José Hernández-Orallo - 'La IA que no sabía decir que no sabía'

The talk provides a compelling and well-argued perspective on a critical issue in AI: the lack of epistemic humility in large language models. Hernández-Orallo, a leading researcher in AI evaluation, brings substantial expertise to the topic. He effectively illustrates the problem with concrete examples, such as the anagram case, where a model confidently provides a wrong answer. His argument that current training and evaluation procedures incentivize plausible but potentially incorrect responses is well-founded and supported by recent research, including his own Nature commentary. The talk is not a formal scientific presentation but rather an expert opinion lecture, which is appropriate for the venue. However, it lacks detailed methodological explanations and empirical data, which might be expected in a more technical setting. The speaker references several sources, including a recent OpenAI paper and his own work, but does not provide a comprehensive literature review. The title accurately reflects the content, and the talk successfully raises awareness about the importance of developing AI systems that can say ‘I don’t know’. Overall, the talk is insightful and valuable for a general audience interested in AI reliability, though it could benefit from more depth in certain areas.

The title accurately reflects the central theme: AI systems' inability to admit ignorance, a key aspect of reliability.

8/10

The speaker is a recognized expert in AI evaluation, with publications in Nature and other top venues. The talk is based on his own research and references to recent papers, but it is an opinion/lecture rather than a peer-reviewed study.

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

• 02:00Introduction: the problem of AI not admitting ignorance, analogy to 'cuñados'.
• 10:00Discussion of OpenAI's paper on why models hallucinate.
• 20:00Examples of models giving confident wrong answers, such as the anagram.
• 30:00Historical trend: models are becoming less likely to say 'I don't know'.
• 40:00Discussion of his Nature paper on the reliability of language models.
• 50:00The concept of 'ultracrepidarianism' and the need for epistemic humility.
• 1:00:00Proposal for a paradigm shift in AI design to include uncertainty.
• 1:10:00Conclusion: call for more science in AI to understand intelligence.

• Fundación Ramón Areces event page — Event page for the talk, providing context and registration information.

• OpenAI paper on why language models hallucinate — Referenced in the talk as explaining the causes of hallucination.
• Nature commentary by Hernández-Orallo et al. — Referenced as his own research on the reliability of language models.