Topic Modeling of the AI Ethics, Generative AI, and Large Language Model Literature with Latent Dirichlet Allocation: Methodology and Preliminary Results
Abstract
The rapid diffusion of generative artificial intelligence (AI) and large language models (LLMs) has produced a fast-growing, interdisciplinary body of scholarship on their ethical, social, regulatory, and safety implications. This paper presents the methodology and preliminary results of a Latent Dirichlet Allocation (LDA) topic model applied to a Scopus corpus of 20,000 records published between 2018 and 2026, of which 19,505 documents were retained after screening. Titles and abstracts were cleaned with a deterministic pipeline combining lemmatization, bigram detection, and four layers of stopword filtering. A 27-topic model with an asymmetric document-topic prior and $ta = 0.5$ ($C_v = 0.576$; NPMI = 0.082) was fitted with Gensim. The topics separate into a technical layer and a normative layer, with publication volume heavily concentrated in 2025–2026 and originating predominantly from the United States and China.
Topic Modeling of the AI Ethics, Generative AI, and LLM Literature with LDA
1. Introduction
Scholarship on the ethics of artificial intelligence has evolved quickly. Early efforts concentrated on mapping and consolidating principles, such as the global landscape of AI ethics guidelines and ethical frameworks. Critical analyses of algorithmic systems were later followed by work on the specific risks of large language models (LLMs). The public release of widely accessible generative systems has accelerated this output: in the corpus analyzed here, 82.8% of the documents were published in 2025 and 2026.
At this scale, manual reading is no longer a viable way to characterize the field. Topic models treat text as data and offer a reproducible way to summarize large collections. Latent Dirichlet Allocation (LDA) remains the reference probabilistic model for this purpose.
2. Methodology & Corpus Construction
The corpus was harvested from Scopus using a targeted query combining terms related to generative AI/LLMs with governance, ethics, and safety terms for publications up to 2026.
Screening Summary
- Records identified: 20,000
- Excluded by length (< 20 words): 495
- Usable documents modeled: 19,505
- Countries represented: 86
- Distinct sources: 5,697
Text Preprocessing Pipeline
- Normalization: Lowercasing and removal of URLs, metadata, and editorial noise.
- Lemmatization: WordNet lemmatization preserving interpretable base nouns and verbs.
- Multi-layer Stopword Removal: Standard English stopwords, editorial terms, methodological terms, and generic domain stopwords.
- Phrasing: Bigram detection (
generative_ai,machine_learning,foundation_model). - Vocabulary Filtering: Token frequency pruned to words appearing in at least 39 documents and less than 50% of the corpus, resulting in a final dictionary size of 4,533 tokens.
3. Model Selection & Results
Across evaluated topic numbers $K \in [4, 40]$, optimal parsimony and semantic coherence were achieved at $K = 27$ with asymmetric document-topic prior $\alpha$ and topic-word prior $\eta = 0.5$.
| Metric | Value |
|---|---|
| Number of Topics ($K$) | 27 |
| Coherence ($C_v$) | 0.5761 |
| Coherence (NPMI) | +0.0818 |
| Top-10 Diversity | 0.7741 |
| Dominance | 0.0844 |
| Normalized Entropy | 0.9475 |
Major Thematic Blocks
The 27 identified topics cluster into five functional blocks across two overarching layers (Technical Layer vs. Normative Layer):
- Applications and Evaluation (30.1% of corpus): Focuses on clinical/healthcare AI, educational integration, benchmark performance measurement, and domain reviews.
- Technical Core (22.7% of corpus): Encompasses training/fine-tuning, retrieval-augmented generation (RAG), multimodal vision-language models, and foundation sequence learning.
- Governance and Ethics (20.1% of corpus): Focuses on systemic governance frameworks, generative AI societal challenges, bias/fairness in prompts, and adoption surveys.
- Interaction and Agents (18.0% of corpus): Agentic system architectures, user interaction, chatbot UX, and cognitive simulation.
- Risk, Security, and Policy (9.1% of corpus): Adversarial attacks, privacy/explainability, public policy, alignment, and legal compliance.
4. Key Findings
- Geographic Distribution: The United States (23.71%) and China (20.32%) combined account for 44.0% of the total literature analyzed.
- Temporal Shift: Biplot analysis shows a clear shift from foundational background papers (2022–2023) toward evaluating retrieval-augmented generation, agentic architectures, and benchmark security (2024–2026).
- Model Stability: Sensitivity analysis across random seeds yielded a mean Jaccard overlap of 0.319 and ARI of 0.241, indicating moderate stability where high-volume technical topics remain robust, while minor sub-topics should be considered exploratory.