Digital Economy Dispatch #185 -- A Safety-First Approach to Adopting AI-at-Scale

Digital Economy Dispatch #185 -- A Safety-First Approach to Adopting AI-at-Scale
26th May 2024

As enterprises increasingly leverage AI to drive innovation and operational efficiency, the imperative of AI safety becomes ever more pressing. AI has vast potential. But it also brings inherent risks that, if not properly managed, can lead to significant consequences for businesses and society at large.

As a result, it is essential that digital leaders and decision makers develop a deep practical understanding of AI safety to ensure safe AI deployment at scale. Where can they turn for help and guidance?

The Challenge of AI Safety

AI safety is a broad concept that encompasses the strategies and measures implemented to ensure that AI systems operate in a manner that is reliable, ethical, and aligned with human values. As organizations deploy AI technologies on an increasing scale, the stakes become higher. We are all becoming increasingly aware that unintended behaviours, biases, and security vulnerabilities in AI systems can lead to severe repercussions, including financial losses, legal liabilities, and damage to reputation.

The importance of AI safety is underscored by several high-profile incidents. For instance, biased AI algorithms in hiring processes have led to discriminatory practices, errors in AI-assisted driving systems have contributed to vehicles crashing, and flawed AI systems in healthcare have resulted in incorrect diagnoses and treatments. Such cases highlight the necessity for robust safety measures to prevent harm and maintain public trust in AI technologies.

Yet, how to address safety in AI is far from clear. Ensuring AI safety is a multifaceted challenge that involves a wide range of technical, ethical, organizational, human, and regulatory dimensions. Furthermore, unlike traditional software systems, AI systems learn and evolve from data, making their behaviour less predictable and harder to control. This complexity is compounded by the following factors:

1. Dynamic Learning: AI systems continually learn from new data, which can introduce new risks over time. Ensuring safety in such a dynamic environment requires ongoing monitoring and updating of safety protocols.

2. Openness and Interconnectedness: Modern AI systems often interact with various other systems and data sources. This interconnectedness increases the risk of unexpected behaviours due to the complex interplay of different components.

3. Autonomy: As AI systems gain more autonomy, their decision-making processes can become opaque, making it challenging to understand and mitigate potential risks.

4. Ethical Considerations: AI systems must align with ethical standards and human values, which can vary across different cultures and contexts. Addressing ethical concerns requires a nuanced approach that considers diverse perspectives.

Taking Steps Toward AI Safety

Recognizing these key concerns, governments and other institutions have directed a great deal of attention on bringing communities together to focus on AI safety. The UK emerged as a frontrunner in AI safety by hosting the first-ever global summit in November 2023. This has continued with a follow-up summit co-hosted by South Korea and the UK in Seoul in May 2024, with another summit planned for France later in 2024.

In addition, demonstrating further commitment, the UK also established the world's first AI Safety Institute in November 2023, aiming to solidify its position as a global leader in AI safety. The AI Safety Institute describes itself as the first government-backed organization solely focused on advancing safe and beneficial AI. It has a three-pronged approach: conducting cutting-edge research and building robust testing infrastructure to assess the safety of advanced AI; engaging the research community, AI developers, and other governments to influence responsible AI development practices; and contributing to shaping global policies on AI safety.

Other countries have followed the UK lead, notably in the US. Following their announcement in November 2023, the US launched its own AI Safety Institute in February 2024 as part of the National Institute of Standards and Technology (NIST). Its vision and objectives have many parallels with those of its UK equivalent.

Interim Report on AI Safety in Advanced AI

One of the first key documents from these AI safety coordination efforts has now been released. The “International Scientific Report on the Safety of Advanced AI” is the first step in bringing together experts from around the world to offer a consolidated view of AI safety issues.

This important international report, coordinated by Yoshua Bengio, brings together inputs from over 30 nations. Timed for the Seoul AI Safety Summit, the report is a critical resource for digital leaders and decision makers looking for a unified view on AI safety.

The full report is worth reviewing. It examines current and anticipated AI capabilities, explores potential risks, and proposes methods for mitigating and evaluating those risks. The report underscores the vital role of international collaboration in advancing AI research and understanding potential risks. This collective effort fosters a unified approach to AI safety, paving the way for responsible and beneficial development efforts around the world.

However, just as important as the details of the report, it also offers a valuable resource for digital leaders navigating the evolving landscape of AI development and in the process of defining their AI-at-Scale strategies. Briefly summarized, a few of the key aspects of the report to note include:

• Synthesis of Existing Research: The report acts as a comprehensive overview, summarizing existing research on frontier AI risks and identifying areas for further investigation. This provides a crucial foundation for informed decision-making.

• Focus on General-Purpose AI: The report emphasizes general-purpose AI due to its rapid advancement and scientific uncertainty. For digital leaders this is an important reminder of the potential future impact of this technology.

• Divergent Expert Opinions: While some experts anticipate a slowdown in progress, others foresee rapid development. This underscores the need for continuous assessment and adaptation of policies as this area evolves.

• Uncertainties and Societal Choices: The report acknowledges disagreements on AI's capabilities and risks due to varying views on mitigation strategies and their effectiveness. It emphasizes that societal and governmental decisions will ultimately shape the future of AI.

However, it also should be emphasized that this is an interim report on the state of AI. The report explicitly acknowledges its limitations due to time constraints. The authors plan to improve and expand the next report in several ways:

• Evaluating and summarizing research: They will consider more studies, assess their quality, and provide a more nuanced synthesis of the findings.

• Input: They will accept submissions from companies and civil society organizations, in addition to the limited input used for this interim report.

• Comprehensiveness: They will delve deeper into specific risks like the global AI divide and environmental impact, while maintaining a balance between covering a wide range of topics and providing in-depth analysis.

• Scope: They will further refine the definition of the type of AI the report covers.

As a result, staying in touch with the work of these institutions and engaging with subsequent reports will be an essential step for every digital leader to keep on top of the AI safety challenges they will undoubtedly need to address in deploying AI-at-Scale.

Safety First

As AI technologies continue to evolve and permeate every sector of industry and society, the importance of AI safety cannot be overstated. Digital leaders play a critical role in ensuring that AI systems are designed, deployed, and managed in a manner that prioritizes safety, fairness, and ethical integrity. By addressing the complex dimensions of AI safety, organizations can harness the transformative power of AI while minimizing risks and fostering trust among stakeholders.

The global focus on AI safety is critical. We should look to the work emerging from the AI Safety Institues appearing worldwide as important indications of the challenges to be addressed. The interim scientific report on the safety of advanced AI is an important first step. But much more work remains. As we move towards an AI-driven future, deepening our understanding and adopting a proactive and comprehensive approach to AI safety will be essential for a sustainable and responsible approach to delivering AI-at-Scale.