# AI writing may save tokens at the reader's expense **Summary:** An AI writing study suggests pressure to save tokens can make prose harder to follow. I would judge explanations by reader understanding, even when clearer wording takes more space. - Canonical: https://markhuang.ai/news/ai-writing-token-savings-reader-cost - Language: en - Author: [Mark Huang](https://markhuang.ai/about) - Published: 2026-10-24 - Section: News - Tags: AI Writing, Large Language Models, Readability, Technical Communication, Reinforcement Learning - Source: [arXiv](https://arxiv.org/abs/2610.29162) - License: https://creativecommons.org/licenses/by-nc/4.0/ --- ![A metal press compresses a blank ivory paper ribbon into tightly packed folds beside a loose, open section](https://cdn.markhuang.ai/news/ai-writing-token-savings-reader-cost/hero.webp) *A conceptual illustration of the tradeoff: packing an explanation more tightly can leave the reader with more to unpack.* A paper submitted to arXiv on October 21, 2026 gives me a reason to reconsider a routine instruction: "be concise." The researchers behind ["Why is AI-generated language so hard to understand?"](https://arxiv.org/abs/2610.29162) find that all nine frontier and near-frontier models they examine use more unusual word sequences than human reference texts. They also find other language patterns associated with comprehension difficulty. My read is that an explanation can save tokens while leaving its reader with more work. I would give a coding agent more room to explain a change if that helped someone review it. I want the reviewer to follow the explanation and have enough detail to challenge its conclusion. The paper offers evidence for that concern and a possible mechanism. It does not establish how much longer people take to read these answers, or prove that a particular frontier model's training caused its writing style. Those distinctions matter before turning the result into another universal writing rule. ## What makes the prose awkward The researchers collect text from 14 models across conversation, coding, and essay settings. They use 15 starting prompts for conversations and coding sessions, alongside essay assignments. Their analysis looks beyond familiar complaints about favorite words. It examines long distances between grammatically related words and strings of nouns modifying other nouns. It also counts inanimate subjects used with verbs that typically describe animate actors. These are ordinary possibilities in English. The concern is how frequently a model chooses them. A compressed expression can leave the reader to reconstruct relationships that a longer phrase would state explicitly. The paper also distinguishes relative clauses that keep an optional word such as "that" from those that omit it. Across many of its comparisons, models favor the version without that extra word. To compare alternative wording, the researchers translate sentences from English into Japanese and back using a translation model. Many of the difficulty-linked features become less frequent after this round trip, especially in chat and coding. The essay results are less pronounced, which makes me cautious about applying the findings to every kind of AI writing. There is a caveat in the comparison itself. The [2021 study "Machine Translationese"](https://aclanthology.org/2021.eacl-main.188/) found that machine translation reduced lexical and syntactic richness in the language pairs it tested. Translation changes style in its own way. I would treat the round trip as evidence that alternative wording exists, while asking for direct reader tests before calling it better wording. The human references here are also approximate matches, including subtitles for chat and StackExchange for coding. ## What the length penalty changed The authors suspect that pressure to pack in information may compete with clear communication. They test a small, controlled version of that idea by training Qwen2.5-1.5B-Instruct to turn structured event descriptions into English sentences, with a penalty for longer responses. The penalty pushes the model toward shorter grammatical alternatives. That preference also carries over to related constructions absent from training. Transfer to more distant constructions is much narrower: it appears in only one of four tested alternatives, and only without the regularization that limits drift from the starting model. That is evidence for a possible mechanism, with limited reach. Penalizing length during training differs from asking for a shorter answer during use. This experiment cannot tell me what a concise prompt will do. The experiment makes me wary of treating fewer tokens as a sufficient measure of a good explanation. The paper also contains an encouraging comparison. Claude Opus 5.5 scores lower than Opus 5 on difficulty-linked measures in 19 of 21 comparisons, with non-overlapping confidence intervals in eight. ## Familiar shorthand can still be useful An expert may understand compressed technical language perfectly well. That is the objection I find most persuasive. In [Waterloo's interview about "Claudish,"](https://uwaterloo.ca/news/media/q-and-experts-what-claudish) computer scientist Yuntian Deng describes expressions drawn from existing technical fields. He also argues that familiarity can help people understand an agent's output without repeatedly asking for simpler wording. The [Hacker News discussion of the paper](https://news.ycombinator.com/item?id=50223647) reflects that disagreement. One commenter describes slowing down to understand the prose; another says the constructions feel familiar as a developer. These are anecdotes, but they point to a question I would keep in the evaluation: who needs to read this? I would keep technical vocabulary that helps the intended audience. An explanation needs work when it assumes knowledge its reviewer does not have. My earlier piece on [Claude's "load-bearing seams"](https://markhuang.ai/news/claude-load-bearing-seams-no-subject) examined prose that never named its subject. This paper raises a different concern: even when the subject is present, compressed grammar may make its relationships harder to follow. ## I would test the handoff to the reader For an agent's change summary, I would ask a reviewer to identify the changed behavior and the condition under which it applies, then check that understanding against the patch. If the explanation needs a rewrite, I would compare whether the revised version makes those points easier to find while preserving the evidence. The paper does not report this evaluation; it is one I would run. A few extra connecting words would be a reasonable cost if they made the explanation easier to check. More text would still need to earn its place. I would reject a longer answer that buries the decision just as readily as a short one that leaves out the reasoning. The paper is limited to English, a selection of models, and a small set of starting prompts. Its language measures give me a reason to investigate reader effort. They do not yet tell me how large the cost is. For now, I would ask an agent to name the behavior and explain the relevant relationships explicitly. I would decide whether the answer was concise enough after checking whether I understood it.