What disappears when the text is short?

A tiny sample may contain only a greeting, a factual answer, or one familiar opinion. Those tasks invite similar wording from many people and models. With so few choices on the page, punctuation and tone can feel meaningful even when they are ordinary.

  • Reasoning disappears. You cannot see whether one claim follows from another.
  • Specificity disappears. Names, constraints, tradeoffs, and corrections often arrive later.
  • Voice becomes unstable. One sentence is too small a sample for a reliable pattern.
  • The task dominates. “Thanks,” “yes,” and simple definitions give every writer little room to differ.
A three-step diagram showing fragment, context, and known source
Figure 1 · More context adds decisions to inspect; it still does not prove authorship.

What the evidence says

OpenAI described its retired classifier as “very unreliable” below 1,000 characters and said reliability generally improved with longer inputs. It also warned that predictable text cannot be reliably identified. The tool was withdrawn in July 2023 because of its low accuracy.

A separate study by Sadasivan and colleagues stress-tested multiple detector families on passages of roughly 300 tokens. Their results showed that paraphrasing could sharply reduce detection rates, which means length alone does not solve the broader reliability problem.

Inference: shortness removes useful observations, while editing and overlapping human–AI language distributions keep longer samples uncertain. More text can improve a comparison without turning it into proof.

A fairer way to play

  1. Compare texts from the same era, topic, and setting.
  2. Use enough text to expose choices, not just polish.
  3. Reveal known provenance after every guess.
  4. Treat the score as a game result, never an accusation.

The Context Ladder uses three sourced Hacker News comments and three synthetic paired passages already present in this site's verified pool. It records only the score in the current browser tab and sends nothing anywhere.

Sources

This page is not a detector

It does not analyze visitor text or determine authorship. It is a browser-only guessing exercise built from examples with known labels.