Adapting for the future: How society can embrace advanced AI responsibly

The rapid diffusion of advanced AI capabilities presents challenges that traditional methods - like limiting AI capabilities or regulating diffusion - can no longer adequately address. As the cost of AI development decreases, more actors, including smaller organizations and individuals, gain access to these technologies.

As artificial intelligence advances and its applications expand, the risks it introduces grow increasingly complex. In their insightful paper, "Societal Adaptation to Advanced AI," Jamie Bernardi, Gabriel Mukobi, Hilary Greaves, Lennart Heim, and Markus Anderljung, from institutions like the Centre for the Governance of AI and Stanford University, argue for a shift in how society manages these risks. Available on arXiv, the study outlines a new framework for societal adaptation to advanced AI, emphasizing resilience over traditional, capability-focused interventions.

Why societal adaptation is crucial

The rapid diffusion of advanced AI capabilities presents challenges that traditional methods - like limiting AI capabilities or regulating diffusion - can no longer adequately address. As the cost of AI development decreases, more actors, including smaller organizations and individuals, gain access to these technologies. This democratization of AI increases the likelihood of misuse and unintended harm, making it impractical to solely rely on capability-modifying safeguards, which are not failsafe. For instance, safeguards like fine-tuning and output filtering can be circumvented, and leaked or stolen models can cause harm despite their original protective measures.

Societal adaptation offers a complementary strategy by preparing communities to manage the risks and consequences of AI deployment. It shifts focus from controlling AI capabilities to mitigating the harm they might cause, enabling society to reap the benefits of AI while minimizing risks.

A framework for AI adaptation

The study introduces a structured framework for adaptation, which addresses the causal chain of AI risks - from development and diffusion to use, harm, and impact. This framework categorizes adaptive interventions into three primary types:

The first category, avoidance interventions, focuses on reducing the likelihood of harmful uses of AI by making such actions more difficult or expensive. This can be achieved through measures such as enforcing strict laws against AI-driven election interference or instituting robust access controls to prevent unauthorized or malicious use of AI systems.

The second type, defense interventions, is aimed at minimizing the severity of harm if harmful AI use does take place. Organizations can implement these measures by deploying tools like anti-phishing software and conducting training programs to help employees identify and counteract AI-generated scams. 

The third category, remedial interventions, addresses the long-term consequences of harm that has already occurred. Examples include implementing compensation schemes for affected parties or developing robust recovery plans to ensure the rapid restoration of critical infrastructure disrupted by AI-related incidents.

The framework is applied to three key scenarios to illustrate its utility in managing AI risks. In election manipulation with generative AI, adaptive measures such as public awareness campaigns, laws against election interference, and transparent electoral investigations aim to combat the misuse of AI-generated synthetic media. For AI-enabled cyberterrorism, adaptation focuses on enhancing cyber defenses, fostering information-sharing networks, and building system redundancies to recover swiftly from attacks on critical infrastructure. Finally, in loss of control to AI decision-makers, measures like requiring human oversight, conducting rigorous audits, and regulating automation in sensitive areas are proposed to ensure high-stakes decisions remain aligned with human values and priorities. These examples demonstrate the framework's versatility in addressing diverse and complex AI challenges.

Building resilience through adaptation

At the heart of the study is the concept of resilience, defined as society’s ability to effectively adapt to the risks posed by advanced AI systems. Resilience is not a static state but a dynamic, ongoing process that involves three essential steps: identifying risks, evaluating potential adaptive responses, and implementing those responses effectively. This cyclical process ensures that society remains proactive in mitigating AI risks and responsive to emerging challenges.

A critical element of building resilience is shared awareness, which requires collaboration among governments, industries, academia, and the general public. Effective adaptation depends on a collective understanding of the risks associated with AI systems and the measures needed to address them. Governments play a pivotal role in promoting transparency and fostering dialogue between sectors, while public awareness campaigns can help individuals and organizations better grasp the implications of AI advancements. For instance, widespread education about the dangers of AI-driven election interference or cybersecurity threats can empower stakeholders to take precautionary actions. Without shared awareness, society risks fragmented and ineffective responses to AI-related challenges.

Another cornerstone of resilience is institutional capacity - the creation of organizations equipped with the technical expertise, financial resources, and authority to monitor and address AI risks. These institutions are vital for managing the rapid pace of AI development, which often outstrips the capacity of existing regulatory frameworks. Institutions must have the ability to assess AI risks, enforce necessary safeguards, and coordinate adaptive measures across sectors and borders. For example, a dedicated agency could monitor the deployment of AI in critical infrastructure, ensuring that defensive measures are in place to prevent misuse or accidents. Furthermore, robust institutional capacity allows for continuous evaluation and refinement of adaptive measures, ensuring they remain effective in a rapidly evolving technological landscape.

Resilience also necessitates international coordination, as many AI-related challenges, such as cyberterrorism or misinformation, transcend national borders. Establishing frameworks for global cooperation enables countries to share knowledge, pool resources, and implement consistent standards for AI governance. Without such coordination, disparities in regulatory approaches could undermine efforts to address global risks.

By fostering shared awareness, institutional capacity, and international coordination, society can build the resilience needed to adapt to the challenges of advanced AI.

Recommendations for Action

To enable effective adaptation to the challenges posed by advanced AI, the authors present a series of actionable recommendations aimed at governments, industries, and other stakeholders. A key priority is the need to enhance risk assessment by investing in research that identifies potential AI risks and evaluates their societal impacts. Governments and organizations must establish frameworks for robust oversight, ensuring that AI capabilities are deployed responsibly. This proactive approach enables early identification of threats and facilitates the development of informed strategies to address them.

Another essential recommendation is to promote AI literacy among the public, decision-makers, and industry leaders. By increasing awareness and understanding of AI’s capabilities and risks, public education campaigns can empower individuals and institutions to engage with AI responsibly. Greater AI literacy helps demystify the technology and equips society with the tools to critically assess its implications.

The authors also emphasize the importance of developing defensive AI systems. Companies are encouraged to create tools designed specifically to mitigate AI-driven threats, such as cybersecurity solutions capable of identifying and neutralizing vulnerabilities. Defensive AI not only enhances security but also fosters trust in the technology by demonstrating its potential to address the very risks it creates.

Another recommendation is to encourage staged AI deployment, a strategy that involves gradually releasing AI systems to allow sufficient time for risk assessment and the implementation of safeguards. This phased approach enables society to better understand how AI systems perform in real-world settings and adapt regulations and protections accordingly.

Finally, the authors highlight the value of establishing incident reporting mechanisms. Systems for reporting AI-related incidents, combined with whistleblower protections, can provide critical insights into the risks and impacts of AI. By documenting and analyzing these incidents, policymakers and organizations can refine their adaptive strategies, addressing vulnerabilities and preventing repeat occurrences.