The Faculty of Fine Arts of the Universitat Politècnica de València (UPV) today hosted the opening of the IV ValgrAI Scientific Council Forum (VSCF), an annual event that promotes dialogue between science, business and institutions to analyse the present and future of artificial intelligence. Organised by the ValgrAI Foundation, the forum brought together more than 300 attendees in a multidisciplinary day that focused on the responsible, transparent and transformative application of AI.
The institutional welcome was given by Ana Cidad, Managing Director of ValgrAI; Vicent Botti, General Director; and Inma Romero, Director of the Library, Documentation and Open Science Area of the Universitat Politècnica de València. All three agreed on ValgrAI’s commitment to artificial intelligence that not only advances technically, but also serves to improve the lives of people and the environments they inhabit.
The first keynote of the day was given by ValgrAI, UPV and University of Cambridge researcher José Hernández Orallo, one of the most authoritative voices in the field of artificial intelligence evaluation. In his talk, entitled ‘AI Evaluation with Explanatory and Predictive Power’, the UPV researcher stressed the need to go beyond traditional metrics – based on hits or static benchmarks – to move towards evaluation systems that explain the performance of models, predict their behaviour in new situations and allow them to be robustly compared with human capabilities.
Hernández Orallo has presented tools such as DeLeAn and ADeLe, which allow for a more granular analysis of the cognitive abilities exhibited by an AI, and to understand what kind of generalisation it performs. This perspective is not only crucial for designing safer and more efficient systems, but also for building more understandable and auditable AI. ‘We need to stop simply asking whether a system gets it right and start asking why it gets it right,’ he concluded.
The future of Artificial Intelligence
The block entitled ‘The Future of AI: ValgrAI Researchers’ was one of the most stimulating moments of the day, with the participation of ten pre- and post-doctoral researchers from Valencian universities, who shared cutting-edge projects with a high social impact, many of them developed in collaboration with medical, educational and industrial centres. These interventions demonstrated how the ValgrAI ecosystem is training a new generation of scientists with an applied, rigorous and committed approach.
Behzad Mehrbakhsh (UPV) opened the block with his study on the ‘Contamination Budget’, a theoretical model that analyses how different levels of noise or contamination in training data affect the performance of AI models. His research proposes an evaluation framework to understand the trade-offs between breadth, depth and difficulty in learning systems. Lucía Gómez (HealthTech-UPV) presented a multimodal depression detection system using voice, language and facial expression analysis in simulated conversations with virtual humans. Her proposal demonstrates how AI can help detect complex emotional states in a non-invasive and personalised way. Salvador López Barajas (UJI) shared the advances of his research in underwater robotics, entitled ‘From Perception to Grasp’, where artificial intelligence is applied to enable autonomous robots to detect, recognise and manipulate underwater objects. A key technical challenge for sectors such as ocean exploration, infrastructure maintenance or environmental conservation.
On the other hand, Carlos S. Galindo (VRAIN-UPV) presented an efficient solution for the automatic extraction of main content and comments in heterogeneous web pages, based on structural analysis of the DOM. His algorithm outperforms other existing methods in speed and accuracy, being of great use for feeding language models or data curation systems. Vicente Quiles (UMH) has impressed with NeuroRunner, a neurorehabilitative mind-controlled video game developed using a brain-machine interface based on EEG signals. Its use makes it possible to personalise cognitive and physical treatment in patients with neurological injuries, making rehabilitation more accessible and motivating; and Carlos Aliaga Torró (VRAIN-UPV) presented AMAIA-EPO, a multimodal adaptive architecture with intelligent agents for the optimisation of the educational process. The system acts as a pedagogical assistant that analyses student interactions and adjusts didactic resources in real time, thus reinforcing the personalisation of learning.
On the other hand, Carlos S. Galindo (VRAIN-UPV) has presented an efficient solution for the automatic extraction of main content and comments in heterogeneous web pages, based on structural analysis of the DOM. His algorithm outperforms other existing methods in speed and accuracy, being of great use for feeding language models or data curation systems. Vicente Quiles (UMH) has impressed with NeuroRunner, a neurorehabilitative mind-controlled video game developed using a brain-machine interface based on EEG signals. Its use makes it possible to personalise cognitive and physical treatment in patients with neurological injuries, making rehabilitation more accessible and motivating; and Carlos Aliaga Torró (VRAIN-UPV) presented AMAIA-EPO, a multimodal adaptive architecture with intelligent agents for the optimisation of the educational process. The system acts as a pedagogical assistant that analyses student interactions and adjusts didactic resources in real time, thus reinforcing the personalisation of learning.
In addition, Sergio Gómez (PHRLT-UPV) has introduced a Python toolkit for the automated evaluation of language models (LLMs), developed within the ARCHER project. The tool standardises benchmarking processes, improves transparency and allows to compare the performance of models in complex tasks in a reproducible way. Ana Sierra García (HealthTech-UPV) showed the results of a pioneering study on electrical and metabolic brain activity during motor imagination, both in healthy people and in patients with altered consciousness. Through the combined use of EEG and fNIRS, their system makes it possible to explore new avenues for neurological diagnosis without the need for motor responses, and Sandra Goizueta Berheide (HealthTech-UPV) has presented a project to assess covert awareness in patients with severe brain damage, using AI to analyse physiological responses to personalised emotional stimuli. Its method allows for the detection of cognitive activity signals that escape conventional clinical tests.
Rodrigo Barriga Porras (HealthTech-UPV) concluded the session with a retrospective study using machine learning to predict the level of functional independence in patients with traumatic brain injury or stroke. Based on variables collected during the first six months of rehabilitation, his model enables highly accurate forecasts of long-term outcomes and supports the planning of personalized treatments.
Following this, it was the turn of the companies, which highlighted their success stories in the application of AI. This segment showcased the growing maturity of AI as applied in key sectors for the economy and social well-being.
Luis Búrdalo, representing S2 Grupo, delivered a talk in which he warned about the increasing digital threats in a context where AI can be both part of the problem and part of the solution. He presented how S2 Grupo uses AI to detect cyberattacks in real time, automate incident response, and strengthen the resilience of critical infrastructures through platforms such as GLORIA and CARMEN. He emphasized that, despite the power of AI, the human factor remains essential.
Manuel Herranz, CEO of Pangeanic, advocated for a more democratic and sustainable AI model in his presentation “The AI That Works: Small Models for Big Problems.” In contrast to the dominant trend toward large, centralized models, Pangeanic is committed to Small Language Models (SLMs) trained for specific tasks, offering better energy performance, technological autonomy, and greater privacy.
From Nunsys Group, Katherin Orozco introduced NARRATE: Intelligent Manufacturing Custodian, a platform that transforms factories into intelligent organisms capable of adapting in real time to changes in demand, logistical disruptions, or emergencies. This solution combines AI, big data, digital twins, and IoT to enhance the resilience, sustainability, and energy efficiency of European industry. Use cases in sectors such as semiconductors and additive manufacturing illustrated its potential.
In the field of water management, Cristina Baixauli from Hidraqua demonstrated how AI is already being used in critical operations to optimize processes and reduce environmental impact. Her presentation, “Artificial Intelligence at the Service of Sustainability in Water Management,” covered everything from predictive models to anticipate turbidity or leaks, to computer vision systems that detect solid waste or issue alerts for marine intrusions in coastal treatment plants.
José Vicente Cercós, from Idrica, presented the project “From Data to Decisions,” based on the Xylem Vue platform powered by GoAigua. This solution integrates real-time data from sensors, SCADA, and GIS to deliver advanced predictions, prescriptive analytics, and automated decision-making across the entire water cycle. Cercós highlighted how this approach is helping to reduce losses, prevent flooding, and anticipate critical scenarios with high accuracy.
Lastly, Pablo Fernández, Head of Innovation at Baleària, outlined the shipping company’s commitment to sustainable digital transformation and detailed how they are applying AI to route optimization, energy consumption forecasting, emissions control, and preventive maintenance, as well as in the development of digital twins for their vessels. He also introduced the Baleària AI Lab as an open innovation ecosystem involving universities and research centers.
Gottlob Closes the Day with a “Psychoanalysis” of ChatGPT
The second keynote of the day was delivered by renowned researcher Georg Gottlob (University of Oxford and University of Calabria), who offered a critical yet constructive perspective on generative language models like ChatGPT, under the provocative title “Psychoanalysis (and Therapy) of ChatGPT.”
Gottlob demonstrated how these models, despite their astonishing linguistic capabilities, exhibit systematic errors, undetected biases, and a tendency to “hallucinate” information—fabricating facts or references. He questioned the blind trust sometimes placed in such systems and introduced a solution: Chat2Data, a system of intelligent agents that breaks down complex tasks into verifiable subtasks, with built-in mechanisms for review, cross-validation, and traceability. This approach aims to align AI advancements with the demands for truthfulness and control in scientific, legal, or administrative settings.
The event concluded with a roundtable discussion moderated by Ramon López de Mantaras (CSIC), featuring international figures such as Carme Torras (CSIC), Tom Dietterich (Oregon State University), and Georg Gottlob himself. The session focused on the scientific, ethical, and social challenges posed by AI, addressing topics ranging from transparency in research to the urgent need for strong and democratic governance of these technologies.
