Engineering Trustworthy Agentic Intelligence | Francisco Herrera | ADIA Lab Symposium 2026

Value of the Information & Strength of the Argument

The talk provides a valuable conceptual synthesis, moving beyond conventional trustworthy AI definitions to address the specific challenges posed by autonomous agents. The argumentation is well-structured: it establishes the problem (reliability gap), proposes a three-dimensional framework, and supports it with illustrative studies. The notion of ‘residual human states’ is a thoughtful contribution, highlighting a neglected aspect of human-AI interaction. However, the presentation is largely programmatic, and the empirical evidence cited is drawn from preliminary studies with students, not from deployed clinical systems. The argument would be stronger with more concrete data or validation of the proposed framework.

Scientific Rigor, Source Quality, Title Accuracy

Herrera is a senior professor and director of DaSCI, giving the talk high credibility. He references European and American trustworthy AI guidelines, his own publications (e.g., on habituation), and two studies by colleagues (one in International Journal of Human-Computer Studies, one in arXiv). However, specific citations are not systematically provided in the slides or description; the only URL is the symposium page. The title is accurate and matches the content well. The transcript is highly imperfect (many transcription errors), but this does not affect the substance of the reasoning.

Le titre reflète précisément le contenu : l'ingénierie de l'intelligence agentielle fiable, depuis la fiabilité des modèles jusqu'à l'assurance continue, avec un accent sur l'évaluation de l'état résiduel humain.

8/10

High-level expert presentation from a recognized academic, introducing a coherent framework with three pillars (reliability, verifiability, continuous assurance) and a novel concept of residual human states. Claims are supported by references to own and colleagues' ongoing studies, though no detailed empirical data are presented in the talk.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 01:00Introduction to the talk: background on agentic AI and the importance of human interaction.
• 03:00Presentation of the three fundamental dimensions: reliability, verifiable AI, and continuous assurance.
• 07:00Discussion on the shift from classical trustworthy AI to agentic scenarios, highlighting the need for new evaluation criteria.
• 10:00Introduction of the concept of verifiable AI, extending explainability with evidence, traceability, and causality.
• 14:00Explanation of the interactive human verification loop and thresholds for autonomous action.
• 18:00Focus on continuous assurance: monitoring, performance evaluation, and dynamic reconfiguration of guardrails.
• 23:00Introduction of the 'residual human states' concept and its importance for maintaining human oversight.
• 28:00Presentation of the first study: 86 students, AI-assisted vs classic training, error rate increase from 1% to 10%.
• 33:00Second study: effect of interactive explanation and confidence feedback on decision quality.
• 40:00Conclusions: need for verifiability, continuous assurance, and evaluation of human residual states; final question about trustworthiness if humans lose oversight ability.

• ADIA Lab Symposium 2026 — Official page of the symposium where the talk was presented.
• Habituation and AI (paper by F. Herrera) — Mentioned as a published paper discussing habituation in human-AI interaction.
• Study on AI training and medical students (International Journal of Human-Computer Studies) — Cited as the first experiment with 86 students showing increased error rates after AI-assisted training.
• Residual human states (arXiv preprint) — Mentioned as a paper currently under review introducing the concept of residual human states.