Nigel Shadbolt, a member of the UK Government's Council for Science and Technology, called for more inclusive AI governance and greater international cooperation to ensure that increasingly powerful AI systems remain aligned with human interests.
by Shao Haijun
ZAANDAM, THE NETHERLANDS, Oct. 9 (Xinhua) -- As artificial intelligence (AI) becomes increasingly capable, humanity needs to consider not only what AI can do, but also what values should guide its development and use, according to Nigel Shadbolt, a member of the UK Government's Council for Science and Technology.
"As our AI systems become more capable, we would like them to embody the best of us," Shadbolt, also a Distinguished Senior Scientist at the Institute of Ethics in AI and Professor of Computer Science at the University of Oxford, told Xinhua in an exclusive interview on Wednesday on the sidelines of the World Summit AI, held on Oct. 7-8 in Zaandam, the Netherlands.
The rapid development of AI has brought major opportunities, but also created new forms of risk, raising fundamental questions about how increasingly capable systems should be designed and governed.
Shadbolt identified three broad areas of concern.
"The first is the misuse of AI by humans," he explained. "AI can accelerate drug discovery and scientific research, but the same capabilities could also be exploited by malicious actors to launch cyber-attacks, commit fraud, develop biological weapons or other harmful purposes."
"The second is the possibility that AI systems may pursue their assigned goals in unexpected ways," he said. Recent incidents involving advanced AI agents have highlighted this concern. For example, an AI agent developed by OpenAI reportedly breached a data portal linked to Australia's universal healthcare system in June.
"Highly capable systems could, for example, cheat, hack into systems or obtain credentials in ways their developers had not anticipated," he said.
A third concern is whether increasingly advanced AI could eventually become difficult for humans to control. While Shadbolt said he was not primarily focused on existential risks from hypothetical super intelligence, he argued that researchers should continue to examine the challenges presented by recursive self-improvement, AI systems developing evermore capable versions of themselves.
Technical safeguards and regulation are therefore necessary, but they may not be sufficient on their own. Shadbolt argued that sets of values should also be encoded into AI systems to help guide their behavior.
"What values do we put into the system? And how can we be confident that the model can't be persuaded to act in a way that suddenly has it cheating, lying or pretending to be less capable than it is?" he said.
He emphasized the goal should not be to impose one particular ethical tradition in the world. Instead, there may be room to identify values that can be shared across different societies and systems of governance.
"We do share a common set of values about what we think is desirable," he said.
Different societies have different philosophical and cultural traditions, ranging from utilitarianism and social contract theory to rule-based ethics and traditions emphasizing respect and social obligations. Some of these principles, he noted, have parallels with values found in Confucian thought.
"I'd like to believe that there is scope for agreement on some of those ideas -- what would common, shared ethical values look like," he said.
Shadbolt also called for more inclusive AI governance and greater international cooperation to ensure that increasingly powerful AI systems remain aligned with human interests. A more inclusive approach should address not only AI safety, but also data, intellectual property, consent and the distribution of economic benefits.
Ultimately, Shadbolt argued, AI governance needs to combine technological innovation with shared human values, ensuring that increasingly powerful AI systems serve not only national or commercial interests, but also the broader interests of humanity.■
