AI Writing Is Getting Harder to Read — Research Says It's Trainable
lmoroney · x · 2026-10-25
Researchers at NYU and UT Austin (Thomas Hikaru Clark, Matias Martin, Kyle Mahowald) quantified why AI writing keeps getting harder to read:
- Across chat, coding, and essay tasks, frontier models increasingly use sentence habits that slow human readers: rare word pairings, stacked nouns ("reader-patience budget"), dropped optional words like "that", and objects doing things only people do ("this asymmetry bites")
- Their explanation: post-training rewards correct answers in math and code, and shorter answers cost less. Fine-tuning a small Qwen model with a length penalty made it pick shorter phrasing almost every time, and the habit generalized to constructions it never saw in training
- The good news: Claude Opus 5.5 scored better than Opus 5 on 19 of 21 readability comparisons, so the habit can be trained back out
Practical tip: if you write with AI, read the draft aloud and put back the small words that make it easy to follow.
Related event: Study Finds AI Writing Optimized for Information Density Over Readability(2 posts)→
More from Models
- Google searches, LLMs predict: an ex-Meta AI engineer explains why the two work differently — HarperSCarroll · 2026-10-26
- Dev builds Nemotron Bolt gatekeeper to block harmful agent tool calls, weighs abliterated models — gnukeith · 2026-10-26
- Top-b sampling: entropy-adaptive decoding beats Top-p on GPQA and cuts variance 10x — cneuralnetwork · 2026-10-26
- Anthropic token volume on OpenRouter up 3.6x since Opus 5.5 launch — petergyang · 2026-10-26
- Claude runs 53 hours unsupervised, beats human designs on 7 of 8 quantum problems — FinanceYF5 · 2026-10-26
- DesignArena data: Gemini 3.8 Flash quality degrades sharply weeks after launch — legit_api · 2026-10-26