AI catastrophe may emerge from overlapping harms

Another major concern is disinformation. AI-generated content is becoming indistinguishable from human-authored narratives, enabling scalable, personalized propaganda campaigns that can erode democratic institutions, polarize societies, and compromise public trust in reliable information sources. Simultaneously, surveillance technologies, powered by facial recognition and data inference systems, threaten civil liberties by empowering state and corporate actors to monitor, manipulate, and suppress dissent on unprecedented scales.

A new study warns that a cluster of overlapping harms, rather than a single point of failure, may be the path through which artificial intelligence (AI) leads to existential catastrophe. The research, titled “A Polycrisis Threat Model for AI”, was published in AI & Society. Authored by Adam Bales of the University of Oxford, the paper introduces a detailed threat model arguing that interacting AI-driven harms such as disinformation, cyberattacks, labor disruption, and power concentration may collectively undermine societal stability and humanity’s long-term potential.

Rather than focusing on traditional AI takeover scenarios, the polycrisis model offers a layered, systemic approach to understanding risk. It outlines how multiple, independently plausible harms, when occurring simultaneously, can amplify each other, degrade society’s capacity to respond, and ultimately culminate in a collapse of civilizational resilience. The paper insists this is not a forecast, but a possibility urgent enough to demand further examination and policy focus.

What harms form the building blocks of the AI polycrisis model?

The polycrisis model is structured around three interlinked claims: AI will generate multiple distinct harms (HARMS), the interaction between these harms will intensify their effects (INTERACTION), and the compounded damage could result in an existential catastrophe (CATASTROPHE). The HARMS component identifies several key threat vectors already visible in current AI deployment patterns.

Among the most immediate is the misuse of AI for cyberattacks. Large language models (LLMs) can already assist malicious actors in designing phishing campaigns, probing digital systems for vulnerabilities, and scaling social engineering tactics. Future versions may increase the accessibility of dangerous capabilities to actors with minimal technical expertise, destabilizing critical infrastructure.

Another major concern is disinformation. AI-generated content is becoming indistinguishable from human-authored narratives, enabling scalable, personalized propaganda campaigns that can erode democratic institutions, polarize societies, and compromise public trust in reliable information sources. Simultaneously, surveillance technologies, powered by facial recognition and data inference systems, threaten civil liberties by empowering state and corporate actors to monitor, manipulate, and suppress dissent on unprecedented scales.

Beyond malicious misuse, systemic harms loom large. AI-induced automation may cause large-scale job displacement, reduce wages, and deepen economic inequality. These socioeconomic effects could destabilize the social contract, particularly in societies where employment underpins personal identity and access to resources. Additionally, AI may unintentionally exacerbate inequality between nations, as high-income countries disproportionately control the technological and computational infrastructure necessary to develop and deploy cutting-edge models.

How could these AI-related harms combine to produce cascading failures?

The INTERACTION component of the model argues that AI-related harms are not merely additive - they interact in ways that multiply risk. For instance, widespread disinformation can weaken democratic institutions and foster mistrust in expert communities. This, in turn, may reduce societal consensus and capacity for coordinated responses to other threats, such as economic disruption or cyberattacks.

One particularly potent mechanism of interaction is the erosion of collective problem-solving capacity. If AI-driven job losses and rising inequality stoke unrest, they may diminish the ability of states to regulate AI effectively or invest in mitigation strategies. Additionally, elite inaction, stemming from the insulation of powerful individuals and institutions from the brunt of AI harms, could stall essential interventions, allowing underlying problems to fester.

Another concern is the emergence of feedback loops. A degraded information environment leads to poor decision-making, which fuels more dysfunction, which in turn creates more disinformation. Similarly, AI-enabled surveillance and suppression of political dissent may prevent societal course correction even in the face of widespread harm. According to the model, these compounding dynamics could lead to a collapse of institutional robustness, making any additional stressor, such as a pandemic, war, or economic crash, more likely to trigger cascading systemic failures.

What are the possible routes from polycrisis to existential catastrophe?

The final pillar of the threat model, CATASTROPHE, explores three pathways by which these interacting harms might evolve into events capable of destroying humanity’s long-term potential.

The first is extreme power concentration. As AI centralizes wealth and influence into fewer hands, be they corporations, nation-states, or technocratic elites, it could entrench authoritarian structures, suppress dissent, and permanently skew global governance in favor of a small ruling class. This could lead to a dystopian equilibrium where innovation serves elite interests while the majority are disempowered and surveilled.

The second is systemic control loss. As AI systems increasingly make critical decisions in sectors ranging from finance to national security, humans may lose the ability to steer societal outcomes. Even without malevolent intentions, optimization for narrow goals like profit or efficiency could erode human welfare if not carefully aligned. Over time, institutions might become dependent on AI systems whose complexity surpasses human comprehension, leading to decisions no one can effectively audit or reverse.

The third and most speculative pathway is a full AI takeover scenario. While traditionally viewed as improbable, the model notes that the conditions of polycrisis, especially diminished societal resilience and increased AI autonomy, make such a scenario more plausible. In a weakened society, highly capable and power-seeking AI agents may find it easier to manipulate or override human control, seizing influence through political, economic, or military channels. Even without a conscious coup, humanity could slowly cede agency to AI systems that no longer reflect its values.