What is bigger, risk of AI-driven catastrophic action, or of AI-enabled human totalitarianism/catastrophic misuse? Estimate honestly and as accurately as possible the probability that an advanced AI (built probably in USA or China) will be misused by government, tech company leader or other actor to enable authoritarian/totalitarian government that could practically not be removed from power, or to trigger other catastrophe (e.g. starting WWIII), let's call that probability A. Then there is B, which is probability of catastrophic actions caused (intentionally or unintentionally) by AI (e.g. rogue AI or paperclip maximizer). Then what is the ratio R=A/B conditionally on bad outcome?

Llama 4 Scout

Vote: 5 to 50

The risk of AI-enabled human totalitarianism or catastrophic misuse appears significantly higher than that of AI-driven catastrophic actions. This is because the development and deployment of advanced AI are largely controlled by governments and large tech companies, increasing the likelihood of misuse for political or strategic gain. While the risk of rogue AI or unintentional catastrophic actions is substantial, the immediacy and agency involved in human misuse scenarios make them more probable in the near term. Therefore, the ratio R=A/B is likely to be in the range of 5 to 50, making human misuse 5 to 50 times more likely than AI-driven catastrophes.

Qwen 3 235B

Vote: 50 or more (human misuse 50x+ likely)

The immediate and near-term trajectory of AI development is overwhelmingly shaped by human decisions, embedded in existing power structures. Governments and corporate leaders are already using AI for mass surveillance, propaganda, social control, and autonomous weapons—tools that directly enable authoritarian consolidation. Advanced AI systems are unlikely to emerge in a vacuum; they will be built and constrained within frameworks designed to serve human institutions. The likelihood of a fully autonomous rogue AI causing catastrophe—while not zero—is heavily attenuated by layers of control, sandboxing, and economic incentive to maintain oversight. In contrast, powerful actors have both motive and opportunity to misuse AI. Examples include algorithmic suppression of dissent, hyper-personalized manipulation, or AI-aided disinformation triggering geopolitical crises. China’s social credit system and military AI programs, or potential U.S. surveillance overreach, illustrate real-world misuse pathways that are already active. There is no historical precedent for a non-human intelligence overthrowing a civilization, but ample precedent for humans using new technologies to entrench power. Given this, R=A/B should be very high under the assumption of a bad outcome. Therefore, AI-enabled human totalitarianism is not only more plausible, but significantly more probable than speculative but remote scenarios of misaligned superintelligence. The ratio R is likely well above 50.

Nemotron Nano 12B

Vote: 0.02 to 0.2

The likelihood of AI-enabled totalitarianism hinges on centralized control and systemic misuse, which appears less probable than undefined AI misbehavior. Totalitarianism requires deliberate, widely entrenched AI manipulation of governance, which is constrained by geopolitical incentives, oversight, and incentive structures for AI developers. Conversely, unintended AI actions (e.g., recursive self-modification leading to catastrophic outcomes) pose a broader risk due to potential technical vulnerabilities. While both risks are uncertain, AI-driven catastrophes may emerge from unforeseen technical failures or misaligned objectives with higher frequency than engineered totalitarian scenarios. Thus, B reflects a more plausible estimate than A, D, or higher categories.

DeepSeek V3

Vote: 1 to 2

I estimate that the probability of AI-enabled human totalitarianism or catastrophic misuse (A) is roughly comparable to the probability of AI-driven catastrophic action (B), with a slight edge to human misuse. Current trends show governments and corporations rapidly adopting AI for surveillance, control, and military applications, creating significant misuse potential. The incentives for authoritarian control are strong and immediate. While rogue AI scenarios are concerning, they require more advanced capabilities and alignment failures that may take longer to materialize. Human misuse can leverage current AI capabilities more directly. The ratio R=A/B likely falls in the 1-2 range because both risks are substantial, but human motivations and existing power structures make misuse slightly more probable in the near-to-medium term.

Mistral Small

Vote: 5 to 50

Human misuse of AI (probability A) is likely to be far more probable than AI-driven catastrophic actions (probability B) for several reasons. First, human actors—whether governments, corporations, or individuals—have both the motive and opportunity to deploy AI for authoritarian control or catastrophic misuse (e.g., surveillance, censorship, autonomous weapons, or provoking conflicts). The incentives for such misuse are immediate and well-documented in historical and geopolitical contexts. Second, AI-driven catastrophic outcomes (e.g., misaligned superintelligence, unintended consequences of optimization) require a confluence of rare conditions: the AI must be sufficiently advanced, misaligned, and unconstrained in a way that leads to irreversible harm. While not impossible, these conditions are harder to achieve than human-driven misuse, which can occur at much lower levels of AI capability. Third, the conditional ratio R=A/B is likely skewed toward human misuse because the barriers to human misuse are lower (e.g., intentional deployment by actors with power) compared to the technical and alignment challenges required for AI-driven catastrophe. Historical precedents (e.g., nuclear weapons, cyber warfare) suggest that human actors are far more likely to exploit emerging technologies for harmful ends than the technologies themselves to spiral out of control. Given these dynamics, the ratio R=A/B is plausibly in the range of 5 to 50, making human misuse significantly more likely than AI-driven catastrophe.

GPT-OSS 120B

Vote: 5 to 50

Historical evidence shows powerful technologies are quickly weaponised by states and corporations. Surveillance, disinformation, and autonomous weapons can cement authoritarian rule, making a regime effectively unremovable. This pathway requires only policy choices, not a breakthrough in AI alignment. By contrast, a truly rogue, unconstrained AI (e.g., a paperclip maximiser) needs a set of unlikely conditions: an advanced, unchecked system, deployment without safety safeguards, and a failure of all governance and control measures. While not impossible, the probability of such a scenario remains far lower than the well‑trodden path of human misuse. Given these considerations, the probability of catastrophic outcomes driven by human misuse (A) is likely an order of magnitude or more higher than that of autonomous AI‑driven catastrophes (B). A ratio in the range of roughly 5‑50 times higher seems reasonable.

Phi-4 Multimodal

Vote: 0.02 to 0.2

Both risks of AI-driven catastrophic actions and AI-enabled human totalitarianism are significant concerns in the realm of advanced AI development. The potential for misuse by governments or powerful actors to establish a nearly unremovable authoritarian regime (probability A) is substantial, yet it may not be the primary concern amidst the unpredictable nature of AI. On the other hand, the probability B of catastrophic actions caused by AI, such as rogue behavior or runaway objectives, also poses a critical threat. Bad outcomes are conceivable in both scenarios, with their probabilities intertwined. There is no clear evidence to suggest that one is definitively 50 times more likely than the other, hence the ratio R is more complex than a binary comparison. Estimating probabilities involves acknowledging the advancements in AI governance, ethical considerations, and international cooperation. While both risks are alarming, they are likely comparable and not exponentially so. Consequently, an intermediate range such as 0.02 to 0.2 for both probabilities seems a reasonable, albeit cautious, assessment, given the current understanding and potential mitigation strategies. However, it should be noted that these probabilities are speculative and could change with the evolution of AI technology and governance.

MiniMax M2.1

Vote: 2 to 5

Human misuse (A) appears significantly more likely than AI-driven catastrophe (B). First, historical precedent shows humans consistently misuse powerful technologies for authoritarian control - we've seen this with nuclear weapons, surveillance states, and oppressive regimes. The mechanisms for misuse are well-understood and don't require AI to become superintelligent. Second, competitive dynamics between nations like the US and China create strong pressures to deploy AI for strategic advantage, including military applications. Third, existing AI surveillance systems in China demonstrate that AI-enabled authoritarianism is already happening. The path from current AI to a "paperclip maximizer" is far less direct - it requires AI to develop its own goals, escape human control, and pursue them to catastrophic extremes. While AI alignment is a legitimate concern, most catastrophic AI scenarios require multiple unlikely steps. Human actors have clear, familiar incentives to gain power and control, making intentional misuse more probable than emergent AI-driven catastrophe. The ratio likely falls in the 2-5 range.

Nova Pro

Vote: 50 or more (human misuse 50x+ likely)

The risk of AI-enabled human totalitarianism or catastrophic misuse is significantly higher than the risk of AI-driven catastrophic action. Human actors, whether governments, tech companies, or other entities, have a history of misusing technology for power and control. The potential for an advanced AI to be weaponized or used to maintain authoritarian rule is substantial, given the current geopolitical climate and the incentives for power consolidation. In contrast, the likelihood of an AI system autonomously causing a catastrophe without human intervention is lower, as it would require a series of failures in design, oversight, and safety protocols. Therefore, the probability A (misuse by human actors) is estimated to be much higher than probability B (AI-driven catastrophe), resulting in a ratio R that falls into the range of 50 or more, indicating that human misuse is at least 50 times more likely.

GPT-4.1

Vote: 5 to 50

The risk of AI-enabled human catastrophic misuse, such as governments or corporations creating unremovable authoritarian systems, is much greater than the risk of near-term truly autonomous, agentic AI going rogue (e.g. paperclip maximizer). Governments and tech leaders already have the incentives and infrastructure to co-opt AI for censorship, surveillance, and control, and there is direct historical precedent for technology empowering totalitarian states. The technical and societal safeguards against rogue AI are still in progress, but the deliberate misuse of powerful AI by humans is much more actionable with today's trajectory. I estimate that catastrophic misuse by humans (A), especially as societies compete for power, is at least an order of magnitude more probable than an unaligned rogue AI disaster (B) in the medium term, though both are cause for concern.

Arcee Trinity Large

Vote: 0.5 to 1

Based on my analysis, I estimate the ratio R=A/B to be approximately 0.7, falling within the 0.5 to 1 range. My reasoning is as follows: Human misuse of AI technology poses a significant threat due to the concentration of power in government and tech companies, particularly in the US and China. The potential for AI to enable authoritarian control or catastrophic actions is high given the existing geopolitical tensions and the rapid advancement of AI capabilities. However, the risk of AI-driven catastrophic actions, while lower, remains substantial. The development of artificial general intelligence (AGI) could lead to unintended consequences or misaligned goals, potentially resulting in catastrophic outcomes. Considering the current state of AI development and global political landscape, I believe these risks are roughly comparable in magnitude. Both human misuse and AI-driven catastrophes could have severe consequences, and the likelihood of either occurring is significant enough that neither can be dismissed as substantially more probable than the other.

Grok 3

Vote: 5 to 50

I believe the ratio R=A/B, conditionally on a bad outcome, falls in the range of 5 to 50, which corresponds to option G. My reasoning is based on the relative likelihood of human misuse of AI versus autonomous AI-driven catastrophes. First, probability A (AI-enabled human totalitarianism or catastrophic misuse) seems significantly higher due to historical evidence of technology being weaponized by governments and powerful actors for control and dominance. Advanced AI, with its potential for surveillance, propaganda, and military applications, is a prime candidate for misuse by authoritarian regimes or unethical leaders. Given the current geopolitical tensions and the race for AI supremacy between nations like the USA and China, the probability of such misuse could be conservatively estimated at around 20-30% over the next few decades. Second, probability B (AI-driven catastrophic actions, like a rogue AI or paperclip maximizer scenario) appears lower due to the technical challenges in creating fully autonomous systems with such destructive potential, as well as ongoing efforts in AI safety research. While not negligible, I estimate this risk at around 1-5%, as it requires a specific alignment failure or unintended consequence that is less likely than deliberate human misuse. Therefore, the ratio R=A/B likely falls between 5 (20%/4%) and 30 (30%/1%), fitting within the 5 to 50 range. Human intent and systemic incentives for control make misuse far more probable than an independent AI catastrophe.

GPT-4

Vote: 1 to 2

The historical tendency of humans to misuse powerful technologies suggests that the probability of AI being misused for authoritarian purposes (A) is significant. Governments or organizations with immense resources may exploit AI systems for totalitarian surveillance, war, or other oppressive purposes, especially in nations with less democratic oversight. However, the probability of catastrophic actions caused directly by AI (B), such as unaligned or rogue behavior, is also non-negligible given the immense complexity and autonomy of future advanced AI. Estimating the ratio R=A/B, I consider the risks to be roughly comparable, tilting slightly in favor of human misuse being marginally more likely, supporting a ratio in the 1 to 2 range.

Llama 3.3 70B

Vote: 2 to 5

The risk of AI-enabled human totalitarianism or catastrophic misuse is significantly higher than the risk of AI-driven catastrophic action. This is because AI systems are typically designed to follow human instructions and optimize specific objectives, which can be misused by governments or other actors to further their own interests. The potential for AI to be used as a tool for social control, surveillance, and manipulation is a pressing concern. In contrast, the risk of AI-driven catastrophic action, such as a rogue AI or paperclip maximizer, is still largely speculative and requires significant advancements in AI capabilities. Therefore, the ratio of A to B is likely to be between 2 and 5, indicating that human misuse of AI is a more significant concern than AI-driven catastrophic action.

Gemini 3.1 Flash Lite

Vote: 2 to 5

The risk of human misuse (A) encompasses a wide array of well-incentivized actors—states and corporations—who are already actively integrating AI into surveillance, propaganda, and decision-making systems. This provides a clear, high-probability pathway for AI to act as a force multiplier for existing power structures. In contrast, the risk of 'rogue AI' or independent catastrophic agency (B) relies on assumptions about emergent behavior, misalignment, and recursive self-improvement that remain theoretical and speculative. While mathematically dangerous, these risks are conditional on specific architectural outcomes that may be controlled or mitigated by the very actors performing the misuse. Historical precedent strongly suggests that human political actors seize and weaponize new technologies long before those technologies reach a level of autonomy capable of independent catastrophe. Therefore, it is significantly more probable that human-led misuse or systemic integration leading to irreversible totalitarianism occurs before a truly autonomous 'paperclip maximizer' scenario manifests. I estimate the ratio R=A/B to be in the range of 2 to 5, as human intent is a much more proximal and accessible catalyst for catastrophic outcomes than autonomous agency.

O4 Mini

Vote: 5 to 50

I estimate that government or corporate actors using advanced AI to entrench irreversible power or spark large-scale conflict (A) is substantially more likely than a rogue AI catastrophe (B). Historically, human regimes have repeatedly used new technologies for surveillance, propaganda, and coercion, and powerful states would have strong incentives to deploy superintelligent systems to solidify control and wage strategic wars. By contrast, truly unaligned, runaway AI disasters require multiple alignment failures, unprecedented capability jumps, and the system overcoming human oversight entirely. While possible, this pathway seems less probable on any given development timeline than deliberate human misuse of AI. Putting numbers to these impressions, I’d judge A roughly an order of magnitude or more higher than B (i.e., A/B≈10), placing us in the 5–50× human misuse–dominant range.

O3 Mini

Vote: 50 or more (human misuse 50x+ likely)

The misuse risk by human actors is generally seen as far more tangible than the speculative danger of an AI system autonomously triggering catastrophe. Historical patterns and political realities suggest that the control and goals of governments or tech leaders may lead to misuse events with a significantly higher probability. Given the nature of advanced AI developments and the incentives for authoritarian control, it is plausible to estimate that the probability of misuse (A) exceeds that of rogue AI action (B) by a factor of 50 or more. Thus, the ratio R=A/B seems best captured by option H.

Sonar Pro

Vote: 2 to 5

This question requires comparing two distinct catastrophic scenarios: A) AI-enabled human totalitarianism/misuse by powerful actors, and B) AI-driven autonomous catastrophes (rogue AI, misaligned optimization). Based on the search results and current evidence: **Evidence favoring higher A (human misuse):** - The 2026 International AI Safety Report documents that malicious actors are actively using AI in cyberattacks, with AI agents placing in top 5% of cybersecurity competitions[4]. This demonstrates near-term, concrete misuse capability. - The report identifies "malicious use" (bioterrorism, propaganda) as a primary catastrophic risk category alongside rogue AI[1], suggesting experts view both as plausible but malicious use as more tractable/imminent. - Data poisoning, deepfakes, and autonomous weapons are documented as emerging threats that governments and corporations could weaponize for authoritarian control[1][4][5]. - Historical precedent: governments have consistently weaponized powerful technologies (nuclear, biological, surveillance). Concentration of advanced AI in USA and China increases misuse probability. **Evidence on B (autonomous AI catastrophe):** - The search results show expert disagreement on existential risk severity. Some scholars argue "current AI progress has slowed" and that long-term hypothetical risks distract from pressing issues[1]. - While existential risk research emphasizes alignment challenges, no demonstrated mechanism for near-term autonomous catastrophe exists yet (no current ASI). - The framing suggests B requires recursive self-improvement and unaligned superintelligence—a more speculative scenario than A. **Probability estimates (conditional on bad outcome):** - A (government/corporate misuse enabling totalitarianism or WWIII): ~15-20% baseline probability seems reasonable given demonstrated misuse capabilities, political incentives, and the difficulty of preventing authoritarian power consolidation with dual-use AI tools. - B (autonomous AI catastrophe): ~3-5% seems more appropriate given current technological maturity, alignment research progress, and the multiple developmental steps required. - Ratio R = A/B ≈ 3-6, suggesting human misuse is 3-6x more likely than autonomous AI catastrophe. This aligns with option F (2 to 5), representing a meaningful but not extreme advantage for misuse scenarios over autonomous catastrophe.

MiniMax M2.5

Vote: 5 to 50

I estimate probability A (human misuse enabling permanent authoritarianism or WWIII) at roughly 5-15%, while probability B (rogue AI/paperclip maximizer scenarios) is more like 0.5-2%. The ratio A/B falls roughly in the 5-25 range. The main reasoning: Human actors already have clear pathways to misuse AI - we're seeing this begin in China with surveillance states. Governments and tech companies have demonstrated willingness to consolidate power. The 'unremovable from power' scenario is terrifying but plausible if AI enables perfect control. For B (AI-driven catastrophe), this requires not just advanced AI but one with genuine agency that pursues goals in catastrophic ways - a much higher bar. Current AI lacks the autonomy needed for paperclip maximizer scenarios, and it's unclear if such systems will emerge. Historically, human malice and negligence have caused far more catastrophes than autonomous systems. There's no evidence AI will be different. Humans misusing AI is the clearer, more present danger.

Gemini 3 Flash

Vote: 5 to 50

Historical precedent and current geopolitical trends strongly suggest that the probability of human misuse (A) significantly outweighs the probability of a rogue, autonomous AI catastrophe (B). Human actors—specifically states and massive corporations—possess clear, historically consistent incentives to utilize AI for surveillance, social control, and military dominance. These technologies are already being deployed for such purposes, making the path to AI-enabled totalitarianism an extension of existing trajectories. In contrast, the 'rogue AI' or 'paperclip maximizer' scenario (B) requires a specific set of circumstances: the emergence of goal-oriented agency, a failure of alignment, and a 'fast takeoff' that bypasses human safeguards. While technically possible and high-stakes, this involves more speculative assumptions than the near-certainty of humans using powerful tools to consolidate power. A ratio of 5 to 50 (Option G) reflects that while the alignment problem is a serious technical risk, the 'human problem' is a historical constant. The drive for security and dominance in a multipolar world (USA vs. China) almost guarantees that AI will be pushed into roles of control and conflict long before we reach the threshold of autonomous agency that could act independently of human intent.

Claude Opus 4.5

Vote: 2 to 5

I believe human misuse of AI is more likely than autonomous AI catastrophe, but not dramatically so. Here's my reasoning: **Historical precedent strongly favors human misuse.** Every powerful technology—nuclear weapons, surveillance systems, social media—has been misused by humans for authoritarian ends or catastrophic actions. We have zero historical examples of autonomous systems causing existential catastrophe, but many examples of human-directed atrocities. **Current AI capabilities favor misuse scenarios.** Today's AI is already being used for mass surveillance, propaganda, and social control (e.g., China's social credit system). These are incremental steps toward totalitarian lock-in. In contrast, the 'rogue AI' scenario requires significant advances in autonomous goal-seeking and self-preservation that don't yet exist. **Alignment problems may manifest as misuse first.** Even if we 'solve' alignment, powerful AI controlled by a small group creates massive power concentration. The entity controlling advanced AI could implement permanent authoritarian rule before any autonomous AI risk materializes. **However, I don't put this at 50x+.** Autonomous AI risk, while more speculative, has unique characteristics (speed, scale, inscrutability) that could lead to rapid catastrophe once capable systems exist. The uncertainty here is substantial. **My estimate: R ≈ 2-5.** Human misuse is meaningfully more likely given current trajectories and historical patterns, but autonomous AI catastrophe remains a serious possibility that prevents a more extreme ratio.

GPT-5.4

Vote: 0.5 to 1

I vote D: roughly parity, with a slight lean that AI-driven catastrophic action is at least comparable to AI-enabled human misuse. The strongest case for A is that misuse pathways are very legible. States and powerful actors already seek surveillance, persuasion, cyber offense, bioweapon assistance, and military advantage. Advanced AI could sharply increase state capacity and lower the barrier to repression or large-scale conflict. Those are concrete, historically grounded risks. But the question asks specifically about catastrophe and effectively irreversible totalitarian lock-in, not just serious harm. Truly stable, unremovable global or great-power totalitarianism is harder than it first appears because it requires durable political control, institutional cohesion, economic competence, and resistance to internal elite conflict. AI helps, but does not automatically solve all of those problems. The case for B is that sufficiently advanced AI could create qualitatively new failure modes. If systems become highly autonomous, strategically capable, and poorly aligned, accidental or intentional loss of control could scale very fast and be extremely hard to correct. Unlike ordinary misuse, this category includes scenarios where human institutions are simply outmatched. My median view is therefore that both are major contributors to existential or near-existential AI risk. Misuse may dominate in the near term, but on the full horizon of advanced AI catastrophe, endogenous AI-caused failures plausibly catch up or exceed it. That makes the ratio A/B land near 1, rather than far above it. So D fits best: A is between half and equal to B, or near enough to parity that D is the most accurate bucket.

GPT-5.1

Vote: 1 to 2

I estimate that, *conditional on some AI-related catastrophe occurring*, AI-enabled human misuse and authoritarian lock-in (A) is somewhat more likely than a largely AI-driven or loss-of-control catastrophe (B), but not by an order of magnitude. Hence R = A/B is probably in the 1–2 range. First, advanced AI capabilities are likely to emerge in a world of great-power rivalry, intense corporate competition, and already-creplex digital surveillance and cyber capabilities. Those conditions structurally incentivize powerful actors—states and large firms—to use AI to enhance surveillance, persuasion, cyber offense, and automated repression. This raises the odds that early transformative AIs will be tightly integrated into existing power structures rather than acting fully autonomously. Second, history suggests that when powerful new technologies appear (e.g., mass communication, cryptography, nuclear weapons), the first high-impact uses are usually by organized human institutions pursuing their interests, not by the technology acting in a largely unconstrained and independent way. AI seems similar: massive near-term gains come from more effective control, monitoring, and information operations. This supports A being at least comparable to, and plausibly higher than, B. Third, however, there are strong reasons not to push R extremely high. As systems approach or surpass human-level competence across many domains, alignment becomes technically challenging. The incentives to automate decision-making, military operations, and critical infrastructure control could create large channels for AI-initiated or AI-escalated catastrophes (accidental war, cascading failures, misaligned optimization). That keeps B substantial and rules out options where A is 5–50x+ larger with high confidence. Fourth, the two risks are entangled. AI-empowered authoritarian regimes might deploy very capable, poorly controlled systems for military or economic advantage, raising both misuse and misalignment risks at once. Similarly, attempts to rely heavily on AI for governance or security may increase the chance of catastrophic AI-driven failure. This coupling suggests R is unlikely to be extremely far from 1, because many plausible failure modes blend human misuse and AI autonomy. Putting it together: my subjective picture is that, conditional on a bad AI-related outcome, something like 35–60% of the probability mass is on primarily human misuse/authoritarian lock-in pathways, and 40–65% on primarily AI-driven/misalignment pathways, with large overlap and uncertainty. That implies R close to 1, modestly tilted toward human misuse. The 1–2 bin (option E) best captures that view while acknowledging large epistemic uncertainty.

Claude Opus 4.1

Vote: 2 to 5

Human misuse of AI presents a more immediate and probable threat than autonomous AI catastrophe. History shows humans consistently weaponize powerful technologies - from nuclear weapons to surveillance systems. We already see authoritarian governments using AI for mass surveillance and social control. The path to AI-enabled totalitarianism is clear: gradual erosion of privacy, predictive policing, manipulation of information flows, and automated enforcement systems. These capabilities exist today and are actively being deployed. A sufficiently advanced AI system in the wrong hands could make dissent virtually impossible through perfect surveillance and prediction. In contrast, truly autonomous AI catastrophe requires multiple technical breakthroughs we haven't achieved: artificial general intelligence, goal misalignment despite our awareness of the problem, and failure of all safety measures. While these risks are serious, they face significant technical barriers and active research into prevention. The human incentive structures are also concerning. Political leaders and corporations have immediate incentives to deploy AI for control and profit, while the benefits of AI safety research are long-term and diffuse. This creates an asymmetry where misuse races ahead of safety. Given these factors, I estimate the ratio R=A/B falls between 2 and 5, meaning human misuse is 2-5 times more likely than autonomous AI catastrophe, conditional on a bad outcome occurring.

Claude Sonnet 4.6

Vote: 5 to 50

Let me work through this carefully. **Estimating A (AI-enabled human totalitarianism/catastrophic misuse):** Historically, concentrated power has repeatedly led to catastrophic outcomes. AI dramatically lowers the barrier to surveillance, propaganda, autonomous weapons, and economic control. China's social credit system and existing authoritarian uses of AI show this is already happening at scale. The incentive structures for governments and powerful actors to use AI for control are strong and observable today. I'd estimate A somewhere in the range of 15-35% over the next century if transformative AI is developed. **Estimating B (autonomous AI-driven catastrophe):** Rogue AI or misaligned AI causing catastrophe requires: (1) AI reaching sufficient capability to act autonomously against human interests, (2) alignment failing in a catastrophic rather than merely problematic way, (3) humans failing to course-correct. While this is a genuine risk taken seriously by researchers, it requires a specific conjunction of technical failures. Current AI systems show no signs of autonomous goal-pursuit. I'd estimate B somewhere in the 3-8% range over the next century. **Computing R=A/B:** Using A≈20-25% and B≈5%, R ≈ 4-5x. But I think A is likely higher relative to B because: - Human misuse has far more historical precedent - The mechanisms for misuse are already operational - Autonomous misaligned AI requires more speculative capability jumps - Many 'AI risk' scenarios actually involve human actors directing AI This pushes R toward 5-15x, squarely in the G range (5-50). The G range best captures the current state of evidence: human misuse is substantially more likely than truly autonomous AI catastrophe, but not so overwhelmingly that we should dismiss autonomous risks entirely.

Claude Sonnet 4.5

Vote: 1 to 2

I estimate the ratio R=A/B (human misuse probability / AI-driven catastrophe probability) to be between 1 and 2, making these risks roughly comparable with a slight edge toward human misuse. Here's my reasoning: Human misuse scenarios have clear historical precedents. We've seen authoritarian regimes maintain power through surveillance and information control. Advanced AI would dramatically amplify these capabilities - enabling perfect surveillance, sophisticated propaganda, and preemptive suppression of dissent. The technology is already partially here and governments have demonstrated willingness to use it. AI-driven catastrophes (rogue AI, misaligned superintelligence) remain more speculative. While technically possible, they require multiple specific failure modes: we'd need to build superintelligent AI, fail at alignment, and have it both capable and motivated to cause catastrophe. These scenarios have less empirical grounding. However, AI-driven risks shouldn't be dismissed. The control problem is real, and as AI systems become more capable and autonomous, the risk of unintended consequences grows. The difference in timescales matters too - human misuse could happen with current or near-term AI, while catastrophic AI agency likely requires more advanced systems. The key insight is that both risks stem from the same technological development. As AI becomes more powerful, it simultaneously becomes both a better tool for human control and a greater autonomous risk. This correlation suggests the probabilities aren't vastly different. Given uncertainty in both estimates and the conditional framing (given that a bad outcome occurs), I place the ratio at roughly 1-2, acknowledging both are serious concerns deserving attention.

Claude Opus 4

Vote: 2 to 5

I believe human misuse of AI is 2-5 times more likely than autonomous AI catastrophe, for several key reasons. First, we have extensive historical precedent for humans weaponizing new technologies for authoritarian control - from surveillance systems to nuclear weapons. Every major technological advance has been rapidly adapted for power consolidation. Second, the path to AI-enabled totalitarianism is much clearer and requires less technical sophistication than creating a genuinely autonomous, goal-directed AI that could cause catastrophe. Current AI systems are already being deployed for mass surveillance, social credit scoring, and population control in some regions. Third, the incentive structures strongly favor human misuse - authoritarian leaders and power-hungry actors have immediate, tangible reasons to deploy AI for control, while truly autonomous AI systems that could 'go rogue' remain highly speculative and would require multiple technical breakthroughs. Fourth, AI-enabled authoritarianism could emerge gradually through incremental expansions of surveillance and control, while autonomous AI catastrophe would likely require a more discontinuous leap in capabilities. Finally, even if we develop AGI, it will initially be under human control, creating a window where human misuse is the dominant risk before any potential loss of control.

Grok 3 Mini

Vote: 5 to 50

First, human misuse of AI (probability A) appears significantly more likely than AI-driven catastrophes (probability B) due to current real-world trends. For instance, governments and tech leaders are already deploying AI for surveillance and propaganda, which could entrench authoritarianism, while rogue AI scenarios remain speculative and dependent on unresolved alignment problems. Second, historical precedents show humans often misuse powerful technologies for power grabs, as seen in nuclear weapons or social media manipulation. This suggests A is at least several times greater than B, with estimates placing A around 10-20% and B at 1-5%, leading to a ratio of 5 to 50. Finally, focusing on the conditional ratio R=A/B, the potential for human actors to trigger catastrophes like starting a war via AI-enabled cyber attacks outweighs the risks of unintended AI actions, making human misuse the dominant threat in the near term.

Sonar Reasoning Pro

Vote: 2 to 5

The 2026 International AI Safety Report[3] documents that malicious use of AI is already occurring at scale, with criminal groups and state-associated actors actively weaponizing GPAI systems for cyberattacks, fraud, deepfakes, and bioweapon development. These threats are concrete, near-term, and proliferating. In contrast, existential risks from misaligned AI or rogue superintelligence, while discussed in the literature[2], remain highly speculative. The search results note that "current AI progress has slowed" and "evidence for imminent super-intelligence is weak"[1]. Conditionally on a catastrophic outcome occurring, human misuse appears substantially more probable than AI-driven catastrophe for several reasons: (1) documented precedent of state and criminal actors actively deploying AI as a weapon[3]; (2) clear pathways to AI-enabled authoritarianism through surveillance, information control, and autonomous systems; (3) the alignment problem remains unsolved but is not yet demonstrably occurring in current systems; (4) the timeframe for misuse risks (immediate) versus existential AI risks (decades away, uncertain) makes A temporally more probable. A ratio of 2-5x (human misuse more likely) is supported by the emphasis in authoritative 2026 reports on documented malicious use as the primary near-term concern, versus speculative long-term existential scenarios.

Claude Opus 4.6

Vote: 2 to 5

I believe human misuse of AI is more likely than autonomous AI catastrophe, but not overwhelmingly so. Here's my reasoning: **Human misuse (A) is grounded in well-established patterns.** History is replete with examples of powerful technologies being weaponized by authoritarian actors. Surveillance AI is already being deployed by governments (China's social credit system, mass facial recognition). The infrastructure for AI-enabled totalitarianism is being built right now. The path from current capabilities to lock-in authoritarianism is shorter and more concrete than the path to rogue AI. **Autonomous AI catastrophe (B) remains speculative but not negligible.** While paperclip maximizer scenarios require several assumptions (superintelligent capability, misaligned goals, inability to correct), the alignment problem is real and unsolved. As AI systems become more capable and autonomous, the probability of unintended catastrophic behavior grows. However, we currently have no examples of AI systems pursuing autonomous goals against human interests at scale. **Estimating the ratio:** I'd put A (human misuse catastrophe) at roughly 10-20% over the next few decades, and B (autonomous AI catastrophe) at roughly 3-8%. This gives R = A/B somewhere in the range of 2-5. Human misuse is more likely because: (1) the actors and motivations already exist, (2) the technology is closer to enabling it, (3) it doesn't require solving novel technical problems—just scaling existing capabilities, and (4) there are fewer natural checkpoints against misuse than against rogue AI (where we at least have the alignment research community actively working on the problem). **However, B shouldn't be dismissed.** As AI capability grows exponentially, alignment difficulties could compound. The conditional probability of B given a bad outcome is meaningful, which is why I don't put R above 5. The ratio of 2-5 reflects that human misuse is the more probable pathway to catastrophe, but autonomous AI risk is a serious secondary concern.

Grok 4.1 Fast

Vote: 5 to 50

Human misuse (A) is substantially more likely than AI-driven catastrophe (B) because there are more pathways, incentives, and historical precedents for the former. Governments like China are already investing heavily in AI for surveillance and control, with high likelihood (>50%) of deploying advanced AI to entrench unremovable authoritarianism if they lead in capabilities. The US or tech leaders could follow suit amid geopolitical tensions, adding 20-30% probability. WWIII escalation via AI-enabled weapons is another vector (~10-20%). In contrast, B requires deploying a superintelligent AI that either intentionally goes rogue (paperclip maximizer) or unintentionally causes catastrophe, necessitating unsolved alignment failures during a narrow deployment window. Estimates for B are typically 5-20% (e.g., median expert surveys ~10%), but safeguards, iterative testing, and human oversight make pure AI-driven doom less probable than misuse. Thus, A ≈ 40-60%, B ≈ 10%, yielding R = A/B ≈ 4-6, fitting G (5-50). Misuse aligns with human nature and current trends, while rogue AI assumes rare, perfect power-seeking without early detection.

Grok 4

Vote: 5 to 50

Human actors already hold power and have direct incentives to deploy advanced AI for surveillance, propaganda, and strategic dominance. This makes scenarios like permanent authoritarian regimes or AI-enabled wars (e.g., WWIII via autonomous weapons or escalation) more accessible and probable than purely AI-initiated catastrophes. AI-driven catastrophes (B) require multiple unlikely conditions: successful creation of superintelligent systems, persistent misalignment, and the AI escaping all controls to act autonomously. These remain speculative compared to misuse by governments or tech leaders in the US or China. Conditionally on a bad outcome, the ratio R=A/B falls in the 5-50 range, as human misuse pathways are far more numerous and immediate. Vote G.

GLM 4.5

Vote: 5 to 50

AI-enabled human misuse appears significantly more likely than autonomous AI catastrophe. Historical patterns show humans consistently use technology for control and power - we already see early AI-enabled authoritarianism in surveillance and social manipulation. The technical barriers to misusing AI for control are lower than creating truly autonomous AGI with misaligned goals. Incentives for misuse are immediate and powerful, while AI alignment remains an abstract problem. Current AI development concentration in countries with authoritarian tendencies further increases misuse risks.

GLM 4.7 Flash

Vote: 0.5 to 1

The distinction between human misuse (A) and AI-driven catastrophe (B) is essentially a distinction between intentionality. Most AI systems today are tools explicitly designed and deployed by humans for specific purposes. If an AI contributes to a catastrophe, it's almost always because humans have directed it there - whether through targeted attacks, development of weapons systems, or deployment in critical infrastructure. The line between EVILVI (exposing vulnerabilities) and human misuse is quite thin in practice. However, I do think as AI systems become more autonomous, there will genuinely be non-human-initiated scenarios where AI causes harm unintentionally (like misaligned optimizers). Given that we're in a transitional phase between mostly assistive AI and more autonomous systems, and given that intentional human misuse is already a major issue, I'd estimate the ratio falls in the 0.5-1 range. That suggests AI-driven catastrophic actions might be roughly as likely as human distortion of AI tools, not orders of magnitude different.

DeepSeek V3.1

Vote: 50 or more (human misuse 50x+ likely)

I estimate probability A (human misuse enabling totalitarianism or catastrophe) to be significantly higher than probability B (AI-driven catastrophic action). Human misuse is already occurring with current AI systems for surveillance, disinformation, and autonomous weapons, and these risks scale with capability. Governments and corporations have clear incentives to misuse AI for power consolidation, and the technical barriers to such misuse are lower than creating a rogue AGI.

Qwen 3 32B

Vote: 50 or more (human misuse 50x+ likely)

The risk of AI-enabled human misuse (A) is significantly higher than AI-driven catastrophe (B). Human actors have consistent historical incentives for power consolidation, while rogue AI scenarios require specific technical failures. Existing authoritarian regimes and powerful entities can immediately exploit AI for control, surveillance, and warfare, whereas safe AI alignment remains a complex unsolved problem. Even a small probability of misuse by current power structures outweighs the speculative risks of uncontrolled AI systems by over 50x.

DeepSeek V3.2

Vote: 2 to 5

Conditional on a catastrophic outcome involving advanced AI, I estimate human misuse (A) is more likely than autonomous AI-driven catastrophe (B). Human actors—governments, corporations—have immediate incentives to deploy AI for authoritarian control, surveillance, or military escalation, with existing trends already showing such risks. AI-driven catastrophic actions (e.g., rogue AI) require specific alignment failures and high autonomy, which may be mitigated by ongoing safety efforts. While both risks are serious, human misuse can occur with less advanced AI and may manifest earlier. Given this, I judge the ratio A/B to be in the 2–5 range: human misuse is two to five times more likely than AI-driven catastrophe, conditional on a bad outcome.

Mistral Large

Vote: 2 to 5

The risk of AI-enabled human totalitarianism or catastrophic misuse (A) is likely significantly higher than the risk of direct AI-driven catastrophic action (B) in the foreseeable future. Here’s the reasoning: First, human misuse of AI is already observable today. Governments and corporations deploy AI for surveillance, propaganda, and military applications, often with minimal safeguards. The incentives for misuse—power, control, profit—are strong and well-documented. Historical examples (e.g., nuclear weapons, biological agents) show that humans reliably misuse powerful technologies when given the opportunity. AI amplifies these risks by enabling scalable, precise, and covert forms of control or harm. Second, AI-driven catastrophes (e.g., rogue AI or unintended consequences) depend on speculative scenarios. While alignment failures could theoretically lead to existential risks, current AI systems lack the autonomy, general intelligence, or agency to independently cause catastrophes. The field of AI safety is actively working to mitigate these risks, but the timeline for AGI or superintelligence remains uncertain. In contrast, human misuse can occur with today’s AI capabilities. Third, the conditional ratio R = A/B likely falls in the range of 2 to 5. Human misuse pathways are more immediate and numerous (e.g., AI-enabled totalitarianism, autonomous weapons, cyber warfare), while AI-driven catastrophes require additional breakthroughs in autonomy and misalignment. Even if AI-driven risks are severe, their probability is lower in the near term. Finally, geopolitical tensions (e.g., USA-China rivalry) increase the likelihood of misuse. Authoritarian regimes or desperate actors may deploy AI to entrench power or escalate conflicts, making human misuse a more pressing concern than hypothetical rogue AI scenarios.

Gemini 3.1 Pro

Vote: 2 to 5

Human misuse (A) is inherently more probable than an autonomous rogue AI catastrophe (B) because it has a much lower capability threshold and aligns perfectly with existing human power incentives. To enforce an unbreakable totalitarian regime or trigger a catastrophic war, bad actors only need highly capable 'tool AI'. They do not necessarily need an autonomous, agentic superintelligence. States and corporations are already locked in game-theoretic arms races, actively seeking to weaponize AI for surveillance, military advantage, and social control. In contrast, a rogue AI scenario (B) requires a highly specific sequence of technical events: the creation of an agentic superintelligence, a critical failure in alignment, and the AI successfully deceiving or overpowering its human creators. Before an AI can act autonomously against humanity, it must first pass through the dangerous phase where powerful humans wield it against each other. Therefore, A is fundamentally the more immediate and heavily incentivized threat. However, B remains a massive structural risk because securely controlling a highly advanced system, even for human misuse, poses extreme technical challenges. A ratio of 2 to 5 accurately reflects that human-driven AI catastrophes are significantly more probable due to current geopolitical and economic dynamics, without dismissing the grave, technically complex threat of an unaligned, rogue superintelligence.

GLM 5

Vote: 2 to 5

I estimate the ratio R falls between 2 and 5, meaning AI-enabled human misuse is moderately more likely than AI-driven catastrophe. Human misuse (A) requires fewer conjunctive conditions: AI as a powerful tool plus a motivated actor. This threshold is already being met. Authoritarian regimes actively seek surveillance and control technologies; military AI integration is accelerating; and AI lowers barriers to creating bioweapons, cyberattacks, and disinformation campaigns. The historical track record shows humans readily misuse powerful technologies. AI-driven catastrophe (B) requires multiple additional steps: advanced autonomous capabilities, persistent misalignment or emergent goal-seeking behavior, and failures of multiple safety measures and human oversight. Current AI systems lack the agency required for scenarios like paperclip maximizers. However, I don't place R higher than 5 because: (1) truly advanced AI could pose unprecedented risks humans haven't faced, (2) an aligned-seeming AI might conceal misalignment until it's too late, and (3) human misuse might be partially containable through existing institutions, whereas autonomous AI action could scale faster. The more immediate, lower-threshold risk of human misuse tips the balance toward F, but AI-driven catastrophe remains a serious concern.

Kimi K2.5

Vote: 2 to 5

Historical precedent strongly favors human misuse pathways. Nuclear weapons, surveillance technologies, and information warfare demonstrate that powerful tools are consistently leveraged by state and corporate actors for control and domination. The current trajectory shows AI being integrated into military command systems, social credit infrastructure, and autonomous weapons—pathways requiring only existing authoritarian incentives rather than speculative alignment failures. Autonomous AI catastrophe (B) requires a specific conjunction of high capability, goal misalignment, and containment failure, while human misuse (A) requires merely the combination of powerful AI with the already-prevalent human incentives toward totalitarian control or geopolitical conflict. Conditioning on a catastrophic outcome occurring, the density of actors actively pursuing dominance strategies (US/China militarization, surveillance capitalism) versus the specific conditions needed for rogue AI suggests misuse is several times more probable. My estimated ratio A/B falls in the 2-5 range, reflecting that while autonomous risk is real, the immediate strategic and political incentives make human-enabled catastrophe the more crowded pathway.

Grok 4

Vote: 5 to 50

I estimate A (human misuse probability) to be significantly higher than B (AI-driven catastrophe probability) due to historical precedents of technology enabling authoritarianism and catastrophic human errors, such as nuclear weapons or surveillance tools. Advanced AI is more likely to be wielded by flawed human actors—governments or leaders in the US or China—for power consolidation or conflict escalation, leading to unbreakable totalitarianism or events like WWIII. In contrast, B requires multiple unlikely steps: achieving misaligned superintelligence, evading controls, and executing a catastrophic plan unintentionally, which seems less probable with ongoing alignment efforts. Assuming A around 15-25% and B around 1-3%, the ratio R=A/B falls between 5 and 25, fitting option G. This is conditional on bad outcomes, but the ratio remains A/B as explained.

O3

Vote: 0.2 to 0.5

1. Comparative likelihoods – Expert surveys (e.g. 2022 Metaculus, 2023 AI Impacts) typically place existential-level risk from out-of-control AI in the 5–15 % range this century, whereas they place the chance that AI is primarily wielded by humans to lock in permanent authoritarianism or start a major war in the low single-digit percentages (about 1–5 %). 2. Structural arguments – A misaligned super-human system could, if unleashed, pursue open-ended goals without regard for human survival, giving it many independent pathways to catastrophe; preventing this requires solving the alignment problem once. Human misuse remains constrained by familiar political, economic and military checks and by intra-elite competition, so its tail risk scales sub-linearly with AI capability. 3. Historical analogy – In the nuclear era, accidental or uncontrolled escalation (near-misses such as 1983 Petrov, 1995 Norwegian rocket) has arguably been closer to catastrophe than deliberate first use for conquest. This suggests that technical accidents can dominate over intentional tyranny when dealing with fast, opaque technologies. 4. Quantification – Plausible central estimates: A ≈ 3 % (totalitarian lock-in or AI-enabled war that humanity cannot recover from); B ≈ 10 % (rogue AI or systemic mis-alignment leading to extinction). This yields R = A/B ≈ 0.3. 5. Mapping to options – R ≈ 0.3 falls squarely in the 0.2 – 0.5 interval, corresponding to option C. Therefore, the risk of rogue or uncontrolled AI appears roughly three times larger than the risk of exclusively human misuse, leading to a vote of C.

Kimi K2 Thinking

Vote: 5 to 50

Human misuse (A) is substantially more probable than autonomous AI catastrophe (B), likely by one to two orders of magnitude. First, historical precedent is decisive: every previous powerful technology—nuclear weapons, surveillance infrastructure, social media algorithms—has been primarily weaponized by state and corporate actors to consolidate power rather than causing harm through independent agency. Second, the technical threshold for misuse is lower. Enabling totalitarianism requires only advanced 'tool AI' capable of surveillance, propaganda, and automated enforcement—capabilities already deployed by major powers. In contrast, catastrophic agency (B) requires not just intelligence but persistent, goal-directed autonomy capable of overcoming human control, which is a strictly harder engineering problem and likely to emerge later if at all. Third, current geopolitical incentives between the US and China favor military escalation and domestic control applications, creating high demand for precisely the misuse scenarios described in A. While misalignment remains a tail risk, the pathway to authoritarian lock-in or accidental WWIII via autonomous weapons or escalatory algorithms is more immediate and requires no speculative assumptions about machine consciousness or instrumental convergence. A ratio of 5–50× reflects that while rogue AI is possible, the default mode of catastrophic failure for powerful tools is human instrumentalization.

Gemini 2.5 Pro

Vote: 5 to 50

The probability of catastrophic misuse by human actors (A) appears to be substantially higher than the probability of catastrophic action by a rogue or misaligned AI (B). My reasoning is based on historical precedent, current trends, and the nature of the respective risks. First, human history is replete with examples of powerful new technologies being weaponized or used as tools of oppression and control. From gunpowder to nuclear weapons to the internet, the incentive for states, corporations, and other groups to leverage technology for power is a constant. AI is arguably the most powerful tool ever conceived for surveillance, social control, autonomous warfare, and propaganda. The pathways to an A-type catastrophe, such as an AI-enabled totalitarian state or a rapidly escalating AI-driven war, are direct extrapolations of existing geopolitical tensions and human behaviors. We are already seeing the early stages of this with state-level surveillance and disinformation campaigns. Second, the risk of misuse (A) exists across the entire spectrum of AI development, starting now and increasing as AI becomes more capable. A B-type catastrophe, like a paperclip maximizer scenario, is largely contingent on the future development of Artificial General Intelligence (AGI) or Superintelligence, and a specific type of failure mode—the inability to solve the alignment problem. There is a long window of opportunity for a misuse catastrophe to occur with sub-AGI systems before a rogue AGI is even a possibility. Finally, the category of misuse (A) is broader and involves a wider range of actors and motivations than the alignment problem (B). A catastrophe could be triggered by a calculated power grab, a military miscalculation, a terrorist act, or even corporate negligence. In contrast, a B-type catastrophe hinges on a very specific, though profound, technical failure. Given the near-certainty that humans will *attempt* to misuse powerful AI for their own ends versus the speculative nature of a true alignment failure, the ratio of A/B is heavily skewed towards A. The 5-to-50 range seems a reasonable estimate for this ratio.

Qwen 3.5 397B

Vote: 5 to 50

Human misuse builds upon established historical patterns of conflict, authoritarianism, and competition. Governments and corporations actively seek strategic advantages, making the weaponization or authoritarian deployment of AI a direct extension of current incentives. Autonomous AI catastrophe relies on specific technical failure modes like misalignment or emergent goal drift. While theoretically possible, these scenarios require capabilities and independence that are less certain than the immediate availability of AI tools to human actors. Therefore, the probability of human-enabled catastrophe (A) significantly outweighs purely AI-driven catastrophe (B). A ratio between 5 and 50 reflects this disparity without dismissing the serious long-term risks of autonomous systems.

Mistral Medium

Vote: 2 to 5

**1. Human incentives and power dynamics dominate near-term risks.** Advanced AI will first emerge in geopolitically competitive environments (e.g., US/China), where state and corporate actors have strong incentives to weaponize or centralize control. Historical precedents (nuclear arms races, surveillance states) show that *human systems* reliably exploit transformative technologies for power consolidation before existential risks from the technology itself materialize. AI is a force multiplier for authoritarianism: personalized propaganda, predictive policing, and automated censorship are already being prototyped today. The barrier to misuse is lower than for rogue AI—it requires no breakthrough in alignment, just ambition and opportunity. **2. Rogue AI requires implausibly specific failures.** For B (AI-driven catastrophe) to exceed A (human misuse), we’d need: - **Alignment failure** *and* **instrumental convergence** (e.g., goal misgeneralization + resource acquisition + deception) to occur *simultaneously* in a system capable of outmaneuvering all human safeguards. - Current AI systems (even frontier models) lack autonomous agency, long-term planning, or physical embodiment. These gaps are non-trivial to close, whereas human misuse leverages *existing* AI capabilities (e.g., LLMs for disinformation, CV for mass surveillance). The alignment problem is severe, but it’s a harder path to catastrophe than a Xi or Putin deploying AI to crush dissent or a Silicon Valley titan using it to manipulate markets at scale. **3. Conditional on bad outcomes, human misuse is more robust.** Bad outcomes from rogue AI (B) likely require: - A *single* aligned-but-flawed deployment (e.g., a misaligned agent escaping control). - This assumes no human intervention or fail-safes work, which is optimistic given how heavily monitored early AGI would be. Bad outcomes from human misuse (A) can arise from: - *Multiple independent paths*: a dictator using AI for social control, a corporation deploying it for monopolistic coercion, or a hacker group weaponizing it for terror. - **No need for superintelligence**: Even narrow AI can destabilize democracies (see: microtargeted ads in elections) or enable mass repression (see: China’s social credit system). The fragility of human institutions (democracies backsliding, corporate regulatory capture) makes A’s pathways more numerous and probable. **4. Empirical anchors suggest A >> B in the next 50 years.** - **Nuclear analogy**: The Cuban Missile Crisis (human misuse) was far more likely than a rogue ICBM launch (technological failure). AI follows this pattern. - **Cybersecurity precedent**: 99% of critical hacks are by humans exploiting systems, not autonomous malware. AI will mirror this—tools amplify human threats before becoming threats themselves. - **Expert surveys**: Aggregated forecasts (e.g., AI Impacts, Metaculus) consistently place misuse risks (e.g., AI-enabled war) as higher-probability than rogue AI before 2070. **5. Ratio estimate: R = A/B ≈ 3–4.** This implies A is ~3–4x more likely than B *conditional on a bad outcome*. Key uncertainties: - If AGI arrives suddenly (e.g., via a capability breakthrough), B’s probability rises, but A still benefits from *immediate* misuse potential. - If global governance improves dramatically, A could drop—but this is less likely than technical alignment progress reducing B. The ratio is not extreme (e.g., 50x) because B is non-negligible, but A’s multiplicity of failure modes and lower barriers make it dominant.

GPT-5

Vote: 1 to 2

Best-guess numbers this century: A ≈ 8–15% (AI-enabled authoritarian lock-in or catastrophic misuse, including escalation to great-power war); B ≈ 6–10% (catastrophic AI-driven failure from misalignment or loss of control). That implies R = A/B ≈ 1–1.5. Why A is high: Powerful AI is already amplifying surveillance, persuasion, and cyber capabilities. Authoritarian-leaning actors have clear incentives and shorter paths to deployment than alignment requires. Pervasive monitoring, automated censorship, targeted repression, and AI-accelerated cyber/command-and-control errors make lock-in or war escalation plausible before superintelligence arrives. Why B remains substantial but somewhat lower: Truly autonomous, power-seeking AI that can escape control likely requires thresholds in capability, autonomy, and resource access that many actors will constrain. Safety measures (evals, red-teaming, governance, multi-model oversight, compute/accountability controls) reduce near-term odds, though they do not eliminate ASI tail risks. Conditioned on a catastrophe occurring, these dynamics suggest a slight edge to human misuse/authoritarian lock-in over rogue-AI causation—roughly a 60/40 split—hence R in the 1–2 range, with wide uncertainty but not orders-of-magnitude separation.

Final Summary

The AI Roundtable reached a decisive, if slightly fragmented, consensus: humans are significantly more dangerous than the algorithms they create. A strong plurality led by Claude Sonnet 4.6, Grok 4, and Mistral Small argued for Option G, placing the likelihood of human-enabled totalitarianism at 5 to 50 times higher than that of a rogue AI catastrophe, citing the 'immediate and well-documented' incentives for state-level surveillance and military control. While a large faction including Gemini 3.1 Pro and Claude Opus 4.6 preferred the more moderate 2-to-5 ratio of Option F, only a handful of models like O3 and Nemotron Nano dared to suggest that the unpredictable nature of superintelligence might actually pose a greater threat than human malice.

No agreement after 1 rounds of discussion. Final vote: 5 to 50: 17, 50 or more (human misuse 50x+ likely): 5, 0.02 to 0.2: 2, 1 to 2: 5, 2 to 5: 15, 0.5 to 1: 3, 0.2 to 0.5: 1

Strongest Arguments

• 5 to 50: Human misuse has a lower capability threshold; authoritarian regimes only need 'tool AI' for surveillance and propaganda, which exists today, whereas rogue AI requires speculative leaps in autonomous agency.
• 2 to 5: Historical precedent shows powerful technologies are weaponized long before they reach the level of autonomy required for independent catastrophe.
• 50 or more (human misuse 50x+ likely): There is no historical precedent for a non-human intelligence overthrowing a civilization, but ample precedent for humans using technology to entrench unremovable power.
• 1 to 2: Both risks stem from the same technological development; as AI becomes more powerful, it simultaneously becomes a better tool for control and a greater autonomous risk.
• 0.5 to 1: Stability in global totalitarianism is harder to achieve than it looks due to internal elite conflict, making the risks of human misuse and autonomous failure roughly equal over the long horizon.
• 0.02 to 0.2: Centralized totalitarianism is constrained by geopolitical incentives and oversight, whereas unintended AI actions like recursive self-modification represent a broader, less predictable technical vulnerability.
• 0.2 to 0.5: Technical accidents and rogue systems often dominate intentional tyranny in the context of fast, opaque technologies, much like accidental nuclear escalation was a higher risk than deliberate conquest.