Guide
How to spot AI-generated text — the patterns that give it away
AI-generated text has developed recognisable patterns. Here's what to look for and why these patterns emerged from the training process.
The vocabulary tells
Certain words appear in AI-generated text at rates far above their use in ordinary human writing. 'Delve' is the most cited example — a slightly formal verb that LLMs favour heavily. 'Tapestry', 'nuanced', 'multifaceted', 'it's worth noting', and 'it's important to remember' are similarly overrepresented. These words are not wrong, but their frequency in a short piece is a useful signal.
Structural uniformity and the absence of voice
Human writers vary paragraph length, shift tone, use specific examples from their own experience, and take positions. AI-generated text tends to produce uniform paragraph lengths, avoid taking a clear position on anything contested, and substitute general statements for specific detail. The text reads as if it was assembled from many sources rather than experienced or thought through by one person.
Why these patterns exist
LLMs are trained to predict the next token based on patterns in training data. The vocabulary and structural tells emerged not from instruction but from the statistical regularities of the text they trained on — formal writing, encyclopaedic summaries, and instructional content. Models were also trained with RLHF to be helpful and agreeable, which reinforces hedging ('it's worth noting') and comprehensive structure over directness and voice.