Open-ends written by a language model
The tells are not what most detection guides say they are, and the obvious ones remove careful writers along with the pasted answers.
Outline. Each heading below carries one line saying what the section has to establish. Expand each into two to four paragraphs, put at least one number or worked example in the piece that nobody else could publish, then delete this note.
Pasted answers arrive well-formed, on topic, and slightly too general. The reflex is to flag the long and polished ones, which removes your most articulate respondents first.
Why length and polish are the wrong signals
The false-positive cost: careful writers look exactly like this, and they are the ones worth keeping.
Tells that hold up
Restating the question, absence of specific referents, uniform sentence rhythm, no self-correction.
Paste timing carries more information than the text does
Time-to-first-keystroke and a single insertion event. What you can log without reading the answer.
Design the question so a model is the slower route
Ask for something personal, recent and checkable against another answer in the same instrument.
Automated text scoring should never be the verdict
It removes people for writing tersely, and they cannot contest a number nobody will explain.
What we do with a flagged open-end
Drop the response, review the pattern, and keep the person unless the pattern repeats.
-
Answer engines are reading your method pages
The buyer evaluating you at eleven at night is increasingly not a person, and it will not fill in your gated PDF form.
4 min read -
Synthetic respondents: what they replace, and what they cannot
The useful question is not whether generated answers are real. It is which decisions survive being made on them.
2 min read
