Institutional Repository of Vilnius University
A lexical comparison of ai-generated texts: investigating patterns across gemini, claude, and llama /
Abstract
dc:descriptionArtificial intelligence has become an important part of written communication, as various large language models are now widely used to generate texts for different purposes. AI-generated writing seems fluent and natural, but different models do not necessarily produce language in the same way. Examining lexical features of AI writing can provide a better understanding on what kind of language these systems produce. While previous research has compared AI-generated writing with human-produced language, less attention has been given to a comparison between different AI models. The present study therefore focuses on the lexical features of texts produced by Gemini 3 Flash, Claude 4.6 Sonnet, and LLaMA 4 Scout. The subject of this bachelor thesis is lexical diversity and lexical density. Lexical diversity refers to the variety of vocabulary used in a text, whereas lexical density shows the proportion of content words in relation to the total number of words. By analysing these two features the study seeks to determine whether the selected AI models differ in their lexical behaviours and whether written task type have an effect on their generated output. The aim is to compare the two chosen lexical features in texts generated by the three large language models. To achieve the aim, the following objectives have been set: 1. To review theoretical approaches to lexical diversity, lexical density, prompts, and AI-generated texts. 2. To analyse lexical diversity in the texts produced by Gemini 3 Flash, Claude 4.6 Sonnet, and LLaMA 4 Scout. 3. To analyse lexical density in the texts produced by Gemini 3 Flash, Claude 4.6 Sonnet, and LLaMA 4 Scout. 4. To compare the type-token ratio and lexical density of the texts produced by the three AI models across different writing tasks. The empirical part of the study is based on 54 AI-generated texts in English. The data was collected by using three sets of six prompts representing different task types: argumentative, narrative, descriptive, and expository writing. As the length of AI-generated texts is not always the same, all texts were trimmed to the first 245 words to make the comparison with TTR as reliable as possible. The study uses a quantitative comparative method for analysis. Lexical diversity was measured using the type-token ratio, while lexical density was calculated using Ure’s (1971) formula. For the type-token ratio calculations, Sketch Engine was used to obtain the number lowercased word types, to make the analysis case-insensitive. Lexical density was measured using the AnalyzeMyWriting website. The findings reveal that the three analysed models do not behave the same when giving outputs to various task types. Gemini 3 Flash produced the highest lexical diversity values, which indicates better vocabulary variation and less repetition. Claude 4.6 Sonnet texts contained the largest proportion of contents words and displayed the highest lexical density values. LLaMA 4 Scout had produced the lowest values in both measures. The analysis also showed that the lexical behaviours differed not only across models, but also across writing tasks. In the type-token ratio analysis Gemini 3 Flash reached its highest values in the expository tasks, while Claude 4.6 Sonnet and LLaMA 4 Scout performed best in the descriptive tasks. For lexical density, the highest results were with expository prompts, especially in the outputs generated by Gemini 3 Flash and Claude 4.6 Sonnet. The results of the study show that different large language models produce different lexical patterns even when responding to the same prompts. The findings also suggest that the type of the writing task plays an important role in AI-generated writing. This study contributes to a better understanding of how different AI systems use language.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jakštaitė, Justina,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308317055&prefLang=en_US