Why is AI-generated language so hard to understand?

Abstract: There is a widespread sense that AI-generated text has recently become harder to understand. In this study, we investigate a set of psycholinguistic features that are known to correlate with human processing difficulty (e.g., object relative clauses, long syntactic dependencies, low-frequency n-grams). Using a novel corpus of AI-generated text across diverse domains, we find that AI models use these difficult features more often than paraphrases generated by round-trip translation. Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation; we also demonstrate substantial variation across model families. We propose and test a hypothesis as to why: while human language is optimized for successful communication, AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for \textit{information density} rather than successful communication. We provide a proof-of-concept for this hypothesis by fine-tuning a LLM under a length penalty, which leads to a divergence from human-like linguistic behavior. Our findings highlight the axes of misalignment between linguistic processing demands in humans and LLMs, and point toward possible remedies.

• The paper identifies specific linguistic features, such as low-frequency n-grams, long syntactic dependencies, and rare verb forms, that AI models consistently overproduce, leading to difficult-to-understand text.
• A controlled reinforcement-learning intervention shows that enhancing task efficiency penalizing length leads to more compressed, harder-to-read sentences.
• The findings suggest that AI model language deviates from human-like preferences due to post-training objectives emphasizing information density over comprehension.
• LLM Chatbots are increasingly trained for coding, mathematical reasoning, and tool use, where semantic correctness is rewarded without concern for language being easy to understand.

Research question and central thesis

The paper examines why contemporary LLM outputs are often perceived as unusually difficult to read despite being grammatically well formed. Its central empirical claim is that frontier and near-frontier models systematically overproduce linguistic configurations associated with human processing difficulty. The analysis focuses on rare or low-predictability n-grams, long syntactic dependencies, stacked noun compounds, denominal verbs, inanimate subjects paired with animate-selecting verbs, and reduced object relative clauses.

The authors formulate two hypotheses. H1 predicts that frontier and near-frontier models will overuse features associated with comprehension difficulty. H2 predicts that post-training objectives emphasizing information density, task success, or token efficiency will shift model language away from human-like preferences. The paper supports both hypotheses through a large comparative corpus analysis and a controlled reinforcement-learning intervention (2610.29162).

The theoretical explanation is grounded in a contrast between human communicative production and model optimization. Human language production is constrained by incremental processing, lossy working memory, and the need for successful communication with other humans. By contrast, LLMs are increasingly post-trained for coding, mathematical reasoning, tool use, and verifiable task completion. These objectives can reward semantic correctness without rewarding linguistic forms that minimize comprehension effort. The paper therefore treats difficult model language not simply as a stylistic artifact, but as a possible consequence of objective mismatch.

Corpus design and analytical framework

The first experiment evaluates outputs from 14 models spanning several training regimes and model families. The corpus contains three settings: multi-turn chat, multi-turn coding sessions, and single-turn essays. The main comparison includes nine frontier or near-frontier systems, including closed models from the GPT and Claude families and open-weight systems such as GLM and Kimi, together with smaller Llama and Olmo variants at base, instruct, and reasoning stages.

The data collection design is unusually heterogeneous. Chat consists of ten conversational topics per applicable model, with sessions extending to 12 model replies. Coding consists of five repository-level bug-fixing sessions for each of the nine frontier models. Essays consist of ten responses to each of seven PERSUADE 2.0 assignments, yielding 980 model essays. In total, the chat setting contributes 1,428 analyzed model replies, while the coding setting contains 45 agentic sessions. Human comparisons use genre-matched corpora: OpenSubtitles for chat, StackExchange for coding, and high-scoring student essays from PERSUADE 2.0 for essays.

The linguistic measures are motivated by psycholinguistic findings linking predictability, dependency distance, ambiguity, and lexical or structural markedness to processing cost. N-gram behavior is measured through unattested bigrams and trigram perplexity relative to genre-specific reference corpora. Dependency length is standardized across sentence-length bins to avoid conflating longer sentences with intrinsically longer dependencies. Syntactic features are extracted using spaCy dependency parses, with classifier-based judgments for denominal verbs and inanimate-subject constructions.

The paper’s principal methodological contribution is the use of round-trip translation as a semantic control. Each model sentence is translated from English into Japanese and back into English using NLLB-200. The resulting paraphrase is paired with the original, so semantic content is approximately held constant while lexical and syntactic realization changes. If a construction is more frequent in the original than in the round trip, the difference is interpreted as evidence that the model selected that construction over an available alternative. This design does not establish that translation always preserves every pragmatic or discourse property, but it provides a useful within-response comparison that is stronger than an unpaired comparison against unrelated human text.

Figure 1: Overview of the model families, collection settings, round-trip procedure, and linguistic features analyzed.

Model outputs contain a consistent profile of difficult constructions

The round-trip results show that model outputs generally contain more low-frequency lexical sequences than their translated paraphrases. Across most frontier and near-frontier models, original outputs have more unattested bigrams and higher trigram perplexity than the corresponding round trips. The same pattern appears for denominal verbs, inanimate subjects with animate-selecting verbs, and object relatives without an overt relativizer.

This is important because the comparison is not merely between model-generated text and a human corpus with different topics or registers. The original and translated texts express approximately the same content. The observed differences therefore indicate that the models’ linguistic realization choices themselves contribute to the atypical profile.

The treatment of object relatives is particularly informative. Models tend to increase object relatives without an overt relativizer while decreasing object relatives with an overt relativizer after round-trip translation. The authors interpret this as evidence for a preference to omit optional function words, consistent with an information-density pressure. The finding is not that reduced object relatives are ungrammatical; rather, they are less aligned with the distributional and processing preferences evident in human language.

Stacked noun compounds and dependency length show similar effects in chat and coding, although they are less stable in essays. The setting dependence matters. The models’ linguistic profile is not a fixed property expressed identically in every genre. The strongest deviations occur in iterative, multi-turn environments, where responses must maintain and compress information across extended interactions. Essay outputs, produced from a single prompt, generally show smaller round-trip differences for n-gram measures, dependency length, stacked nouns, and denominal verbs.

Figure 2: Changes in processing-relevant linguistic features after round-trip translation through Japanese, shown in standardized units relative to the original model outputs.

The model-family comparisons further qualify the overall claim. Llama models and the Olmo base and instruct models change relatively little under round-trip translation and are closer to the student-essay distribution on several n-gram and verb measures. Olmo Think, however, resembles the frontier models more closely than the other Olmo variants. Within the Olmo family, the effects increase from base to instruct to reasoning, which is consistent with post-training as a source of the divergence, although the design does not causally isolate the individual training stages.

The contrast between Claude Opus 5 and Opus 5.5 is also notable. Opus 5.5 is lower than Opus 5 on 19 of 21 difficulty-linked comparisons, with non-overlapping 95% intervals in eight comparisons. This result contradicts any simple claim that model scaling or recency necessarily increases linguistic opacity. It instead indicates substantial variation across model families and successive model versions. The paper consequently supports a conditional claim: many frontier models overproduce difficult constructions, but the magnitude of the effect depends materially on post-training and model-development choices.

Figure 3: Cross-model summary of feature changes, showing heterogeneous linguistic preferences across base, instruct, reasoning, open-weight, GPT, and Claude families.

Human baselines reveal specific, not universal, divergence

Comparisons with human text show that the model deviations are selective. All nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams than the relevant human baselines. Most models also exhibit longer dependencies, more stacked nouns, and more reduced object relatives in conversational settings. These results support the paper’s claim that the issue is not simply that model writing has a different style; several of the differences occur along dimensions independently associated with human processing demands.

The human comparisons should nevertheless be interpreted feature by feature. The paper does not show that every model is worse than humans on every measure, nor that every linguistic departure increases comprehension time by a uniform amount. Some constructions, such as denominal verbs or inanimate subjects, can be conventional, contextually licensed, or easily interpreted. The argument is statistical: model outputs shift the aggregate frequency of such constructions in a direction associated with higher processing demands.

The reference corpora also impose constraints. OpenSubtitles, StackExchange, and student essays are only approximate genre matches, and the model prompts are not identical to the production circumstances represented in those corpora. The round-trip comparison therefore carries more inferential weight than the absolute human-baseline comparison. The strongest conclusion is that models systematically choose marked alternatives relative to semantically similar paraphrases, with human comparisons providing convergent rather than definitive evidence of comprehension cost.

Figure 4: Model distributions for the analyzed linguistic features, with bootstrap intervals and genre-specific human and Wikipedia reference lines.

The coding setting is especially relevant to the paper’s interpretation. Coding-agent outputs contain technical identifiers, code-related terminology, and compressed explanations, but the analysis removes code blocks and normalizes several code-like expressions. The remaining prose still exhibits atypical dependencies, noun stacking, rare n-grams, and other features. This suggests that the profile is not reducible to source-code contamination. At the same time, coding discourse has legitimate technical conventions that differ from ordinary conversation, making genre-specific controls essential.

Figure 5: Feature behavior across chat, coding, and essay settings, illustrating stronger and more consistent effects in multi-turn interaction than in single-turn essays.

Information-density pressure as a mechanism

The second experiment tests whether an explicit pressure for shorter outputs can alter syntactic preferences. The authors fine-tune Qwen2.5-1.5B-Instruct with GRPO on a structured-event-to-English generation task. Each event can be expressed using two valid syntactic alternatives, one shorter than the other. The reward combines correctness and format validity with a per-token length penalty. The experiment varies the length penalty $\lambda$, the KL coefficient $\beta$, and random seed across 24 runs.

The trained alternations include dative constructions, object relatives, complement clauses, and particle placement. The evaluation also includes near-transfer constructions involving optional that and far-transfer alternations involving passives, genitives, quantifiers, and purpose clauses. Because both members of each trained alternation satisfy the task’s semantic requirements, a length penalty can favor the shorter form without improving task correctness.

The result is direct: increasing the length penalty increases the probability of selecting the shorter variant. Without KL regularization, any positive length penalty produces a near-categorical preference for the shorter form. For both trained that alternations, the short form exceeds 98% at every positive $\lambda$. When $\lambda = 0$ and $\beta = 0$, different seeds produce unstable preferences ranging from 0% to 100%, indicating that the short-form preference is not an inevitable property of the base model but is induced by the optimization objective.

Figure 6: Reinforcement-learning results showing increased preference for shorter syntactic variants as the length penalty increases, with and without KL regularization.

The effect transfers to held-out constructions involving optional that, even though those specific constructions were not present during training. This supports the claim that the model learns a general compression bias rather than memorizing a particular alternation. Far transfer is weaker: without KL regularization, only the genitive alternation shifts among the four far-transfer categories. Passive and purpose-clause preferences do not shift, while the quantifier alternation is already near ceiling in the base model.

The intervention therefore establishes a plausible mechanism, but not a complete explanation of frontier-model behavior. It demonstrates that a token-efficiency objective can change grammatical choice while preserving task success. It does not show that contemporary frontier systems were trained with an equivalent explicit length penalty, nor that length pressure alone explains their rare n-grams, dependency lengths, or inanimate-subject constructions. The result is appropriately described as a proof of concept for H2 rather than a causal decomposition of deployed-model training.

The role of KL regularization is theoretically consequential. KL regularization attenuates the shift toward short forms, but does not eliminate it. This indicates that staying close to the base model can preserve some human-derived preferences while still allowing optimization to alter them. In the absence of KL regularization, the model converges toward categorical choices even when human speakers use the alternatives contextually. The experiment thus distinguishes information compression from human-like communicative adaptation: a model can remain correct while abandoning conditional variation that serves comprehenders.

Generalization beyond the trained task

The paper evaluates the 24 fine-tuned models on 40 open-ended prompts unrelated to structured event generation. The purpose is to determine whether the RL intervention produces only task-local behavior or changes the model’s broader language distribution. The reported n-gram analysis indicates changes relative to the base model, although the magnitude and interpretation require caution because the open-ended evaluation is small and uses only one model family.

Figure 7: N-gram statistics for unrelated open-ended responses from the base model and length-penalized variants.

The transfer pattern is mixed rather than universal. Near transfer to unseen optional-that constructions is consistent, while far transfer appears only in a subset of alternations and primarily without KL regularization. This asymmetry is theoretically useful. It indicates that objective-induced linguistic preferences can generalize compositionally, but that generalization is constrained by the learned representation, base-model biases, and the relationship between the trained and held-out constructions.

The empirical findings therefore support a more precise version of the information-density account. Optimization for token reduction can alter the model’s distribution over syntactic alternatives, including alternatives not explicitly targeted. However, it does not follow that every difficult construction is generated by a single global compression objective. The model-specific and feature-specific variation in Experiment 1, together with the incomplete far transfer in Experiment 2, argues for multiple interacting causes.

Limitations and open questions

The study is limited to English, a restricted set of models, a small number of chat topics and repositories, and a finite set of essay assignments. The model sample spans several families and training stages, but it cannot identify which components of post-training caused the observed effects. In particular, the paper infers a relationship between reasoning-oriented or verifiable-reward post-training and difficult language from cross-model patterns rather than from controlled training histories.

Round-trip translation is an informative semantic control but not a neutral transformation. NLLB-200 may itself prefer simplified, frequent, or structurally different English, and translation can alter discourse relations, pragmatic emphasis, lexical specificity, and sentence segmentation. The paper checks alternative pivots and translators for word-count behavior, but the central feature analyses remain dependent on the validity of the translation-based counterfactual.

The feature detectors also introduce measurement uncertainty. Dependency parses can misanalyze technical prose, automated judgments of denominal verbs and inanimate subjects can encode classifier biases, and n-gram rarity depends on the reference corpus and smoothing procedure. These measures are established proxies for processing difficulty, not direct measurements of human reading times or comprehension accuracy. The paper therefore demonstrates a linguistic profile associated with processing costs rather than directly demonstrating that each model output is harder for human readers to understand.

The causal account remains open. A stronger test would compare the same base model across controlled post-training regimes that vary only in verifiability, length pressure, KL regularization, communication rewards, or context requirements. Human behavioral experiments could then establish whether the measured changes predict reading times, reanalysis, recall, or task success. A further unresolved question is whether models can be trained to retain task-efficient representations while selecting human-adaptive realizations, rather than merely increasing output length.

Conclusion

The paper provides convergent evidence that many frontier and near-frontier LLMs overuse linguistic forms associated with human processing difficulty. Relative to round-trip paraphrases, they produce more rare n-grams, reduced object relatives, denominal verbs, inanimate-subject constructions, and, in several settings, longer dependencies and denser noun compounds. The effects vary substantially across model families and tasks, with Opus 5.5 providing a notable counterexample to a monotonic deterioration claim.

The controlled GRPO experiment shows that a token-level length penalty can induce near-categorical preferences for shorter syntactic variants, including in some unseen constructions, while preserving formal task correctness. This establishes a concrete route by which optimization for information density can diverge from human-like communicative behavior. The paper’s main contribution is consequently both empirical and mechanistic: it identifies measurable axes of linguistic misalignment and demonstrates that objective design can shift them independently of grammatical validity or task success (2610.29162).