Why fluent AI prose can still be hard to read

Here is an invented example of fluent prose doing unnecessary work: ‘The report expands the team decision-option landscape.’ ‘The report gives teams more options’ conveys much the same idea with less decoding. A new study examines whether language models systematically favour constructions that impose this kind of effort on readers.[1]

Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald compare English outputs from nine frontier and near-frontier models with human writing and round-trip-translation paraphrases. They find more features associated in earlier research with processing difficulty. This is a corpus study of linguistic proxies, not an experiment measuring how long people took to read the generated answers.[1]

Words that belong together but arrive too far apart

One reported tendency is greater use of long syntactic dependencies. These occur when a reader has to hold an unfinished grammatical relationship in memory while navigating intervening material. A long relative clause can force the reader to remember the subject and its missing predicate through several interruptions. Correct grammar does not remove the effort. The length and arrangement of a sentence can burden working memory while each individual word remains familiar.[1]

The researchers also find more low-frequency sequences of adjacent words, often described as n-grams. A rare sequence can be perfectly valid, yet unfamiliar combinations require more interpretation than expressions readers have encountered repeatedly. Most of the examined models also use more stacked nouns in conversation: several nouns piled together in the service of one concept. Such strings compress meaning but leave the relationship between their parts for the reader to reconstruct.[1]

Another contrast is semantically odd agenthood. The paper reports that all nine models use more inanimate subjects with verbs that typically take animate actors than humans do. A policy 'deciding' or a framework 'believing' can function as metaphor; overused, it makes an abstract thing behave like a person without saying who actually made the decision.[1]

What the comparisons can and cannot tell us

The round-trip-translation comparison is important. It gives the authors an alternative form of paraphrase that transforms wording without being a native speaker creating an original text from scratch. The pattern of differences is therefore harder to dismiss as a generic side effect of any rewriting. The human-writing comparison provides a second reference, though genre, prompt design and topic still matter whenever corpora are compared.[1]

The authors do not claim uniform behaviour across model families. Some constructions appear more strongly in some systems than others. That variation is useful, because a finding that differs across models is harder to attribute to an unavoidable mathematical property of neural language generation. Training objectives, style preferences and reinforcement procedures may all influence output habits.[1]

A hypothesis about density

The authors propose an explanation: human language has been shaped by communication with other minds, while language models are increasingly rewarded for strong performance on code, mathematics and reasoning. Those tasks often favour compact, information-dense representations. The result may be prose that looks efficient to the model but transfers decoding work to the human reader. A proof-of-concept fine-tuning experiment that penalises length produces a departure from human-like patterns, lending some support to the hypothesis.[1]

That experiment is suggestive, not decisive evidence that length penalties alone caused the language patterns of today's models. Many changes occur between pretraining, instruction tuning, system prompts and user interfaces. Nor is density always bad. Technical experts legitimately use shorthand to communicate precise ideas in their own field. The problem arises when a system employs that shorthand for readers who have no reason to know it.[1]

A practical editorial lesson

A useful edit makes the relationship visible: name the person who decided, unpack a stack of nouns, or move a qualification closer to the statement it changes. Shortness alone is no guarantee of clarity. Sometimes the extra words are what make a sentence easy to understand.[1]

Report an error or suggest a correction