AI text fatigue: six patterns NYU and UT Austin researchers identified

What happened

Researchers from New York University and the University of Texas at Austin have found that frontier language models systematically overuse linguistic patterns that are hard for people to parse. According to reporting by The Decoder, the study identifies six recurring patterns behind the sense of fatigue many readers feel when working through AI-generated text.

The researchers trace the problem to how these models are trained. Training that rewards code and mathematics tends to reward correctness, but it does not reward clarity in the way a human reader needs. The study also found that smaller models write in a way that is closer to human prose.

The findings point to a gap between what models are optimised for and what makes their output easy to use. The study has been covered as a research release rather than a product announcement, so its implications are still being tested across the industry.

Why it matters

As organisations embed language models in customer service, internal knowledge tools and document workflows, the readability of machine-written text becomes a service-quality issue. Text that is technically accurate but tiring to read adds cognitive load for customers and staff, and that load shows up in abandonment, repeated queries and lower trust.

For leaders in experience and digital transformation, the study shifts the question from whether a model is accurate to whether its outputs are usable by the people who receive them. It also suggests that model choice is a service decision as well as a technical one, since model size and training emphasis appear to shape how people experience the output.

The Renascence take

Most discussion of AI writing quality focuses on accuracy and hallucination. The more useful question for operators is whether a reader can act on the text within a few seconds, because fatigue compounds across every interaction.

Readability is a design decision, not a finishing touch, and it is easy to overlook when outputs look fluent on the surface. Behavioural economics would point to the effort a reader must spend decoding a message as a hidden cost that drives drop-off long before anyone measures it. A customer-obsessed operator should test AI-generated responses with real readers against a plain-language standard, and treat the smaller or more human-like model as a serious candidate when clarity matters more than breadth. Accuracy that exhausts the reader is not yet good service.

Meta title: NYU and UT Austin study: why AI text is hard to read Meta description: A NYU and UT Austin study links hard-to-parse AI writing to training on code and math, where clarity goes unrewarded. Here's what it means for CX teams.