Why is AI generated language so hard to understand?
- ID
- 39248
- Status
- summarized
- Published
- 25 Oct 2026, 2:04 AM
- Fetched
- 25 Oct 2026, 7:56 PM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://arxiv.org/abs/2610.29162
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 25 Oct 2026, 7:56 PM
- Tags
- Audience
- developersvibe_codersai_ml_learnersai_agent_userssaas_founders
What happened
An arXiv paper (2610.29162, submitted 21 Oct 2026) by Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald measures psycholinguistic features linked to human reading difficulty in AI-generated text. Across a new multi-domain corpus, all nine frontier and near-frontier models tested used more inanimate subjects with animate verbs and more low-frequency n-grams than human writing, and most also produced longer syntactic dependencies, more stacked nouns, and more reduced object relatives in conversation — with substantial variation between model families. The authors hypothesise that training pressure toward information density for math, code, and reasoning (rather than successful communication) explains this, and show a proof-of-concept where fine-tuning under a length penalty pushes a model further from human-like linguistic behaviour.
Why it matters
If your product ships model-written copy, docs, chat replies, or support macros, this says the density problem is systematic across all nine models tested, not a quirk of one vendor — so 'just use a better model' is not the fix. The concrete lever the paper points at is the objective: models tuned toward density (math, code, reasoning) read worse, so add an explicit readability pass or human paraphrase step for anything customer-facing, and check stacked nouns and long dependencies specifically rather than skimming for tone.
Discussion angle
The paper's density hypothesis implies clearer output may cost you reasoning quality — so where in your stack do you want the tradeoff: a reasoning-heavy model plus a rewrite pass, or a lighter model that writes more plainly from the start? Worth testing on one real customer-facing surface this week.