TUG studies and designs the relationship between people and opaque artificial intelligence systems, so that it stays functional, traceable and auditable. It treats trust as infrastructure: something that can be observed, measured and audited, not merely declared.
Purpose
When an organisation brings AI into its decisions, the question is no longer how well a model performs. It becomes what relationship a person maintains with a system they do not fully control.
That relationship is the line’s object. TUG supports the creation of sociotechnical processes mediated by ICT, including generative AI, aimed at managing uncertainty and trust.
The line works on three planes that reinforce one another. It investigates how trust is formed, lost and recovered in AI-mediated environments. It develops artefacts and processes that make it operational. And it trains teams and organisations, as a first step towards building these capabilities in.
Approach
TUG separates three tasks that are often conflated: measuring trust, designing it and governing it.
Measuring means characterising an operator’s actual trust in an AI ecosystem, beyond what they state, and telling effective oversight apart from nominal oversight.
Designing means turning that measurement into decisions about how systems and the processes around them are built.
Governing means placing all of the above within its framework: the meaningful human oversight required by the European AI Act and the safeguards of the General Data Protection Regulation.
Instruments
The line works with its own instruments, which are at once research results and means of intervening in real processes.
ConfAI. A psychometric instrument that measures an operator’s actual trust in an AI ecosystem and tells effective oversight apart from nominal oversight. It serves as both a measurement and a design guide.
CHORUS. Orchestrates several heterogeneous language models and translates their disagreement into an interpretable index. That index marks where human scrutiny should be concentrated.
RAG and formalisation of expert knowledge. Anchor model outputs in verified sources. They underpin the knowledge management work with retrieval-guided LLMs developed for the Madrid Research and Innovation Area.
MEDLEY-SEC. The security strand of MEDLEY, the international collaboration’s system. It targets the detection of manipulation and data poisoning through disagreement between models. Reported in a publication in the Journal of Medical Internet Research on AI safety.
Scope in complex settings
Sociotechnical trust is tested in settings of high uncertainty and serious consequences.
The coordination from UPM-ETSIST of the European project NEURO-REBUILD (Horizon Europe) addresses the sociotechnical design of reconstruction processes in contexts such as natural disasters or conflict, combining multimodal measurement and digital twins. There, managing trust at a sociotechnical level is a condition for the process to work.
This capacity draws on the group’s track record in cybersecurity and cyberdefence, which brings first-hand knowledge of the environments where trust is put under strain: virtualised 5G environments for training, predictable and dependable industrial systems, cyber-situational awareness, and cyberdefence training through simulation. The group also sits on the advisory board of GUARDIANES, a network of excellence in AI technologies for security and defence.
Training
Training is one of the line’s avenues of work and, often, the first step an organisation can take.
The course NLP for Emotional AI, already delivered, works on improving labelling and on capturing tone and emotion in interactions, in order to adapt system discourse to the communicative situation. A course on Agentic Systems is in preparation, planned for September.
Ongoing research
There are ongoing doctoral theses on governance and uncertainty in AI systems and on the application of local language models in clinical settings.
This work builds on the group’s track record of doctoral supervision, including a thesis on the security and privacy of connected devices in the digital home, defended in 2023, and one on machine-learning-based intrusion detection for cybersecurity, defended in 2021.