ScienceAI

AI Is Transforming Science — but Does It Understand Any of It? | with Claire Malone

Royal Institution
Royal Institution•Jun 12, 2026

Why It Matters

AI’s predictive power can dramatically speed discovery, but without aligning with scientific standards of falsifiability and explanation, it risks reshaping research into pattern‑finding rather than insight‑driven inquiry.

Key Takeaways

  • •AI accelerates data analysis, boosting efficiency across scientific fields.
  • •Generative models predict patterns but lack true understanding of meaning.
  • •Science demands falsifiable hypotheses; AI must align with verification standards.
  • •Transformers enable large-scale pattern extraction, transforming experimental workflows.
  • •Integrating AI raises philosophical questions about the nature of scientific insight.

Summary

The talk explores how artificial intelligence, especially generative models and transformers, is reshaping scientific research—from particle physics at CERN to climate modeling—while questioning whether machines truly grasp the meaning behind their predictions. Claire Malone frames the discussion with a nod to Douglas Adams’ Deep Thought, warning that answering the "ultimate question" without understanding the question itself could mislead science.

She outlines how AI tools have moved from supervised learning, where models are guided by labeled data, to unsupervised and deep‑learning approaches that discover hidden structures without explicit instruction. The transformer architecture, introduced in 2017, is highlighted for its self‑attention mechanism that predicts the next token, pixel, or data point at scale, enabling rapid pattern detection across massive datasets.

Concrete examples include a live sorting exercise that mirrors unsupervised learning, a probability‑wheel demo illustrating how language models generate text by sampling likely continuations, and the contrast between physics‑driven hypothesis testing and AI’s data‑driven pattern extraction. Malone emphasizes that while AI excels at prediction, it lacks the capacity for explanation, intent, or falsifiable hypothesis generation.

The implications are profound: scientists must decide whether AI merely accelerates existing methods or fundamentally redefines what counts as understanding. Integrating AI demands new standards for transparency, reproducibility, and philosophical clarity about the role of machines in the scientific method.

Original Description

What does it mean to make a scientific discovery — and can a machine really do it?
📅 Filmed at the Ri on 2 May 2026
AI is now simulating particle collisions at CERN, predicting protein structures that stumped biologists for decades, and proposing experiments that human researchers never thought to run. But here's the unsettling question underneath all of it: what if these systems are generating correct answers without understanding why those answers are true?
In this talk, particle physicist and science journalist Claire Malone cuts to the heart of one of the most pressing questions of our time. Drawing on her own experience at CERN — searching for supersymmetry at the world's most powerful particle accelerator — she unpacks how today's most advanced AI systems actually work, from the transformer architecture powering large language models like ChatGPT, to the generative diffusion models now being used to simulate the ATLAS detector at the Large Hadron Collider tens of times faster than traditional physics-based methods.
____
Subscribe to the Royal Institution for weekly talks from the world's leading scientists and thinkers ► http://bit.ly/RiSubscRibe
Become a Royal Institution member and support our work ► https://www.youtube.com/channel/UCYeF244yNGuFefuFKqxIAXw/join
____
Claire examines the philosophy that underpins science itself; from the Vienna Circle's demand for verifiability to Karl Popper's principle of falsification — asking whether systems that operate on statistical pattern-matching rather than causal understanding can ever truly meet those standards. With AlphaFold solving a 50-year problem in molecular biology and anomaly-detection algorithms at CERN flagging physics that our theories cannot yet explain, the stakes of these questions have never been higher.
Claire Malone is a science journalist based in London, a contributing columnist for Physics World, and the STEM Lead for the Lightyear Foundation. She holds a PhD in particle physics from the University of Cambridge, where her research focused on developing novel techniques to search for evidence of supersymmetry beyond the Standard Model. Her 2021 TED talk has been viewed nearly 2 million times.
___
The Royal Institution is a charity. Talks like this one exist because of the generosity of people who believe science should be accessible to everyone. If this video sparked something in you and you're in a position to give, a donation of any size makes a real difference ► https://rigb.my.site.com/membership/s/donate
___
Chapters:
0:00 Introduction: AI and the Future of Scientific Discovery
2:55 What Is Science? The Philosophy Behind the Method
8:06 How Machine Learning and AI Actually Work
16:22 How ChatGPT and Large Language Models Generate Text
24:13 AlphaFold: How AI Solved a 50-Year Biology Problem
30:33 How CERN Is Using AI to Analyse Particle Physics Data
37:27 The Nobel Turing Challenge: Can AI Win a Nobel Prize?
43:05 Why Current AI Can't Make True Scientific Discoveries
47:04 Is AI a Scientific Collaborator or Just a Powerful Tool?
___
Donate to the RI and help us bring you more lectures: https://www.rigb.org/support-us/donate-ri
Subscribe for the latest science videos: http://bit.ly/RiNewsletter
____
#AI #ArtificialIntelligence #MachineLearning #Science #CERN #ParticlePhysics #AlphaFold #GenerativeAI #LargeLanguageModels #ScientificDiscovery #DeepLearning #Physics #ChatGPT #royalinstitution
____
Product links on this page may be affiliate links which means it won't cost you any extra but we may earn a small commission if you decide to purchase through the link.

Comments

Want to join the conversation?

Loading comments...