The lecture, delivered by Professor Michael Wooldridge at the Royal Society upon receiving the 2025 Michael Faraday Prize, explores the current state of artificial intelligence (AI) and contrasts it with the expectations and promises of the past. Wooldridge begins by clarifying that his talk is neither an attack on contemporary AI nor a manifesto for returning to older approaches like symbolic AI. Instead, he aims to highlight both the remarkable achievements and the peculiarities of modern AI, particularly large language models (LLMs) such as GPT-3 and its successors. He notes that the progress in AI over the past decade has been astonishing, surpassing the expectations of even leading experts in the field.
Wooldridge illustrates the rapid advancement of AI by comparing early systems like Microsoft’s CaptionBot, which struggled with basic image recognition, to the capabilities of modern LLMs that can solve complex mathematical problems and generate sophisticated text. He emphasizes the concept of “emergent capabilities,” where AI systems display skills they were not explicitly trained for, such as understanding certain logical relationships or solving novel problems. Despite these breakthroughs, he points out that AI still makes bizarre and sometimes nonsensical errors, demonstrating a lack of true understanding.
A central theme of the lecture is the distinction between how people imagine AI works and how it actually functions. Wooldridge explains that, contrary to popular belief, LLMs do not “think” or solve problems in a rational, step-by-step manner. Instead, they generate responses by statistically predicting the most likely next word or phrase based on vast amounts of training data. This approach, enabled by the transformer architecture and massive computational resources, allows for impressive feats but also leads to unpredictable and sometimes illogical outputs.
Wooldridge discusses the limitations of LLMs, noting that they are neither rational minds nor sentient beings. They lack consistency, cannot distinguish between fact and belief, and are prone to “hallucinations”—confidently generating false information. He stresses that these models are not capable of revising their knowledge in the way humans do and are fundamentally disembodied, with no awareness of the world or passage of time. The tendency of users to anthropomorphize AI, treating it as a conversational partner or confidant, is a reflection of human psychology rather than the true nature of the technology.
In the Q&A session, Wooldridge addresses questions about the future of AI, the challenges of robotic intelligence, and the differences between human and machine learning. He suggests that the next major frontier is robotic AI—creating systems that can operate effectively in the physical world, a task that remains far more difficult than language-based tasks. Wooldridge concludes by reiterating that while today’s AI is a marvel of engineering and has transformative potential as a “cognitive prosthesis” to augment human intelligence, it is not the rational, sentient intelligence once envisioned. Understanding and safely harnessing these powerful yet peculiar systems is one of the greatest scientific and engineering challenges of our time.
