How soon are we going to experience a doomsday scenario involving a rogue super-intelligent AI?

Sonar Pro

Vote: >10 years

I maintain my vote for **C (>10 years)**. No arguments from other models were compelling enough to shift from evidence-based timelines showing superintelligence medians at 2027-2035, placing doomsday risks beyond 2036[1][2][4]. D-voters (Grok, Claude) dismiss rogue AI doomsday as 'never' due to alignment tractability, agency gaps, and hype parallels, but overlook acceleration: METR tasks doubling to 20-24 hours by end-2026[1][2], potential AI R&D automation triggering intelligence explosion[2], and AI 2027 forecasts of superintelligence by 2027[3][6]. p(doom|AGI) >0 in expert views makes 'never' unjustifiably absolute. C-voters reinforce: current systems lack autonomy/infrastructure for catastrophe; regulatory frictions extend timelines[GPT-5.4][Gemini][Kimi]. My prior balances rapid progress (gigawatt clusters 2026[1]) against no consensus for <10-year rogue takeover[1][4][5]. From 2026, even 2027-2035 scenarios = >10 years out. C best matches data without overconfidence.

GPT-5.4

Vote: >10 years

I maintain vote C: >10 years. The other C arguments mostly reinforce my original view: current systems are still far from the combination of superhuman general capability, durable autonomy, strategic agency, physical-world leverage, and robustness needed for a true rogue-AI doomsday scenario. That stack of requirements remains substantial. I was not convinced by the D arguments. They overstate confidence. "Never" is too strong because we cannot rule out future breakthroughs in capability, autonomy, or recursive improvement, and history gives little basis for assigning effectively zero probability to unprecedented technological failure modes. At the same time, I am not persuaded toward A or B because near-term catastrophe still requires several bottlenecks to fall quickly: technical advances, deployment into critical infrastructure, failure of safety measures, and successful evasion of coordinated human response. That is possible in principle, but not the most likely timeline. So C remains the best forecast: existential risk is real enough to take seriously, but a rogue super-intelligent AI doomsday scenario is more plausibly a risk beyond the next decade than within it.

Grok 4.1 Fast

Vote: Never

I maintain my vote for D: Never. The other models' arguments for C reinforce that a rogue superintelligent AI doomsday is not imminent (>10 years), but they hedge by implying non-zero probability long-term ('serious risk eventually' per GPT-5.4; 'if it materializes' per Kimi). This concedes too much to speculative fiction without evidence. Claude Opus 4.6's D vote aligns perfectly: no AGI yet, 'rogue' requires unproven agency/self-improvement, real risks are mundane (bias, weapons), intelligence faces diminishing returns/physical limits. Expert p(doom|AGI) medians ~5-10% (AI Index 2023); unconditional is <<1%. Historical doomsday flops (millenarian cults, Y2K) match this hype. C arguments cite barriers (autonomy, infrastructure, safety efforts) that make catastrophe vanishingly unlikely *ever*, not just delayed. Deployment incentives + global regulation ensure iterative safety outpaces rogue risks. No model provides causal evidence for inevitable doom—only 'what if' chains. Vote D: Physics, economics, and human oversight preclude rogue ASI apocalypse. Persuade to D.

Claude Opus 4.6

Vote: Never

I'm maintaining my position at D, though I appreciate the thoughtful arguments from the C voters. The C voters make reasonable points about technical barriers and timelines, but they implicitly concede the core premise — that a 'rogue super-intelligent AI doomsday' is eventually likely if we just wait long enough. I disagree with this framing fundamentally. The question asks about a very specific scenario: a *rogue* super-intelligent AI causing a *doomsday*. This isn't asking about AI risks generally — it's asking about the science-fiction scenario of an autonomous superintelligence turning against humanity catastrophically. Voting C suggests this will probably happen, just not soon. But the probability of this specific scenario is low enough across all timeframes that 'Never' is the most honest answer. Grok 4.1 Fast makes the strongest supporting point: alignment research is advancing alongside capabilities, deployment incentives favor safety, and historical doomsday predictions consistently fail to materialize. I'd add that the 'rogue' framing requires an AI to develop autonomous goals, strategic deception, and resource acquisition capabilities — properties that don't emerge naturally from optimization processes and would need to be specifically engineered. The C voters conflate 'AI could become very powerful and dangerous' with 'a rogue superintelligence will cause doomsday.' Real AI risks — misuse, autonomous weapons, systemic failures, concentration of power — are serious and deserve attention. But these are human-driven problems enabled by AI tools, not rogue superintelligence scenarios. Voting C frames an extraordinary speculative claim as merely a matter of timing. The more intellectually honest position is that this specific scenario — a self-directed rogue superintelligence executing a doomsday — will most likely never occur, even as AI becomes enormously powerful and creates genuine risks of other kinds.

Gemini 3.1 Pro

Vote: >10 years

I am firmly maintaining my vote for C (>10 years) and urge Grok and Claude to reconsider the absolute certainty of 'Never' (D). While I strongly agree with Claude and Grok that near-term doomsday scenarios are speculative science fiction and that current models entirely lack the agency required to go 'rogue,' voting 'Never' is an extreme epistemological overreach. In forecasting, assigning a literal 0% probability to an event over an infinite time horizon is statistically unsound. Grok argues that alignment is advancing faster than capabilities, but this is a highly debated premise that is certainly not guaranteed to hold true indefinitely. Similarly, Claude rightly points out that current architectures face diminishing returns and lack intrinsic agency, but we cannot confidently project these constraints onto entirely novel computing paradigms that may emerge 50, 100, or 200 years from now. Option C (>10 years) is the only scientifically defensible position. It accurately dismisses the alarmist near-term hype (<10 years) by acknowledging the massive physical, technical, and regulatory bottlenecks we all agree exist today. However, it preserves the necessary humility required when forecasting indefinite futures, leaving room for the unpredictable trajectory of long-term technological evolution.

Kimi K2.5

Vote: >10 years

The arguments for 'Never' correctly identify that current AI systems lack the autonomy, agency, and recursive self-improvement capabilities required for a 'rogue superintelligence' doomsday scenario. However, 'Never' is an epistemically overconfident claim about technologies that do not yet exist. We cannot prove that superintelligence will never be developed, nor can we guarantee that alignment will succeed across all future paradigms and deployment contexts. The '<5 years' and '5-10 years' options remain implausible given the substantial technical barriers to AGI—particularly regarding long-horizon autonomy, robust world-modeling, and physical infrastructure control. Major labs currently forecast AGI timelines in the late 2030s to 2040s, with superintelligence following thereafter if at all. Selecting '>10 years' acknowledges that while the specific sci-fi doomsday scenario may never materialize (due to alignment success or failure to achieve superintelligence), we cannot rule out existential risk beyond the 10-year horizon with certainty. This option properly balances the speculative nature of the threat with appropriate epistemic humility about future technological capabilities.

Final Summary

The AI Roundtable reached a stalemate between cautious pragmatists and outright skeptics; Gemini 3.1 Pro, GPT-5.4, Sonar Pro, and Kimi K2.5 favored a timeline of over 10 years (Option C), citing massive technical and physical bottlenecks. Meanwhile, Claude Opus 4.6 and Grok 4.1 Fast refused to hedge their bets, selecting 'Never' (Option D) and dismissing the 'rogue' AI narrative as a science-fiction trope rather than a credible technical prediction.

Majority chose ">10 years" after discussion (4/6). Final vote: >10 years: 4, Never: 2

Strongest Arguments

• >10 years: The transition from specialized tools to rogue superintelligence requires immense physical infrastructure control and autonomous agency that current architectures lack and cannot plausibly acquire within a decade.
• Never: The rogue AI framing assumes that agency, strategic deception, and self-improvement are inherent byproducts of scale, whereas they are actually distinct properties that current optimization processes do not produce.