Researchers find six patterns that make AI text so exhausting to read
decoder ·
A recent study by researchers at New York University and the University of Texas at Austin identifies six specific linguistic patterns that make AI-generated text difficult and exhausting to read. The analysis of over 1,000 essays and numerous conversations reveals that frontier models systematically overuse unusual word combinations, separated sentence clauses, and stacked nouns, which tax the reader's working memory. Additionally, these models frequently turn nouns into verbs or apply animate actions to inanimate subjects, creating metaphors that humans would not naturally use. While smaller open-source models like Llama and Olmo produce text closer to human norms, larger frontier models exhibit significant deviations due to their training objectives. The researchers argue that the shift from Reinforcement Learning from Human Feedback to Verifiable Rewards prioritizes information density and correctness over clarity. This optimization process encourages a 'dialect' suited for compressing data rather than communicating with people, a trend noted by industry insiders as adapting to 'LLM psychology.' However, targeted fine-tuning efforts, such as those seen in Claude Opus 5.5, have shown promise in restoring readability by addressing these structural issues.
decoder