals together can raise reasonable suspicion: generic content with few specifics, uniform sentence rhythm, stock phrases, confident errors, and citations that don't exist. AI detectors add a probability estimate, not proof. The fair approach is to treat signals and scores as reasons to look closer and talk to the writer, never as grounds for an accusation by themselves.

Signals in the writing

None of these proves anything on its own. Plenty of careful human writers show some of them. They matter when several appear together, and especially when they don't match what you know about the writer.

1. Generic content, few specifics

AI text often makes reasonable points that could apply to almost anything. Watch for:

• Examples that are hypothetical and vague ("a company might use this to improve efficiency").
• No reference to class discussions, assigned readings or the writer's own experience.
• Claims that sound right but don't engage with the actual prompt in detail.

2. Uniform structure and rhythm

Paragraphs of similar length, each opening with a topic sentence and closing with a mini-summary. Sentences of similar length and shape. Lots of lists in threes. Human writing is usually lumpier.

3. Stock vocabulary and phrasing

Some words and phrases appear in AI output far more often than in typical writing: "delve," "tapestry," "it's important to note," "navigate the complexities," "plays a crucial role." One of them means nothing. A cluster of them is worth noticing.

4. Hedging without a position

The text balances every point ("on the one hand... on the other hand") and concludes that the answer "depends on many factors." An argument assignment with no argument is a signal, though it's also a common habit of students who are unsure.

5. Confident errors and fabrications

This is the strongest textual signal. Look for:

• Citations to articles or books that don't exist, or that exist but don't say what's claimed.
• Quotes attributed to real people that you can't find anywhere.
• Plausible but wrong dates, numbers or definitions.

Check two or three references. If they don't exist, that's concrete, verifiable evidence of a problem, whatever its cause.

6. A mismatch with the writer

For teachers and editors, the most useful comparison is with the person's other work. A sudden jump in vocabulary, structure or polish compared with in-class writing is worth a conversation. So is a writer who can't explain a term they used repeatedly.

Signals outside the text

Sometimes the context tells you more than the words. These are worth considering, again with care:

• Process evidence. A document's version history may show the text arriving in one large paste rather than growing over hours or days. That's not conclusive (some people draft elsewhere and paste in), but it's a reason to ask about the drafting process.
• Leftover prompts or chatbot phrases. Occasionally text includes remnants like "Certainly! Here is an essay..." or "As an AI language model." These are rare in careful submissions, but they're hard to explain otherwise.
• Prompt drift. The text answers a slightly different question from the one you asked, often a more generic version of it.
• Formatting quirks. Unusual heading styles, bolded labels at the start of every paragraph, or markdown symbols that didn't render in the final document can suggest text was copied from a chat interface.
• Out-of-date or out-of-scope material. The writer cites ideas or examples that weren't in the course and that they can't explain where they found.

As with textual signals, look for patterns rather than single clues, and keep an open mind. Each of these has an innocent explanation in some cases.

What AI detectors can and can't tell you

AI detectors analyze statistical properties of text, such as how predictable the word choices are and how much variation there is between sentences, and estimate how likely it is that a model produced it. ZeroTell's AI Detector, for example, highlights sentences that look most machine-like so you can see where the score comes from rather than relying on a single number.

AI DetectorSentence-level AI score

Used well, a detector helps you decide where to look. Its limits are important:

• It's a probability, not a fact. A high score means the text resembles AI output, not that AI wrote it.
• False positives happen. Short texts, formulaic genres, non-native writing and heavily edited prose are more likely to be misjudged.
• False negatives happen. Edited or mixed AI text can score low.
• Different tools disagree. The same text can get very different scores from different detectors.
• Scores don't explain intent. A student may have used AI in a way the course allowed, such as grammar help or brainstorming.

Many universities and detection providers advise against using a score as the sole basis for a misconduct finding. That's sound practice.

A fair process when you're suspicious

Whether you're a teacher, an editor or a manager, here's a reasonable sequence.

1. Note the specific signals. Write down what made you suspicious: particular passages, missing specifics, unverifiable citations.
2. Verify facts and sources. Look up the references. Fabricated sources are concrete; they don't depend on interpreting style.
3. Compare with known work. Look at the writer's earlier assignments, in-class writing or previous reports.
4. Optionally run a detector. Treat the result as one more data point, and note which tool and settings you used.
5. Talk to the writer. Ask about their process, sources and choices in a neutral, curious tone. Ask to see drafts or version history.
6. Follow your institution's procedure. If concerns remain, use the official process rather than an informal penalty.

"I'd like to hear more about how you put this essay together. Can you walk me through how you found your sources, and show me an earlier draft or your notes? I want to understand your process before I finish grading."

That framing gives an honest writer a fair chance to show their work, which is what you want, and it often clarifies things quickly either way.

Making AI use easier to judge in the first place

Detection is hard, so prevention and clear expectations do more:

• Write clear AI policies for each assignment: what's allowed, what must be disclosed.
• Ask for process artifacts: outlines, drafts, annotated sources or a short reflection on how the work was made.
• Design specific prompts that connect to class discussions, local data or personal experience.
• Include some in-class writing to build a baseline of each student's voice.

If you're a writer worried about being misjudged

Keep drafts and use a tool with version history. Write with specifics from your own reading and experience. If you're questioned, our guide on how to prove you did not use AI explains what evidence helps.

Key takeaways

• Signals in the text can raise suspicion but rarely settle the question.
• Fabricated citations and facts are the most concrete evidence.
• Detector scores are probabilistic and can be wrong both ways.
• Compare with known work, verify sources and talk to the writer.
• Never accuse someone on a score alone.