The rapid advancement of Generative Artificial Intelligence, particularly Large Language Models (LLMs), has significantly influenced contemporary software development practices through automated code generation, intelligent autocompletion, and algorithm synthesis. Despite their growing adoption in both academic and industrial environments, LLM-based code generation systems raise persistent concerns related to code correctness, reproducibility, reinforcement of suboptimal programming practices, and ethical and legal issues such as intellectual property ownership, attribution, bias propagation, and accountability. Addressing these challenges is essential for the responsible and reliable deployment of AI-assisted software engineering tools. This thesis proposes a multi-layered framework that integrates Large Language Models with Automated Machine Learning (AutoML) techniques to improve the robustness, reliability, and trustworthiness of machine learning code generated from natural language specifications. Natural Language Processing and Natural Language Understanding methods are employed to analyse user prompts and extract structured semantic representations, which are subsequently used as formal inputs to an AutoML pipeline. This integration enables systematic hyperparameter optimisation, model selection, and pipeline configuration, thereby reducing ambiguity in model specification and improving the correctness and adaptability of the generated code for real-world deployment. In parallel, the thesis addresses ethical considerations in AI-assisted programming by introducing an ethical prompt recommendation layer based on collaborative filtering. This recommender mechanism learns from historical user interactions and contextual information to suggest ethically aligned follow-up prompts, encouraging adherence to principles such as fairness, transparency, non-discrimination, and responsible use. A synthetic dataset is constructed to evaluate both the predictive performance and fairness characteristics of the proposed framework, supporting reproducibility and enabling comparative analysis with existing ethical oversight approaches. Furthermore, a dedicated filtering component is incorporated to detect and suppress potentially harmful, biased, or legally non-compliant outputs, including those that may violate intellectual property constraints or established regulatory guidelines. The experimental results demonstrate that the proposed framework achieves high predictive accuracy while maintaining practical feasibility and interpretability. Overall, this thesis contributes a novel and scalable approach that bridges technical innovation in LLM-based code generation with structured AutoML optimisation and explicit ethical governance, advancing the state of responsible and accountable AI deployment in software engineering.
Weyland: A novel and autonomous approach to developing machine learning, using artificial intelligence
Nelson, J. (Author). Feb 2026
Student thesis: Doctoral Thesis