Silvia Mirri

Artificial intelligence (AI) and large language models are powerful tools but come with risks and caveats for proper usage. In this work, we are concerned with effective human-AI collaboration. As in any team, it is key that all team members properly understand the problem at hand, the environment, and each other. In order to scaffold a mature human-AI collaboration, we propose a process framework based on continual, proactive, bidirectional assessment, fostering co-evolution of the human-AI team. While most research has focused on assessing the strengths and limitations of AI agents, we especially focus on the risks associated with a lack of maturity on the human side. The core idea is to have AI agents help humans assess themselves and to anticipate interaction issues. Furthermore, to place the proposal within a broader vision, we advance two novel concepts: (i) “AI-in-the-human-loop”, the idea of AI agents observing and acting within human reasoning and activity, and (ii) “human-under-test”, a software testing analogy that suggests techniques for assessing human maturity by AI agents. A case study of a human-AI team involved in a statistical analysis project is considered, and analysed through an LLM-as-a-judge setup. Preliminary results highlight collaboration maturity issues and the possibility of proactive corrections, showcasing limitations in current collaborative workflows and opportunities for further research.

User-Centered Design (UCD) relies on the active involvement of end users throughout the design process. However, in complex socio-technical domains, continuous user participation is often difficult to sustain due to time constraints, safety requirements, confidentiality issues, high workload, or limited access to operational environments. This conceptual paper proposes an Agentic UCD framework based on Large Language Model (LLM) User Twins (UTs) to address this user involvement gap. UTs are proposed as generative, interactive, adaptive, and agentic representations of target users or user groups, grounded in user research, domain knowledge, task descriptions, scenarios, and design constraints. Unlike traditional personas, which make user knowledge explicit but remain static, UTs can be queried, updated, and reused across design activities. Within the proposed framework, UTs support design exploration, implementation, synthetic evaluation, and preparation for human validation by acting as role-oriented agents that simulate user perspectives, critique design alternatives, generate scenarios, and identify potential usability, trust, workload, or accessibility issues. The paper emphasizes that UTs do not replace empirical user research, co-design, or usability testing. Their outputs should be treated as simulated feedback and design hypotheses, not as empirical evidence. Instead, UTs are positioned as a complementary mechanism for maintaining continuity in user-centered reasoning between moments of direct user involvement.

The increasing adoption of Large Language Models (LLMs) in software development has enabled automatic code generation from natural language, yet the influence of communicative factors such as prompt tone remains underexplored. This work introduces PromptTone, a controlled dataset designed to investigate how variations in prompt politeness affect LLM-based code generation in web development. The dataset is constructed through a structured experimental design combining three variables: programming paradigm (Vue.js Composition API vs. Options API), LLM provider (GPT, Claude, Gemini), and prompt tone (impolite, neutral, polite), resulting in 396 generated components across 22 implementations. Data were collected in an educational setting under a single-prompt constraint to capture first-shot model behavior, and are provided in both hierarchical and CSV formats, including prompts, generated code, and error annotations. Preliminary analysis reveals that prompt tone influences output characteristics such as verbosity, with model-specific patterns: for instance, some models exhibit increased output length with more polite prompts, while others remain stable. Differences also emerge across programming paradigms, suggesting an interaction between tone and code structure. These findings highlight that LLMs are sensitive not only to semantic content but also to pragmatic aspects of input. Overall, the dataset provides a novel benchmark for studying human–LLM interaction in code generation, supporting future research on prompt engineering, model evaluation, and socially-aware Artificial Intelligence (AI)-assisted development tools.

The scarcity of training data presents a fundamental challenge in applying deep learning to archaeological artifact classification, particularly for the rare types of Chinese porcelain. This study investigates whether synthetic images generated through Stable Diffusion with Low-Rank Adaptation (LoRA) can effectively augment limited real datasets for multi-task CNN-based porcelain classification. Using MobileNetV3 with transfer learning, we conducted controlled experiments comparing models trained on pure real data against those trained on mixed real-synthetic datasets (95:5 and 90:10 ratios) across four classification tasks: dynasty, glaze, kiln and type identification. Results demonstrate task-specific benefits: type classification showed the most substantial improvement (5.5\% F1-macro increase with 90:10 ratio), while dynasty and kiln tasks exhibited modest gains (3-4\%), suggesting that synthetic augmentation effectiveness depends on the alignment between generated features and task-relevant visual signatures. Our work contributes practical guidelines for deploying generative AI in archaeological research, demonstrating both the potential and limitations of synthetic data when archaeological authenticity must be balanced with data diversity.