How to Tell if Something Is Written by AI
You cannot tell for certain whether a piece of text was written by AI just by reading it, and neither can any detector. What you can do is look for a cluster of signals, such as generic claims with no specific detail, uniform sentence rhythm, invented or unverifiable facts, and a polished tone that does not match the writer's other work, and then weigh them alongside context. Treat the result as a reason to ask questions, never as a verdict.
This guide explains the common signals, why each one is unreliable on its own, what detectors actually measure and how to respond fairly.
Signals that text may be AI-generated
1. Lots of words, little specific information
AI-generated text often explains a topic correctly in general terms without saying anything only a particular person would know. A product review that never mentions a scuff, a delivery delay or how the item felt in the hand. A reflective essay with no named place, date or person. When every sentence could appear in a hundred other pieces, that is worth noticing.
2. Uniform structure and rhythm
Paragraphs of similar length, each opening with a topic sentence and closing with a neat wrap-up. Sentences that hover around the same length. Lists that almost always come in threes. Human writing tends to be lumpier.
3. Over-hedged, balanced conclusions
"While there are advantages and disadvantages, ultimately it depends on individual circumstances." Models are trained to be cautious, so they often avoid a firm position unless instructed otherwise.
4. Recurring stock phrasing
Certain words and transitions show up frequently in default AI output: "delve into," "it's important to note," "navigating the complexities," "a rich tapestry," "in today's fast-paced world," "furthermore" and "moreover" in quick succession. The specific favourites shift as models change, so do not treat any list as permanent.
5. Facts that do not check out
This is the most useful signal because you can verify it. AI models can produce citations to papers that do not exist, quotations nobody said, wrong dates and statistics with no source. If a reference cannot be found in any library database, or a quoted page number does not contain the quote, that is concrete evidence something is wrong, whether from AI or carelessness.
6. A mismatch with the writer's usual work
For teachers and editors, this is often the strongest clue. A sudden change in vocabulary, sentence complexity or error patterns compared with earlier work, in-class writing or emails is worth a conversation.
7. Missing the brief
AI responds to the prompt it was given, which may not match the actual assignment. Look for answers that address a slightly different question, ignore specific required readings or miss class-specific content.
Why none of these is proof
Every signal above has an innocent explanation:
- Generic writing is common among people writing about topics they find dull or do not understand well.
- Uniform structure is exactly what many students are taught: the five-paragraph essay, topic sentences, signposting.
- Formal stock phrases appear in a lot of human writing, particularly from second-language writers who learned English from academic texts.
- Polished writing may reflect a good editor, a tutor, a grammar tool or just effort.
- A change in style can come from improvement, a different subject or more time spent.
Also, AI-assisted text that has been edited by a person may show none of these signals. So the absence of signs does not prove human authorship either.
How AI detectors work, and their limits
AI detectors analyse statistical features of text, such as how predictable each word is given the words around it and how much that predictability varies across sentences. They output a score or label estimating the likelihood that the text was machine-generated.
Things to keep in mind:
- They produce estimates. A score of 80% is not a finding that 80% of the text was written by AI, and it is not proof of anything.
- False positives happen. Human writing can be flagged, especially short texts, formulaic writing and writing by non-native speakers.
- False negatives happen. Edited or heavily prompted AI text can score as human.
- Detectors disagree. The same passage can get very different results from different tools.
- Short texts are unreliable. A paragraph or two gives a detector little to work with.
An AI detector can be one input among several, for example to decide whether a closer look is worth the time. For software, an AI code detector has the same strengths and limits. Neither should be the only basis for a decision that affects someone's grade, job or reputation.
A fair way to investigate
If you are a teacher, editor or manager with a genuine concern:
- Check verifiable facts first. Look up the references and quotes. This is objective and often settles the matter.
- Compare with known work. Look at earlier submissions, in-class writing or drafts.
- Ask about process, not guilt. "Can you walk me through how you developed this argument?" or "Can I see your notes and earlier drafts?" A person who wrote the work can usually talk about it in detail.
- Look at version history if the document was written in a tool that records it.
- Follow your organisation's policy. Many institutions have specific procedures for suspected AI misuse and specific guidance on detector use.
Prevention beats detection
For educators, assignment design often helps more than any detector: in-class writing, staged submissions with drafts and outlines, reflective notes on process, oral follow-ups and prompts tied to local, recent or personal material. For organisations, a clear policy on when AI use is fine and when it must be disclosed removes much of the guesswork.
FAQ
What is the most reliable sign of AI writing?
Verifiable factual errors, especially citations or quotations that do not exist, are the most concrete signal because you can check them. Even then, they show the work is unreliable, not definitively that AI wrote it.
Can AI detectors be wrong?
Yes. They can flag human writing (false positives) and miss AI-generated or edited text (false negatives), and different detectors often give different scores for the same passage. Treat their output as an estimate.
Is it fair to accuse someone of using AI based on a detector score?
A score alone is weak evidence. A fairer approach combines it with other information, such as fact checks, comparison with the person's other writing and a conversation about how they produced the work, and follows any formal policy that applies.
Can you tell if text was written by AI and then edited by a human?
Usually not with confidence. Editing tends to remove the patterns detectors and readers rely on. This is one reason clear disclosure policies are more practical than trying to detect every case.