Signs of AI Writing on Wikipedia (flowingdata.com)

Wikipedia has published a catalog of “signs” commonly seen in AI‑generated contributions — not to outlaw words or punctuation but to help editors spot patterns that frequently occur in LLM output. The guidance stresses these are probabilistic indicators, not proof: models were trained on human writing (including Wikipedia), so overlap is expected, and any single trait doesn’t prove machine authorship. The list is tuned to Wikipedia’s style and editorial context, so some items (especially punctuation and formatting habits) may not transfer cleanly to other kinds of writing. A highlighted technical pattern is the distinctive use of em‑dashes: LLMs tend to insert em‑dashes more often and more formulaically than nonprofessional human writers, often where a comma, parenthesis, colon or hyphen would be more typical. Wikipedia attributes this partly to training data biases — novels and editorial prose in the corpus that overuse em‑dashes — which causes models to mimic that rhythm. The guidance therefore recommends combining multiple signals rather than relying on one feature, acknowledging risks of false positives and the need for nuanced human review. For the AI/ML community, this is a practical example of how training data shapes surface stylistic artifacts and how those artifacts can inform detection and moderation workflows.