Human-Centered AI: User-Driven Adapted Language Models (HumanAI-LANG)

The HumanAI project, led by the HULAT-UC3M group in coordination with the NIL-UCM group, aims to address the challenges faced by people with cognitive disabilities when using generative Artificial Intelligence (AI) tools. This project focuses on promoting inclusivity and accessibility in generative AI, reducing the digital divide, and improving the quality of life for these individuals while maintaining a commitment to environmental responsibility.

The goal is to define a framework and methods for multimodal generative AI tools in Spanish, ensuring they are cognitively accessible, customizable, and easy to understand through active user participation throughout the project, from refining language models to designing user interfaces. Achieving these objectives supports the project's research hypothesis that advances in accessible and customizable user interfaces, alongside innovative AI and Natural Language Processing (NLP) approaches, can close the digital divide for people facing cognitive accessibility barriers when using new generative AI tools.

This HULAT subgroup project, HumanAI-LANG, will focus on tackling the challenges posed by the rapid progression of generative AI technologies, particularly those based on large language models (LLMs). The challenges include difficulties in prompt usage, understanding interaction content, and interpreting results. The aim of this subproject is to explore open-source LLMs and adapt them through instruction-based fine-tuning approaches, including people in the process.

Additionally, datasets (corpora) with human-feedback-driven instructions will be created for fine-tuning LLMs, particularly considering people with cognitive disabilities in Spanish—a language that has less representation in LLMs compared to English. This will allow generative AI models to be tailored to meet the needs of these users while generating textual content that complies with plain language and easy-to-read guidelines. The project will also integrate alternative communication methods, such as pictograms and visual elements, to enhance understanding of the obtained results.

Moreover, training and fine-tuning LLMs are energy-intensive tasks, and the operation of generative AI systems has raised environmental concerns.

This subproject has been awarded FPI funding, and a doctoral thesis titled Human-in-the-loop in LLMs fine-tuning in the context of cognitive impairments will be conducted.

Funding: €183,368.74