Lourdes Moreno

LLM-generated interfaces are increasingly used in high-consequence workflows (e.g., healthcare communication), where how information is presented can impact downstream actions. These interfaces and their content support human interaction with AI-assisted decision-making and communication processes and should remain accessible and usable for people with disabilities. Accessible plain-language interfaces serve as an enabling infrastructure for meaningful human oversight. In these contexts, ethical and trustworthiness risks, including hallucinations, semantic distortion, bias, and accessibility barriers, can undermine reliability and limit users' ability to understand, monitor, and intervene in AI-supported processes. Yet, in practice, oversight is often treated as a downstream check, without clear rules for when human intervention is required or who is accountable. We propose oversight-by-design: embedding human judgment across the pipeline as an architectural commitment, implemented via escalation policies and explicit UI controls for risk signalling and intervention. Automated checks flag risk in generated UI communication that supports high-stakes workflows (e.g., readability, semantic fidelity, factual consistency, and standards-based accessibility constraints) and escalate to mandatory Human-in-the-Loop (HITL) review before release when thresholds are violated, or uncertainty is high. Human-on-the-Loop (HOTL) supervision monitors system-level signals over time (alerts, escalation rates, and compliance evidence) to tune policies and detect drift. Structured review feedback is translated into governance actions (rule and prompt updates, threshold calibration, and traceable audit logs), enabling scalable intervention and verifiable oversight for generative UI systems that support high-stakes workflows.

This paper describes the participation of HULAT-UC3M in CLEARS 2025 Subtask 1: Adaptation of Text to Plain Language (PL) in Spanish. We explored strategies based on models trained on Spanish texts, including a zero-shot configuration using prompt engineering and a fine-tuned version with Low-Rank Adaptation (LoRA). Different strategies were evaluated on representative internal subsets of the training data, using the official task metrics, cosine similarity (SIM) and the Fernández-Huerta readability index (FH) to guide the selection of the optimal model and prompt combination. The final system was selected for its balanced and consistent performance, combining normalization steps, the RigoChat-7B-v2 model, and a dedicated PL-oriented prompt. It ranked first in semantic similarity (SIM = 0.75), however, fourth in readability (FH = 69.72). We also discuss key challenges related to training data heterogeneity and the limitations of current evaluation metrics in capturing both linguistic clarity and content preservation.