董汀：让国际人工智能治理成为人类迈向美好未来的契机

董汀

清华大学战略与安全研究中心助理研究员

导语

当地时间2023年11月15日，国家主席习近平在美国旧金山斐洛里庄园同美国总统拜登举行中美元首会晤。两国元首同意推动和加强中美各领域对话合作，包括建立人工智能政府间对话。这是人工智能治理首次进入中美高层沟通。本文认为，鉴于人工智能发展迅速且不可预测，以及全球范围内已有的治理路径各不相同，未来合作宜调整仅仅将技术作为权力工具的固有认知，寻找国家之间最广泛利益的动态共同点，实现人类与人工智能技术优雅而公正地共存。

Fears over AI dangers are hindering the chance to create a better future

图源《南华早报》

Last week, a breath of optimism swept through the news as the United States and China made headlines for encouraging reasons. The highly anticipated meeting between President Xi Jinping and US President Joe Biden concluded with both nations expressing a shared understanding across a broad spectrum of interests. This included not only traditional sectors such as the economy, trade and agriculture but also burgeoning fields such as artificial intelligence.

The discussion on AI seems to resonate with the appeal for intergovernmental cooperation that Henry Kissinger and Graham Allison proposed last month. Echoing those sentiments, there were hopes the two powers would agree on AI arms control or at least a consensus to exclude AI from nuclear command systems.

The enduring peace among nuclear nations since 1945 underlines a mutual recognition of nuclear perils. AI technology has amplified these risks, making their non-use a matter of course. However, the path to effective AI governance is complex and the theory of knowledge differs vastly from practical epistemology, especially in the unpredictable theatre of war.

AI is more than just a substitute for specific human functions. Its essence is shaped by the data it processes and the complexity of its algorithms. Unlike commercial applications, nuclear weapons in warfare cannot readily amass extensive training data. The confidentiality of enemy information and the dynamic nature of warfare further complicate data acquisition.

Even the most advanced AI algorithms fall short of emulating the nuanced decision-making of an experienced general who bases decisions on a wealth of professional knowledge and real-life experience. This differs significantly from current discussions about the military application of AI, often encapsulated in the “human in the loop” concept.

Algorithms and software are very intricate, so it’s difficult to decide who should be in which loop and how they should interact. Moreover, even in a case where China and the US reach an agreement, another conundrum arises. How could we verify compliance, and would any nation permit external inspection of their nuclear launch system’s software codes?

The discourse on AI risks extends beyond military applications. The prevailing debate in China suggests that technological development is often equated with safety and security. AI, while a captivating and succinct term, isn’t a single technology but a fusion of various technologies.

As such, it’s not appropriate to discuss AI as a generic whole in terms of governance needs. Many AI types and applications don’t threaten human safety and should be actively developed. However, particular safety risks arise at the frontier of AI, a concern endorsed by the 29 signatories of the Bletchley Declaration at the UK AI Safety Summit earlier this month.

Current discussions around AI governance focus on misinformation, intellectual property, privacy and virality. However, these challenges aren’t new. False information has existed in media such as newspapers, radio and television since their inception.

This leads to the question of whether the advancement of AI by leading firms can protect humanity from the very threats AI poses. In my classroom, when discussing topics such as energy, climate or public health, my socially conscious students often express dismay at corporations prioritising profits over the public interest.

But merely expressing outrage or attempting to shame companies during a few class sessions achieves little. Corporations exist through legal mandates and societal frameworks. What they need are better incentive mechanisms that reward beneficial actions and penalise harmful ones.

What’s more, technology isn’t the source of all evil, and neither should it be a scapegoat for flawed human decision-making. If we outsource moral and political choices to intelligent machines, we ignore the warning of cybernetics pioneer Norbert Wiener – that these responsibilities will ultimately return to us.

From an international governance perspective, if there’s anything to learn from nuclear arms control, it’s that humanity is capable of figuring out how to manage emerging, unpredictable technologies. Given the rapid evolution of AI and the diverse governance approaches worldwide, our focus should shift to adaptable cooperation.

Rather than pursuing a one-size-fits-all standard or approach to governance, the true challenge lies in consistently finding the broadest common ground for ensuring safety amid these risks. This requires a dynamic balance of security, development, prosperity and fairness.

In the realm of technological governance, simplicity is a myth. The challenge with emerging technologies demands a comprehensive strategy involving skill development, robust support systems and cultural assimilation.

Take nuclear energy, automobiles, aviation and pharmaceuticals. These are not mere triumphs of innovation but milestones of a long, arduous journey towards safety and civility. This complex endeavour requires flexible, sustained collaboration across diverse stakeholders. The recent US-China consensus and declaration at the UK AI Safety summit herald promising starts.

However, there’s a crucial aspect eluding our current discourse on AI governance. The rush to establish rules for managing AI’s risks is predicated on the assumption that AI will soon outpace human intellect.

But can an entity surpassing human intelligence ever be fully subjugated by human control? This might not be explained simply by a path-dependent model of technological governance but rather by a deeper reflection that our perception of technology as a tool of power has obscured its true nature.

We’ve been so captivated by AI’s ability to empower specific people and institutions to create new forms of dominance that we might be overlooking an unprecedented opportunity. AI could be the answer to age-old aspirations to forge a future where we coexist with technology in a world of grace and justice.

本文2023年11月25日首发于《南华早报》网站

向上滑动阅览

相关阅读