top of page
Search

The Cohesion Trap: Why AI-Generated Text Reads Like a Textbook and What It Means for Language Learners

Der Bücherwurm" (The Bookworm) / Carl Spitzweg / 1850
Der Bücherwurm" (The Bookworm) / Carl Spitzweg / 1850

There's a peculiar quality to AI-generated writing that most readers sense but few can articulate. The prose flows smoothly, transitions appear logical, yet something feels mechanical – as if the text were designed for someone who needs every conceptual leap explained. This isn't coincidence. Large language models have been trained predominantly on explicit academic writing, student essays optimized for standardized tests, and instructional content where cohesive devices serve pedagogical rather than stylistic purposes.

The result is prose that treats readers as perpetual novices, signposting every turn with "Furthermore," "In addition," "On the other hand," even when the logical progression is self-evident. Human writers, particularly skilled ones, understand that cohesion often works best when invisible – when ideas connect through semantic proximity, shared lexical fields, or conceptual resonance rather than explicit markers.

Consider how international proficiency tests assess cohesion. Rubrics reward the presence of linking words and transitional phrases as concrete evidence of organizational skill. This creates a measurable standard but inadvertently establishes a baseline that AI models interpret as the default register for all writing. The training data skews heavily toward test-taking contexts where explicitness is rewarded, not the nuanced prose of literary essays, opinion columns, or sophisticated business communication where excessive signposting reads as condescending.

The linguistic term for this phenomenon relates to what Halliday and Hasan identified as textual cohesion versus coherence. Cohesion refers to the surface-level grammatical and lexical connections between sentences – the conjunctions, pronouns, and transitional phrases that explicitly link ideas. Coherence, by contrast, is the deeper semantic unity that emerges when ideas relate meaningfully without requiring explicit markers. A text can be cohesive without being coherent (all the right connectors but nonsensical content) or coherent without heavy cohesion (ideas flow naturally with minimal linking devices).

AI models excel at cohesion because it is pattern-based and formulaic. They struggle with coherence because it requires understanding conceptual relationships at a level that transcends pattern matching. When an AI writes "Moreover" or "Additionally," it has identified a structural slot for elaboration. When a human writer chooses to juxtapose two sentences without a connector, trusting the reader to infer the relationship, that choice reflects confidence in shared context and the reader's interpretive capacity.

This has direct implications for English language learners, particularly at advanced levels. If learners study predominantly AI-assisted materials or model their writing on AI outputs, they internalize a style that may score well on tests but sounds stilted in professional or creative contexts. The overuse of cohesive devices becomes a fossilized feature – a plateau where functional communication is achieved but stylistic sophistication remains out of reach.

Consider the difference between "The policy failed. Leadership changed" and "The policy failed. As a result, leadership changed." The first version relies on juxtaposition and allows readers to infer causation. The second spells it out. Neither is wrong, but the first assumes a more sophisticated audience and creates tighter, more dynamic prose. AI defaults to the second because explicitness reduces ambiguity and mirrors its training corpus.

The issue extends beyond mere style. When every conceptual shift gets announced with "However" or "Nevertheless," readers develop a dependency on these signals. They stop actively processing logical relationships and instead wait for explicit instruction on how ideas connect. This undermines one of the core skills that separates proficient from advanced language users: the ability to navigate implicit connections and construct meaning from context rather than overt markers.

The phrase "let's address the elephant in the room" exemplifies another dimension of this problem. It is not a cohesive device in the grammatical sense but rather a meta-textual gesture – a self-conscious acknowledgment that the writer is about to shift topics or confront something uncomfortable. AI models deploy these phrases as if they were structural necessities, borrowed from the informal register of blog posts and popular writing where they signal candor or relatability. In formal or sophisticated writing, they often read as performative rather than functional.

Human writers choose when to be explicit and when to be elliptical based on audience, purpose, and rhetorical effect. They might use heavy cohesion in instructional writing but sparse connectors in narrative or argumentative prose where momentum matters. AI lacks this contextual judgment because it has learned cohesion as a universal good rather than a tool to be deployed strategically.

For educators and learners working with platforms like Grandomastery (https://grandomastery.com), the challenge becomes clear: how do we develop the meta-awareness to recognize when cohesive devices strengthen communication and when they clutter it? Activities that demand rapid synthesis, such as those involving Random Abstractions (https://grandomastery.com/abstractions) or Random Because (https://grandomastery.com/because), push learners to construct logical relationships without relying on formulaic transitions. The cognitive effort of linking disparate concepts under time pressure builds the kind of associative fluency that allows writers to trust implicit connections.

This is not an argument against teaching cohesive devices. Learners need to understand how conjunctions, referential pronouns, and transitional phrases function. The problem arises when these devices become the primary or only strategy for creating textual unity. Advanced literacy requires recognizing that cohesion is one tool among many – that lexical chains, parallel structures, and thematic consistency can bind a text together just as effectively, often with greater elegance.

The broader implication touches on how AI is reshaping linguistic norms. If learners increasingly encounter prose that over-explains connections, that style becomes naturalized. Future generations may find economical, elliptical writing harder to produce or interpret. The dense, aphoristic style of writers like George Orwell or Joan Didion, where much is conveyed through what is left unsaid, could become less accessible if readers are conditioned to expect every relationship spelled out.

There is also a cognitive cost. Research in psycholinguistics suggests that moderate ambiguity and inferential demand enhance comprehension and retention. When readers must work slightly harder to connect ideas, they engage more deeply with the material. Over-signaled prose, like overly scaffolded learning tasks, can actually reduce cognitive engagement by removing the productive struggle that consolidates understanding.

For language learners aiming at C1 or C2 proficiency, the task involves unlearning some of what standardized tests reward. It means recognizing that "In conclusion" is not always necessary if the concluding nature of a paragraph is evident from its content. It means experimenting with sentence juxtaposition, allowing semantic relationships to emerge from proximity rather than announcement. It means developing the confidence to assume an intelligent reader who does not need every interpretive step laid out.

This shift requires exposure to writing that exemplifies nuanced cohesion – essays, journalism, and literature where ideas flow without constant signposting. It also benefits from activities that disrupt formulaic thinking, forcing learners to construct coherence through conceptual rather than lexical means. The Grandomastery framework (https://grandomastery.com), by emphasizing bisociation and integrative thinking, creates conditions where learners cannot fall back on stock transitions. When asked to connect unrelated abstractions or build narratives from randomized prompts, they must rely on semantic ingenuity rather than mechanical linking.

The ultimate goal is not to eliminate cohesive devices but to develop register awareness – the ability to modulate explicitness based on context. A legal brief may warrant heavy use of "Furthermore" and "In contrast" because precision and clarity are paramount. A personal essay or creative piece may thrive on subtler connections. Advanced language users command this range. They understand that writing for clarity is not the same as writing that assumes reader incompetence.

As AI continues shaping how people encounter and produce English, these distinctions will matter more, not less. The ability to write with economy, to trust readers to follow unstated connections, to vary textual strategies based on rhetorical purpose – these are markers of linguistic maturity that no amount of cohesive device deployment can substitute. For learners and educators alike, the challenge is recognizing when AI's default mode serves as a useful scaffold and when it becomes a constraint that must be transcended.

The language you develop in dialogue with AI may be functional, but functionality alone does not equal sophistication. True fluency includes knowing when to let ideas speak for themselves, when to withhold the obvious transition, when to trust that meaning will emerge from juxtaposition rather than announcement. This is the kind of linguistic judgment that separates competent communication from compelling prose – and it is precisely what gets lost when every sentence relationship comes pre-explained, courtesy of an algorithm trained on test essays rather than the full range of human expression.


Comments

bottom of page