Photo by MJ Duford on Unsplash
The honest answer to “can you tell if text was written by AI” is: sometimes, unreliably, and not with the tools sold for the purpose. Which matters, because those tools are being used to make decisions about people.
What follows is what the research actually shows, who gets hurt by the errors, and the signals that are worth something once you stop expecting a verdict.
The Detectors Fail in a Specific Direction
Researchers at Stanford ran 91 TOEFL essays, all written by humans, through seven AI detectors and published the results in the journal Patterns. The detectors flagged the non-native English writing as AI-generated at an average rate above sixty percent. Essays written by US school students were mostly classified correctly.
This was not one faulty product. It was a consistent pattern across seven of them, which points at the method rather than the implementation.
Why the Method Cannot Work Well
Detectors measure statistical properties of the text, chiefly how predictable each word is given the ones before it, and how much sentence length and structure vary.
Neither measures authorship. Both measure regularity. Generated text is regular; so is a technical manual, a legal summary, a school essay written to a rubric, and a paragraph composed carefully by someone working in a second language. The signal the tool detects is present in plenty of human writing produced under constraint.
Photo by Zoshua Colah on Unsplash
The signal the tool detects is present in plenty of human writing produced under constraint.
And It Is Trivially Evaded
The same mechanism explains why detection fails in the other direction. Because the target is a statistical fingerprint, disturbing the statistics is enough. Passing text through a second model, editing a portion by hand, or translating it and back will generally do it.
Which produces the perverse outcome: the person who wrote honestly in a second language gets flagged, and the person who took two minutes to launder the output does not.
Signals Worth Something
No single item below proves anything. Several together are worth a conversation.
| Signal | Why it appears |
|---|---|
| Even paragraph rhythm throughout | Human writing varies as attention and interest vary |
| Confident claims with no checkable detail | Generated prose favours the general over the specific |
| Attribution to nobody in particular | “Experts agree”, “studies show”, with no study named |
| Citations that do not resolve | Plausible authors, plausible journals, no such paper |
| Nothing surprising anywhere | Models converge on the consensus middle by construction |
The last one is the most reliable and the least mechanical. Generated text rarely contains the odd detail, the aside, or the mild inconsistency that comes from a person having actually thought about something.
Photo by Kaitlyn Baker on Unsplash
What to Do Instead of Accusing
Ask about process. It works whether or not AI was involved, and it does not require you to be right in advance.
- “What did you cut from the first draft?”
- “Which part took longest?”
- “Where did you check this figure?”
- “Why this structure rather than the obvious one?”
Someone who did the work answers easily. Someone who did not usually stops around the second question. No score is produced, nobody is accused of anything, and the conversation is useful either way.
For educators the same reasoning applies with higher stakes, which I set out alongside the student-data rules in AI Tools for Teachers.
It is a weak signal being presented with a decimal point.
Where This Is Heading
Detection after the fact looks like a losing approach. The more promising direction is provenance: metadata attached at creation recording how a file was made. That shifts the question from guessing at properties of the text to checking a record, though only for content that carries it.
In the meantime, treating a detector score as evidence is the mistake to avoid. It is a weak signal being presented with a decimal point.
FAQ: Frequently Asked Questions
Do AI content detectors work?
Not reliably. Published research found average false-positive rates above sixty percent on human-written essays by non-native English speakers, across seven separate tools.
Can AI-written text be identified at all?
Sometimes, by reading rather than scanning. Uniform rhythm, unattributed generalities and citations that do not resolve are the useful signals, and none is conclusive alone.
Why do detectors flag my writing?
Because they measure predictability, not authorship. Careful writing, second-language writing and formulaic genres all score as predictable, which is the flaw rather than a judgement about you.
What should I do if I am wrongly accused?
Produce process evidence: drafts, version history, notes, search history. A document that can be shown being built over time is the strongest available response, and it is free to keep as you go.
The research cited tested the detectors available at the time of publication; both detectors and models continue to change. Nothing here should be treated as a method for making decisions that affect someone’s academic or professional standing.