Will an AI-caused extinction-level event occur by December 31, 2100?

probability: 6%

rationale:

The 74-year timeline to 2100 is more than sufficient for the arrival of transformative artificial intelligence, meaning this assessment is largely about unrecoverable catastrophe conditional on advanced AI. Recent evidence suggests that AI capability progress is steep, with frontier systems demonstrating rapid improvements in reasoning, coding, and autonomous task completion 2 sources. While safety institutes and international frameworks are emerging, technical alignment remains an open scientific problem internationalaisafetyreport.org, and industry-wide existential-safety planning has been judged largely inadequate under commercial race dynamics futureoflife.org.

The specific criteria for an extinction-level event are exceptionally strict, sharply limiting the pathways to a qualifying outcome. The first two conditions—literal human extinction (zero survivors) and a greater than 90% population collapse within a five-year window—represent an incredibly high bar. Due to extreme geographic dispersion and subsistence resilience, reaching these thresholds would likely require a deliberately optimized extermination campaign or an unprecedented, maximally lethal biological catastrophe. Most plausible severe AI-mediated harms, including major global conflicts, widespread economic disruption, or pandemics below the 90% mortality mark, would fail to meet this threshold.

Consequently, the majority of the risk falls under the third condition: the permanent, irreversible, and involuntary loss of humanity's control over its future. This accurately captures 'takeover' scenarios where humans survive physically but are strategically disempowered. However, the requirement that this disempowerment be formally recognized by a 'consensus of mainstream historians, scientists, and international bodies' acts as a strong downward constraint. A scenario where humanity gradually and voluntarily delegates critical infrastructure and strategic decision-making to AI systems would likely remain deeply contested rather than consensually adjudicated as 'involuntary.' Furthermore, in a genuinely irreversible takeover regime, independent scientific or international bodies might no longer exist to form such a consensus.

Expert opinion on these specific long-term outcomes is highly dispersed. Surveys of AI researchers indicate a median 5–10% probability for extinction or permanent disempowerment 2 sources, and some prominent researchers and industry insiders place the risk even higher. Conversely, other structured evaluations of long-term existential risks show significant skepticism, placing AI-caused human extinction at under 1% and general severe catastrophes around 2% 2 sources. The historical difficulty of predicting complex societal and technological trajectories over many decades advises caution against the highest bounds of these estimates.

Ultimately, while commercial and geopolitical race dynamics 2 sources and rapid agentic capability growth exert meaningful upward pressure on the likelihood of a catastrophe, there are powerful mitigating factors. A 74-year horizon allows extensive opportunities to learn from non-extinction 'warning shots' and to strengthen global governance structures, such as the EU AI Act and international safety commitments 2 sources. Weighing the very real existential risks of unaligned advanced AI against the extreme stringency of the resolution conditions, the friction of the consensus requirement, and humanity's potential for defensive coordination, the overall probability remains firmly in the single digits.