The Path to Safe, Ethical AI: SRI Highlights from the 2025 IASEAI Conference in Paris
SRI Advisory Board member Geoffrey Hinton presented a keynote address remotely on February 6, 2025, for the inaugural International Association for Safe and Ethical AI (IASEAI) conference in Paris.
The mission at the Schwartz Reisman Institute for Technology and Society (SRI) is to deepen knowledge of technology, society, and what it means to be human. As the SRI community navigates the evolving landscape of artificial intelligence (AI), a critical question remains at the forefront: How can AI be not only effective but also safe and aligned with human values?
This question was a central focus of the February 6-7 International Association for Safe and Ethical AI (IASEAI) inaugural conference in Paris. The event, held in advance of the Paris AI Action Summit, brought together experts from academia, civil society, industry, media, and government.
With deep expertise in the risks and challenges of AI technologies, SRI had a strong presence at the conference. SRI Advisory Board member and Nobel Laureate Geoffrey Hinton was a keynote speaker, and Gillian Hadfield, SRI affiliate and professor of computer science at Johns Hopkins University, participated in a compelling panel discussion. Other major figures in the field were also in attendance, such as Turing Award winner Yoshua Bengio, IASEAI President and Distinguished Professor of Computer Science at UC Berkeley Stuart Russell, and MIT Physics Professor Max Tegmark.
The conference featured over 40 talks on key AI governance and development topics. Discussions covered global AI regulation, safety engineering, aligning AI with human values, and the impact of AI on public perception, especially regarding disinformation. Experts also highlighted the need for transparency in AI systems and explored ethical challenges in controlling real-world systems and AI agents.
What Is Understanding?
Geoffrey Hinton is well positioned to examine AI safety. In his keynote What Is Understanding?, he stressed the importance of the scientific community reaching a consensus on whether AI understands in a way similar to humans. This is a critical issue with major safety implications—without consensus on understanding, we risk misjudging AI’s abilities, either overestimating its control or underestimating its risks.
Hinton argued that the debate over AI’s capacity for understanding remains as divisive as the early disputes over climate change, before scientific consensus was reached. He began by outlining two competing schools of thought that have shaped AI research for decades: symbolic AI, which treats intelligence as rigid rule-based reasoning, and biologically inspired AI, which views intelligence as an emergent property of learning—much like human cognition.
Large language models (LLMs) process language like our brains, Hinton stated, reshaping meaning based on context, rather than storing sentences. Words, he said, are like "flexible Lego blocks" that adapt to their surroundings. Some critics call LLMs auto-complete. But Hinton countered: “That’s exactly how human memory works”—we reconstruct memories, sometimes inaccurately, but while preserving the overall meaning.
Despite these similarities, Hinton pointed out a key difference: AI’s unmatched ability to share knowledge. While humans pass information in small chunks, AI synchronizes trillions of bits instantly. “It’s no competition,” he said. If intelligence is about learning and sharing knowledge, AI is set to surpass us in speed and understanding. It’s a “very scary conclusion,” Hinton said—a warning that highlights the need for consensus on AI’s capabilities.
SRI is uniquely positioned to address the “scary” side of AI. The institute’s emphasis on interdisciplinarity is crucial to tackling the divides in AI research and ensuring comprehensive solutions for AI safety. The institute has long been at the forefront of AI safety, with Hinton co-authoring a recent paper in Science on managing AI risks.
Watch SRI Advisory Board member Geoffrey Hinton’s keynote address for the inaugural International Association for Safe and Ethical AI (IASEAI) conference.
Strategic Foresight: The Urgent Need for Regulatory Infrastructure
One of Hinton’s co-authors on that paper was Gillian Hadfield, who spoke on the second day of the conference as part of a panel called "Strategic Foresight for Safe and Ethical AI." Moderated by Carnegie Mellon University’s Atoosa Kasirzadeh, the panel also featured the Director of the Machine Learning Department at Carnegie Mellon University Zico Kolter, Senior Researcher at the Oxford Martin AI Governance Initiative Toby Ord, and Executive Director at The Future Society Nicolas Moës.
During the discussion, which covered topics like the existential risks of AI and who decides when AI is “safe enough,” Hadfield argued for the urgent need to build new legal and regulatory infrastructures to manage AI's rapid advancement. Hadfield emphasized that we must be as innovative in the legal domain as we are in technology and move quickly, even without perfect information. “We don’t currently know what limits and requirements and standards we should be putting in place,” Hadfield said. “But what we can be doing and must be doing right now is laying the groundwork [and] the capacity for governments to respond and act.”
SRI’s interdisciplinary approach is crucial in addressing Hadfield’s concerns. At the institute, technical experts collaborate directly with policymakers, social scientists, legal scholars and more—fostering a rare integration of perspectives. This collaborative environment is essential for developing the regulatory infrastructure that Hadfield advocated, ensuring that AI protections evolve alongside technological advancements.
Also of note from day two was a talk by Evi Micha, former graduate affiliate at SRI. Titled “Axioms for AI Alignment From Human Feedback,” the talk covered reinforcement learning from human feedback (RLHF), where AI systems learn from human comparisons. Micha explained that this process involves aggregating preferences, related to social choice theory, and demonstrated that the Bradley-Terry-Luce Model fails basic principles. New rules for learning reward functions were proposed, introducing the concept of "linear social choice" to improve AI alignment.
An Urgent Call to Action
On the final day of the conference, IASEAI issued a call to action for lawmakers, academics and the public, which was derived from the research presented and the deliberations held during the conference. The ten-point statement was designed to help further IASEAI’s mission to “ensure that AI systems are guaranteed to operate safely and ethically,” and includes calls for recognition of the significance of new developments in AI, prevention of AI-driven institutional and social disruption, and an increase in publicly funded research.
Highlighting the need for action and the importance of the ideas presented at the conference, Stuart Russell said: “The development of highly capable AI is likely to be the biggest event in human history. The world must act decisively to ensure it is not the last event in human history. This conference, and the cooperative spirit of the AI Summit series, give me hope; but we must turn hope into action, soon, if there is to be a future we would want our children to live in.”
SRI echoes Russell’s sentiment and is committed to fostering deeper knowledge of how technological advancements like AI impact human society.
Want to learn more?
Watch Geoffrey Hinton’s talk “What is Understanding” here.
Watch videos of the conference presentations and panels here on IASEAI’s website.
