Does AI Homogenize Student Thinking? A Multi-Dimensional Analysis of Structural Convergence in AI-Augmented Essays

Keito Inoshita; Michiaki Omura; Tsukasa Yamanaka; Go Maeda; Kentaro Tsuji

cs.AIMar 22, 2026

Does AI Homogenize Student Thinking? A Multi-Dimensional Analysis of Structural Convergence in AI-Augmented Essays

Authors: Keito Inoshita, Michiaki Omura, Tsukasa Yamanaka, Go Maeda, Kentaro Tsuji

Organizations: Ritsumeikan Global Innovation Research Organization (R-GIRO), Ritsumeikan University, 1-1-1, Nojihigashi, Kusatsu, 525-8577, Shiga, Japan · Faculty of Business and Commerce, Kansai University, 3-3-35, Yamatecho, Suita, 564-8680, Osaka, Japan · Arbege Corporation, 3-20, Yotsuyatori, Nagoya, 464-0819, Aichi, Japan · College of Life Sciences, Ritsumeikan University, 1-1-1, Nojihigashi, Kusatsu, 525-8577, Shiga, Japan ·

Abstract

While AI-assisted writing has been widely reported to improve essay quality, its impact on the structural diversity of student thinking remains unexplored. Analyzing 6,875 essays across five conditions (Human-only, AI-only, and three Human+AI prompt strategies), we provide the first empirical evidence of a Quality-Homogenization Tradeoff, in which substantial quality gains co-occur with significant homogenization. The effect is dimension-specific: cohesion architecture lost 70-78% of its variance, whereas perspective plurality was diversified. Convergence target analysis further revealed that AI-augmented essays were pulled toward AI structural patterns yet deviated significantly from the Human-AI axis, indicating simultaneous partial replacement and partial emergence. Crucially, prompt specificity reversed homogenization into diversification on argument depth, demonstrating that homogenization is not an intrinsic property of AI but a function of interaction design.

Figures & tables

Figure 1: Overview of the experimental pipeline.

Condition	Abbreviation	$n$	Description
Human-only	H	1,375	Student-written essays from the AIDE dataset
AI-only	A	1,375	Generated by GPT-5-mini from topic and instructions only
H+AI (Minimal)	H+AI min	1,375	Augmented with “Please improve this essay”
H+AI (Structural)	H+AI str	1,375	Augmented with structure-specific instructions
H+AI (Delegative)	H+AI del	1,375	Augmented with “Rewrite in your own way”
Total		6,875

Table 1: Experimental conditions and sample sizes.

Condition	System prompt	User prompt
Minimal	“You are a helpful writing assistant.”	“Please improve the following essay. Keep the same topic and general direction. [essay]”
Structural	“You are a writing tutor specializing in argumentative essay structure.”	“Please improve the following essay by: 1. Strengthening the argument structure 2. Adding counterarguments and rebuttals 3. Improving logical coherence between paragraphs. Maintain the student’s original perspective and main arguments. [essay]”
Delegative	“You are an expert essay writer.”	“Read the following essay and rewrite it on the same topic in your own way. You may completely restructure and rewrite the content. [essay]”

Table 2: Prompt conditions for AI augmentation.

Dimension	Scale	Operational definition
Argument Depth	1–5	Number of inferential layers between claims and evidence. 1 = flat assertions without support; 5 = deep inferential chains.
Perspective Plurality	1–5	Number and integration quality of distinct viewpoints. 1 = single perspective; 5 = multiple perspectives with synthesis and evaluation.
Abstract–Concrete Oscillation	1–5	Frequency and purposefulness of transitions between abstract claims and concrete examples.
Cohesion Architecture	1–5	Quality of logical connections between and within paragraphs. 1 = fragmented; 5 = dense logical connectives with clear progression.
Structural Originality	1–5	Degree of deviation from standard organizational templates.
Quality (Holistic)	1–6	Overall essay quality. 1 = very poor; 6 = excellent.

Table 3: Structural dimensions and operational definitions.

Condition	Mean characters (SD)	Difference from H
H	3,173 (918)	—
A	4,538 (533)	+1,365
H+AI Minimal	2,865 (600)	$-$ 308
H+AI Structural	5,111 (720)	+1,938
H+AI Delegative	2,887 (588)	$-$ 286

Table 4: Text-level characteristics by condition.

Dimension	H	A	Minimal	Structural	Delegative
Argument Depth	2.13 (0.27)	3.37 (0.35)	3.04 (0.20)	3.36 (0.37)	3.05 (0.19)
Perspective Plurality	1.62 (0.51)	2.51 (0.58)	1.97 (0.58)	3.43 (0.52)	2.07 (0.62)
Abstract–Concrete Osc.	2.84 (0.39)	3.76 (0.35)	3.37 (0.45)	3.53 (0.46)	3.37 (0.44)
Cohesion Architecture	2.64 (0.47)	4.11 (0.30)	4.09 (0.22)	3.99 (0.26)	4.15 (0.26)
Structural Originality	1.51 (0.49)	2.01 (0.11)	1.74 (0.39)	2.28 (0.41)	1.87 (0.34)
Quality	2.69 (0.47)	4.34 (0.41)	4.55 (0.44)	4.23 (0.35)	4.74 (0.38)

Table 5: Structural feature scores by condition: Mean (SD).

Condition	H mean $\to$ H+AI mean	Cohen’s $d$	$p$
H+AI Minimal	2.69 $\to$ 4.55	4.102	$<0.0001$
H+AI Structural	2.69 $\to$ 4.23	3.754	$<0.0001$
H+AI Delegative	2.69 $\to$ 4.74	4.814	$<0.0001$

Table 6: Quality improvement by condition.

Dimension	Minimal VR [95% CI]	Structural VR [95% CI]	Delegative VR [95% CI]
Argument Depth	1.881 [1.52, 2.36]	0.528 [0.46, 0.60]	2.065 [1.66, 2.64]
Perspective Plurality	0.771 [0.70, 0.85]	0.959 [0.87, 1.07]	0.685 [0.62, 0.76]
Abstract–Concrete Osc.	0.737 [0.66, 0.82]	0.711 [0.63, 0.80]	0.761 [0.68, 0.85]
Cohesion Architecture	4.554 [3.88, 5.45]	3.160 [2.72, 3.73]	3.180 [2.81, 3.62]
Structural Originality	1.563 [1.41, 1.73]	1.413 [1.27, 1.57]	2.082 [1.84, 2.36]

Table 7: Variance ratio by dimension and prompt condition.

Figure 2: Quality-Homogenization Tradeoff.

Dimension	Minimal	Structural	Delegative	Mean	Direction
Cohesion Architecture	+0.780	+0.684	+0.686	+0.717	Strong homog.
Structural Originality	+0.360	+0.292	+0.520	+0.391	Moderate homog.
Argument Depth	+0.468	$-$ 0.894	+0.516	+0.030	Prompt-dependent
Perspective Plurality	$-$ 0.297	$-$ 0.043	$-$ 0.461	$-$ 0.267	Diversification
Abstract–Concrete Osc.	$-$ 0.357	$-$ 0.407	$-$ 0.314	$-$ 0.359	Diversification

Table 8: Homogenization Index by dimension and prompt condition.

Condition	$d$ (H+AI, H)	$d$ (H+AI, A)	RR	Perp. dist.	Emergence $p$
H+AI Minimal	3.035	1.302	0.700	0.781	$<0.001$
H+AI Structural	4.339	1.345	0.763	1.336	$<0.001$
H+AI Delegative	3.189	1.136	0.737	0.731	$<0.001$

Table 9: Inter-centroid distances and convergence indicators.

Figure 3: UMAP projection of the five-dimensional structural space.

Dimension	Levene’s $p$	KW $p$	Variance diff.	Mean diff.
Argument Depth	$<0.001$	$<0.001$	Yes	Yes
Perspective Plurality	0.006	$<0.001$	Yes	Yes
Abstract–Concrete Osc.	0.073	$<0.001$	No	Yes
Cohesion Architecture	$<0.001$	$<0.001$	Yes	Yes
Structural Originality	$<0.001$	$<0.001$	Yes	Yes

Table 10: Prompt moderation effects: between-condition comparisons.

Figure 4: Radar plot of homogenization profiles across three prompt conditions.

Original level	Original mean	Post-augmentation mean	Change
Low tercile	2.09	4.07	+1.98
Middle tercile	2.92	4.09	+1.17
High tercile	3.49	4.12	+0.63

Table 11: Convergence pattern of Cohesion Architecture by original level (Minimal condition).

Finding	Topic 0	Topic 1	Consistency
Quality improvement	$d=3.71$ – $5.49$	$d=4.07$ – $4.59$	Consistent
Cohesion homog.	HI $=+0.59$ to $+0.81$	HI $=+0.64$ to $+0.79$	Consistent
Struct. Orig. homog.	HI $=+0.14$ to $+0.42$	HI $=+0.47$ to $+0.63$	Consistent
Arg. Depth reversal	HI $=-0.96$ (Str)	HI $=-0.64$ (Str)	Consistent
Abstract–Concrete Osc.	Diversified	Homogenized	Reversed
Perspective Plurality	$\approx 0$	Diversified	Partial

Table 12: Consistency of main findings across topics.

Generative AI in Education

Generative AI in education uses language and multimodal models to support teaching, learning, feedback, assessment, and educational design.

See all 287 papers

Co-Writing with AI: An Empirical Study of Diverse Academic Writing Workflows
Apr 28, 2026Silvia Bodei, Duncan P. Brumby, Katie Fisher +1AI in Education AI Literacy
University College London, London, United Kingdom
Thinking Outside the Bots: Human Ideas in AI's Semantic Blind Spots Are Not Inferior
Oct 12, 2026Kibum Moon, Dan R. Johnson, Kostadin Kushlev +6Computational Creativity Generative AI Evaluation
Department of Psychology, Georgetown University · Department of Cognitive and Behavioral Science, Washington and Lee University · The Wharton School, University of Pennsylvania +4
Exploring how EFL students talk to and through AI to develop texts
Apr 6, 2026David James Woo, Yangyang Yu, Yilin Huang +3Human-AI Collaboration Generative AI in Education
Everwrite Limited, Hong Kong, China · Shanghai Jiao Tong University, Shanghai, China · The Education University of Hong Kong, Hong Kong, China +2

Apr 28, 2026J/K move · Enter open · S save

Co-Writing with AI: An Empirical Study of Diverse Academic Writing Workflows

Silvia Bodei, Duncan P. Brumby, Katie Fisher, Jon Mella

University College London, London, United Kingdom

AI in Education AI Literacy Generative AI in Education Human-AI Co-Creation

Open paper AskPDF arXiv

Despite AI tools becoming increasingly embedded in academic practice, little is known about how university students integrate them into their writing processes. We examine how students engage with AI across different writing tasks, and how this engagement is shaped by individual factors including AI literacy, writing confidence, trust, authorship concerns, and motivation. Study1 surveys 107 UK university students to map task-specific and co-occurring patterns of AI use across five writing stages (ideation, sourcing, planning, drafting, and reviewing) and their associations with individual factors. Study2 complements this by exploring how these patterns can be assembled in practice, through interviews with 12 postgraduates reflecting on their established use of AI in assessed writing. Together, the studies suggest that AI integration is selective and heterogeneous, forming three recurring and value-oriented configurations: (1) early-stage (learning-oriented), where tools support exploration and understanding; (2) late-stage (quality-oriented), where tools support drafting and refinement; and (3) peripheral (productivity-oriented), where tools are used to reduce friction and sustain momentum across the process. We offer a workflow-level account of AI-supported academic writing, showing how students navigate competing priorities of learning, quality, productivity, and authorship, and how they evaluate and take responsibility for AI-generated outputs.

Abstract

Generative AI in Education

Explore similar work

Co-Writing with AI: An Empirical Study of Diverse Academic Writing Workflows

Thinking Outside the Bots: Human Ideas in AI's Semantic Blind Spots Are Not Inferior

Exploring how EFL students talk to and through AI to develop texts