Text written by artificial intelligence models uses language patterns that are harder for people to process than the writing humans produce, according to a new study that tries to explain why. The researchers say the problem may stem from what the models are being trained to do.
The paper, posted to the arXiv preprint server on Oct. 21, was written by Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald. It starts from a sentiment many readers already share: AI-generated text has recently become harder to understand. To test that impression, the authors looked at a set of psycholinguistic features, which are linguistic traits known to correlate with how much effort human readers need to make sense of a sentence.
Those features include object relative clauses, long syntactic dependencies and low-frequency n-grams, which are short sequences of words that rarely appear together in everyday writing. The team built a new corpus of AI-generated text spanning a range of subject areas. They then compared it with paraphrases produced through round-trip translation, a method that converts text into another language and back again.
The AI models used these hard-to-process features more often than the paraphrases did. Set against human writing, all nine models the researchers studied used more inanimate subjects paired with animate verbs, a construction that can make a sentence feel like it is doing something it should not be able to do. All nine also used more low-frequency n-grams.
Most of the models went further. They also relied on longer dependencies, meaning the distance between related words in a sentence grows, and on more stacked nouns, where one noun piles up to modify another. In conversation, most also produced more reduced object relatives, a clause structure that drops words readers would normally expect to see. The authors also found substantial differences between model families, so the effect is not uniform across the field.
The researchers offer a hypothesis for the pattern. Human language, they write, is optimized for successful communication. AI-generated text, by contrast, is increasingly optimized for other goals such as math, code and reasoning. In those domains, the pressure is toward information density rather than clarity for a human reader, and that pressure may push the models toward compressed, harder-to-parse phrasing.
To test the idea, the team fine-tuned a language model under a length penalty, which rewards brevity. The model moved away from human-like linguistic behavior, which the authors present as a proof of concept that the optimization target can drive the drift. The results, they say, highlight where the processing demands of humans and large language models come apart, and they point toward possible remedies. The abstract does not spell out what those remedies would look like.
The findings matter for anyone who relies on AI systems to explain things, summarize documents or answer questions. If the models are drifting toward phrasing that people find taxing, that affects how useful their output is in practice, whatever its accuracy. The study does not measure how readers actually fare with the text, but it gives developers a specific set of features to watch as they tune models for technical tasks.
Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.