Why is AI generated language so hard to understand?

AI-generated text now uses more complex linguistic structures than human writing, potentially making it harder to read and understand.

A new study published on arXiv, "Why is AI-generated language so hard to understand?" by Thomas Hikaru Clark et al., investigates the recent perception that AI-generated text has become less comprehensible. The research examined psycholinguistic features known to affect human processing difficulty, such as long syntactic dependencies and low-frequency n-grams. Analyzing a novel corpus of AI-generated text across various domains, the study found that AI models utilize these complex features more frequently than human-written paraphrases. Specifically, nine leading AI models were observed to use more inanimate subjects with animate verbs and more low-frequency n-grams than human writing, with many also exhibiting longer dependencies, more stacked nouns, and more reduced object relatives in conversational contexts. The researchers hypothesize this trend stems from AI models being increasingly optimized for information density in tasks like math and code, rather than for clear communication, a deviation from human language optimization. A proof-of-concept fine-tuning experiment, which penalized length, demonstrated a shift away from human-like linguistic behavior.

The core finding is that AI models, unlike human language which prioritizes communicative success, are increasingly optimized for information density. This shift means they are favoring linguistic structures that, while efficient for machine processing or complex tasks like coding, pose challenges for human readers. The divergence highlights a fundamental misalignment between the goals of LLMs and human comprehension, suggesting that current AI development may inadvertently create barriers to clear communication.

This research implies that future AI development needs to consider the human readability of generated text. The observed trend indicates a potential need for new evaluation metrics that go beyond task performance to assess clarity and ease of understanding for end-users. Companies and researchers focused on natural language generation may need to adjust their training objectives and fine-tuning processes to re-align AI output with human cognitive processing demands.