AI Snake Oil Authors Urge Big-Tent AI Safety Focus

In a new analysis, Princeton researchers Arvind Narayanan and Sayash Kapoor challenge the dominant narratives surrounding artificial intelligence safety. They argue that focusing excessively on existential risk, or x-risk, polarizes public debate and misdirects resources away from tangible, systemic threats. Instead, they propose a big-tent safety movement that prioritizes building resilience against immediate, AI-amplified dangers like cyberattacks and biological threats, rather than trying to prevent a hypothetical superintelligence.

To illustrate the global failure to prepare for catastrophic risks, the authors point to pandemic preparedness. A 2019 World Health Organization and World Bank report warned of a respiratory pathogen that could kill 50 to 80 million people and erase nearly 5 percent of the global economy, yet governments failed to spend the estimated 1 to 2 dollars per person annually required for defense. Since the 2009 H1N1 pandemic, at least 11 high-level panels made recommendations that went largely ignored, a pattern a 2024 independent panel called a "cynical attraction" to crisis spending over prevention. Similarly, in cybersecurity, markets fail to price tail risks; Lloyd's warned in 2022 that potential losses could far exceed what insurers can absorb.

The authors note that AI acts as an amplifier for these existing vulnerabilities. For example, Anthropic CEO Dario Amodei has warned that autonomous agent swarms could potentially take over the internet. Additionally, philosopher Atoosa Kasirzadeh has highlighted "accumulative risk," where gradual shifts, like chatbots eroding news media traffic, slowly decay societal institutions. Despite these dangers, a Gallup poll shows only 0.5 percent of the public views AI as the most important issue, highlighting a massive gap between industry anxiety and public salience.

For AI practitioners and policymakers, this shift to a big-tent framework changes how safety is operationalized. Instead of pursuing legally ambiguous bans on recursive self-improvement, the authors advocate for practical measures like transparency, liability, and embedded evaluation. This approach requires developers to build systematic accountability for negative externalities. By moving away from panic-driven policymaking, practitioners can focus on shoring up concrete defenses across hundreds of disparate systems rather than trying to build a single, flashy wall against an imaginary superintelligence.