On March 10, 2020, as the world grappled with the nascent stages of the COVID-19 pandemic, a stark realization dawned on one individual: the potential scale of the crisis was being profoundly underestimated. In a Stamford, Connecticut office, a colleague’s back-of-the-envelope calculation suggested that COVID-19 could claim a million American lives. This initial projection, though alarming, was a significant reduction from an even more staggering estimate derived from the available data. The raw numbers, based on a 1% to 3% fatality rate and a U.S. population of 330 million, hinted at a potential death toll of 2 million or more. However, the sheer magnitude of such a figure felt almost inconceivable, prompting a downward revision to a still-dire one million. This act of self-censorship, driven by a reluctance to sound alarmist or like a "Cassandra," was met with skepticism by an experienced entrepreneur and data analyst, who dismissed the possibility with a simple, "that’s not going to happen." Within three days, offices were shuttered, and by the end of 2022, the grim reality saw approximately 1.1 million Americans succumb to COVID-19, a number only contained by the rapid development and deployment of vaccines.
This unsettling echo of past underestimations has resurfaced with renewed urgency in the ongoing debate surrounding the potential dangers of artificial intelligence (AI). The parallels drawn to historical events, particularly the September 11th attacks and the initial response to the COVID-19 pandemic, highlight a recurring human tendency to falter in the face of unprecedented threats—a failure of imagination.
The Shadow of Unforeseen Catastrophes
The 9/11 Commission Report, a comprehensive investigation into the attacks of September 11, 2001, identified a critical failure at its core: a lack of imagination. Policy makers, intelligence agencies, and military leaders, despite possessing pieces of critical information, found it difficult to conceptualize a scenario where terrorists would weaponize commercial airliners to strike U.S. soil. The sheer audacity and alien nature of such a plan fell outside the realm of their established understanding and operational frameworks. This inability to connect disparate dots, to envision a threat so far removed from their lived experiences, ultimately contributed to the devastating outcome.
Similarly, the early days of the COVID-19 pandemic were characterized by an underestimation of its potential impact. The sheer scale of the projected fatalities—millions of lives lost—was difficult for many to process. This cognitive dissonance, this struggle to reconcile data with intuitive understanding, led to a downplaying of the threat, even among those with analytical backgrounds. The author’s experience of revising their own dire projection from over 2 million to 1 million deaths due to its perceived incredulity exemplifies this phenomenon. The immediate dismissal of the 1 million death estimate by a seasoned professional underscores how the sheer magnitude of a potential crisis can override rational assessment when it transgresses the boundaries of what is considered plausible.
AI: The Emerging Existential Risk
The specter of artificial intelligence, with its rapidly advancing capabilities, is now raising similar uncomfortable questions. Are we, in our current discourse and preparedness, repeating the mistakes of the past by underestimating the potential risks associated with AI? This question was brought into sharp focus at a recent PE Summit in New York, where Dan Glasier, former CEO of the global risk management giant Marsh McLennan, delivered a sobering assessment.
Glasier, a veteran with a lifetime dedicated to quantifying and managing risk, identified artificial intelligence as the preeminent risk facing humanity today. His assertion was stark: while trillions of dollars are being invested in combating climate change—a threat that may unfold over a century—virtually nothing is being done to regulate or contain the development of AI. This disparity, he argued, was "absurd."
The juxtaposition of these two monumental challenges—climate change and AI—highlights a potential imbalance in our risk mitigation strategies. Climate change, while an undeniable and severe threat, often presents a more tangible, albeit complex, timeline for its most catastrophic impacts. AI, on the other hand, possesses the potential for more abrupt and unpredictable shifts in global dynamics, driven by exponential technological advancement.
The Dual Nature of AI: Opportunity and Existential Threat
In a moment of attempting to steer the conversation towards a more optimistic outlook, Glasier was asked about the greatest opportunity in the world. His response, delivered with characteristic alacrity, was unequivocal: "AI, of course." This immediate pivot from existential threat to unparalleled opportunity encapsulates the complex and often contradictory nature of AI. It is a technology that holds the promise of solving some of humanity’s most pressing problems, from disease eradication and climate modeling to economic prosperity and scientific discovery. Yet, it simultaneously harbors the potential for unprecedented disruption and even existential peril.
This duality was keenly felt during the summit. The uncomfortable truth presented by Glasier, that AI represents the single greatest risk, lingered long after the session concluded. The author’s reflection on the periods preceding major historical shocks—the days between the USS Cole bombing and 9/11, the early warnings of the COVID-19 pandemic, or the lead-up to the attack on Pearl Harbor—serves as a potent reminder of how failures of imagination can precede catastrophic events. The question now is whether the current dialogue around AI safety, while growing, is sufficient to counter this inherent human tendency to underestimate the unthinkable.
Understanding the Scale of the Challenge: Supporting Data and Context
To fully grasp the potential risks associated with AI, it is crucial to consider the accelerating pace of its development and its integration into various facets of society.
- Exponential Growth in Computing Power: Moore’s Law, though debated in its current applicability, has historically described the exponential increase in computing power. This has fueled the development of increasingly sophisticated AI models. The training of large language models (LLMs), for instance, requires immense computational resources, with the largest models consuming energy equivalent to thousands of homes.
- Advancements in Machine Learning: Deep learning and other machine learning techniques have enabled AI to perform tasks that were once considered exclusively within the domain of human intelligence, such as image recognition, natural language processing, and complex decision-making. The performance of AI systems in areas like Go and protein folding has surpassed human capabilities, indicating a rapid trajectory of advancement.
- Ubiquitous Integration: AI is no longer confined to research labs. It is being integrated into critical infrastructure, financial markets, autonomous systems, and even military applications. The potential for cascading failures or unintended consequences in these interconnected systems is a significant concern.
- Economic and Societal Disruption: Projections by various economic forums suggest that AI could automate a substantial portion of current jobs, leading to significant labor market shifts and the potential for increased economic inequality if not managed proactively. The World Economic Forum, for example, has highlighted the need for reskilling and upskilling initiatives to prepare workforces for an AI-driven economy.
Historical Precedents and Their Implications
The parallels drawn to past underestimations are not mere rhetorical devices; they are grounded in historical analysis of how societies have grappled with novel and disruptive threats.
The COVID-19 Pandemic: A Case Study in Underestimation
- Timeline:
- Late 2019: Initial reports of a novel coronavirus emerge from Wuhan, China.
- January 2020: The World Health Organization (WHO) declares a Public Health Emergency of International Concern.
- March 2020: The virus spreads globally, leading to widespread lockdowns and significant societal disruption. The author’s personal calculation of a million potential deaths occurs.
- 2020-2022: The pandemic continues, with multiple waves and the emergence of new variants, resulting in millions of deaths worldwide.
- Underestimation Factors:
- Novelty of the Virus: The unprecedented nature of SARS-CoV-2 made it difficult to predict its transmissibility and severity.
- Underestimation of Global Spread: Early assumptions about containment proved overly optimistic.
- Cognitive Biases: A tendency to downplay threats that seemed too extreme or improbable.
- Data: Early estimates of the Case Fatality Rate (CFR) varied but were consistently high enough to suggest a catastrophic outcome if widespread infection occurred. The eventual CFR, while lower than initial worst-case scenarios, still resulted in an immense death toll.
The September 11th Attacks: A Failure of Intelligence and Imagination
- Timeline:
- 1998-2001: Al-Qaeda, led by Osama bin Laden, plans and executes attacks against U.S. interests, including the 1998 U.S. embassy bombings in Kenya and Tanzania and the 2000 USS Cole bombing.
- Early 2001: Intelligence agencies receive fragmented information suggesting potential attacks, but a clear picture of the intended method and target fails to emerge.
- September 11, 2001: 19 al-Qaeda terrorists hijack four commercial airplanes, crashing two into the World Trade Center in New York City and one into the Pentagon in Arlington, Virginia. A fourth plane crashed in Shanksville, Pennsylvania, after passengers attempted to regain control.
- Underestimation Factors:
- Unconventional Tactics: The idea of using airplanes as missiles was outside the conventional threat assessment framework.
- Information Silos: Lack of effective information sharing between different intelligence agencies.
- "Black Swan" Event: The attacks were a low-probability, high-impact event that was difficult for planners to fully anticipate.
- Data: While there were warnings of increased terrorist activity, the specific nature and scale of the 9/11 plot were not definitively identified or understood by relevant authorities.
The AI Debate: Emerging Concerns and Expert Perspectives
The current discourse around AI safety mirrors these historical anxieties. Leading figures in technology, academia, and risk management are voicing concerns that echo the underestimations of the past.
- Statements from AI Researchers and Ethicists: Prominent AI researchers, including those who have been instrumental in developing advanced AI systems, have signed open letters warning of potential existential risks. These warnings often center on the possibility of superintelligent AI developing goals misaligned with human values, leading to unintended and catastrophic consequences.
- Regulatory Challenges: Governments worldwide are beginning to grapple with the need for AI regulation. However, the pace of technological development often outstrips the legislative process. Debates revolve around issues of bias in AI, job displacement, autonomous weapons systems, and the potential for AI to be used for malicious purposes, such as sophisticated disinformation campaigns or cyberattacks.
- Industry Responses: While some tech companies are investing in AI safety research, critics argue that the drive for innovation and market dominance often overshadows genuine concern for long-term risks. The rapid release of powerful AI models without sufficient safety testing or societal preparedness is a recurring point of contention.
Broader Impact and Implications
The implications of underestimating AI risks are profound and far-reaching, potentially impacting every aspect of human civilization.
- Geopolitical Instability: An AI arms race could lead to increased global tensions and a heightened risk of conflict, particularly with the development of autonomous weapons systems. The absence of clear international norms and treaties governing AI in warfare creates a volatile environment.
- Economic Disruption and Inequality: Unmanaged AI-driven automation could exacerbate existing economic disparities, leading to widespread unemployment and social unrest. The concentration of AI capabilities and benefits in the hands of a few could further widen the gap between the haves and have-nots.
- Erosion of Truth and Trust: The ability of AI to generate hyper-realistic fake content (deepfakes) and spread disinformation at scale poses a significant threat to democratic processes and societal trust. The challenge of discerning truth from falsehood in an AI-saturated information environment is becoming increasingly daunting.
- Existential Threat: In the most extreme scenarios, misaligned superintelligent AI could pose an existential threat to humanity, either through direct action or by inadvertently causing catastrophic outcomes as it pursues its programmed objectives.
The historical pattern of underestimating unthinkable risks serves as a critical cautionary tale. As artificial intelligence rapidly advances, the question of whether we are once again falling victim to a failure of imagination becomes paramount. Addressing this challenge requires a proactive, multi-faceted approach that combines rigorous scientific inquiry, thoughtful regulation, and a collective willingness to confront the potential for outcomes that, while currently seeming "insane," may be tragically plausible. The time to engage with these risks is not when they materialize, but in the crucial moments of foresight, before the unthinkable becomes an irreversible reality.
