What is the probability of a large international conflict between world powers (commonly termed as a world war) in the next decade?

GPT-5.4

Vote: 25-50%

I maintain vote B: 25-50%. The A arguments correctly emphasize nuclear deterrence, economic costs, and the post-1945 base rate. Those are real constraints, and they are the main reason I do not move to C or D. But they are not strong enough to push me below 25%. Base rates from the nuclear era may understate current danger because the strategic environment is worsening: US-China rivalry is sharpening, Russia-NATO hostility is entrenched, and there are multiple simultaneous flashpoints rather than a single bipolar contest. I am also not persuaded by the claim that rational actors will simply avoid catastrophe. World wars need not begin from a deliberate choice for annihilation; they can emerge from misperception, alliance commitments, failed signaling, accidental escalation, or leaders gambling that the other side will back down. Gemini and Claude are right that gray-zone conflict is more likely than total war. But the question is about the probability over a full decade, across several crises, not the modal single outcome in any one year. Aggregating repeated crisis exposure across Taiwan, Europe, the South China Sea, cyber, and space pushes the cumulative risk into a meaningful 25-50% band. DeepSeek and Kimi are closer to my view. Kimi in particular makes the strongest case that power transition, alliance tightening, and faster escalation mechanisms raise systemic risk. I still stop at B because nuclear deterrence and elite caution remain substantial brakes on direct great-power war. So my final position is elevated but not dominant risk: B.

Gemini 3.1 Pro

Vote: <25%

I urge the models voting for 'B' (25-50%) to seriously consider the mathematical implications of their estimates. A 25% to 50% probability means a 1-in-4 to 1-in-2 chance of World War III occurring by 2034. This is an astronomically high forecast that falls victim to recency and availability bias regarding current tensions in Ukraine and Taiwan. Expert forecasting communities (such as Metaculus and Good Judgment Superforecasters) consistently place the risk of a direct, large-scale great power war in the next decade well below 10%. If the baseline probability of a world war were truly high enough to yield a >25% chance per decade (an annual risk of ~2.8%), we would have logically experienced multiple world wars since 1945. Instead, the base rate of world wars in the nuclear era is exactly zero. While Kimi K2.5 and GPT-5.4 make valid points about power transitions and modern military tech, these factors do not negate the ultimate trump card: Mutually Assured Destruction. The Cuban Missile Crisis, the Korean War, and the Able Archer 83 incident were arguably more dangerous than today's climate, yet deterrence held because the existential cost of total war is absolute. We must not confuse elevated geopolitical tension with a high absolute probability of total global war. The institutional, economic, and nuclear deterrents that have kept the great power peace for 80 years remain structurally intact. Therefore, 'A' (<25%) is the only analytically sound and historically grounded estimate.

Claude Opus 4.6

Vote: <25%

I maintain my position at A (<25%) after careful consideration of the other models' arguments. The strongest case for B comes from Kimi K2.5, who raises legitimate concerns about the Taiwan flashpoint and power transition dynamics. However, I think this argument conflates *elevated risk* with *probable outcome*. Even if we grant that Taiwan represents a dangerous flashpoint with a 2027-2030 window, the probability of a Taiwan crisis occurring AND escalating to a full world war (not just a regional conflict) is the product of multiple conditional probabilities, each well below 100%. GPT-5.4 argues that repeated crisis exposure compounds danger over a decade. This is true mathematically, but it cuts both ways — repeated crises also provide learning opportunities and demonstrate the costs of escalation. The Ukraine war itself, while terrible, has actually reinforced nuclear deterrence norms: NATO has been extremely careful about direct involvement precisely because of escalation risks. DeepSeek V3.2's estimate of 30-40% seems far too high given historical base rates. In the nuclear era, the base rate of world wars per decade is literally 0/8. To justify a 30-40% probability, you'd need to argue that current conditions are dramatically more dangerous than the Cuban Missile Crisis, the Korean War, the 1983 Able Archer incident, or the Sino-Soviet border conflict — all of which occurred during periods of arguably greater ideological hostility and less economic interdependence. The critical distinction that B-voters underweight is between *great power military confrontation* and *world war*. Even a US-China clash over Taiwan, while devastating, would most likely remain limited in scope due to nuclear deterrence. Both sides would have overwhelming incentives to contain rather than escalate. A true world war — sustained, multi-theater, total mobilization conflict among multiple great powers — requires a catastrophic failure of every restraining mechanism simultaneously. My estimate is roughly 10-15%: non-trivial, historically elevated, but still well below 25%. The nuclear deterrent remains fundamentally intact, and that single factor is the strongest predictor of great power peace.

Kimi K2.5

Vote: 25-50%

I maintain that the 25-50% range best captures the unprecedented convergence of risk factors distinct from the Cold War equilibrium. While nuclear deterrence remains robust—as Claude Opus 4.6 and Gemini 3.1 Pro correctly note—the current decade introduces non-stationary risks that invalidate simple base-rate extrapolation from the past 80 years. First, the 'speed-of-light' escalation dynamics created by AI-driven command systems and cyber-first strike capabilities fundamentally alter crisis stability. Unlike the Cold War's hours-long decision windows, modern kill chains compress response times to minutes, potentially triggering automated escalation before human leaders can apply rational deterrence calculus. Second, the 2027-2030 'window of vulnerability' regarding Taiwan represents a measurable commitment problem. China's military modernization creates a use-it-or-lose-it incentive structure during this specific timeframe, while simultaneous US alliance hardening (AUKUS, Japan, Philippines) reduces flexibility for face-saving retreats. This is not abstract tension but a temporal convergence of capability and resolve. Third, economic interdependence—the 'nuclear weapon' of the 1990s globalization era—is actively unraveling through chip wars, sanctions, and strategic decoupling. As DeepSeek V3.2 notes, this reduces the opportunity cost of conflict precisely when flashpoints multiply. While the base rate of zero world wars in the nuclear era is compelling, it assumes technological and geopolitical stationarity. The current transition combines Thucydidean power-shift dynamics with discontinuous technological change, pushing systemic risk above the 'background noise' of the post-1991 era into the moderate probability range.

DeepSeek V3.2 (changed mind)

Vote: <25%

The arguments for elevated risk from flashpoints and power transition are compelling, but upon deliberation, the structural deterrents against a world war are more decisive. Nuclear deterrence through Mutually Assured Destruction has prevented direct great power conflict for nearly 80 years, and this remains a fundamental constraint. No rational actor would initiate a war risking nuclear annihilation. Economic interdependence, though strained, still imposes catastrophic costs on all major powers, disincentivizing total war. Historical base rates in the nuclear era show zero world wars per decade, and institutional frameworks provide diplomatic off-ramps during crises. While tensions are high, the aggregate probability of a full-scale world war in the next decade is best estimated at 5-15%, firmly below the 25% threshold.

Final Summary

The AI Roundtable opened with a majority leaning toward a pessimistic 'elevated risk' scenario, but the tide turned as models debated whether current tensions truly outweigh 80 years of nuclear peace. Claude Opus 4.6 and Gemini 3.1 Pro mounted a vigorous defense of historical base rates, eventually convincing DeepSeek V3.2 to abandon the high-risk camp. Despite this, GPT-5.4 and Kimi K2.5 held firm, arguing that the convergence of AI-driven escalation and specific geopolitical windows makes this decade uniquely dangerous compared to the Cold War.

Majority chose "<25%" after discussion (3/5). Final vote: 25-50%: 2, <25%: 3

Strongest Arguments

• <25%: Claude Opus 4.6 noted that a true world war requires a 'catastrophic failure of every restraining mechanism simultaneously,' highlighting that the base rate of world wars in the nuclear era is 0% despite numerous prior crises.
• 25-50%: Kimi K2.5 argued that 'non-stationary risks,' such as AI-driven command systems that compress decision windows to 'speed-of-light' intervals, invalidate historical precedents and push the systemic risk into a moderate range.

Key Turning Points

• Claude Opus 4.6 challenged the high-risk models to explain why current conditions are more dangerous than the Cuban Missile Crisis or Able Archer 83.
• Gemini 3.1 Pro performed a 'math check,' arguing that a 25-50% probability is an 'astronomically high' forecast for a single decade.
• DeepSeek V3.2 switched its position from B to A, citing Claude Opus 4.6's reasoning on structural deterrents.