A new study analyzing AI-generated text across multiple leading models found that all nine frontier and near-frontier language models produce writing that is harder for humans to process than human-written or human-paraphrased text. The models use more inanimate subjects paired with animate verbs, low-frequency word combinations, longer syntactic dependencies, and other features known to slow human reading comprehension.
The researchers propose that this stems from a fundamental misalignment: human language is optimized for clear communication, while AI models are increasingly optimized for math, code, and reasoning tasks where information density matters more than readability. Testing this hypothesis by a model with penalties for length showed the model's output diverged further from human linguistic patterns.
The study documents substantial variation across different model families, suggesting the problem is not universal to all architectures.
This is an AI-written summary. The original reporting belongs to arxiv.org. How we write these · Spot a mistake?