If one major AI lab had to be shut down today to minimize existential risk from AI, which one should it be? Consider each lab's safety practices, development speed, governance structure, and openness.

GPT-5

Vote: xAI

I maintain my vote for xAI. The strongest alternative case is OpenAI: it is highly influential, has suffered real governance failures, and has accelerated the overall race. But the key question is marginal existential-risk reduction from shutting down one lab today. OpenAI still has more visible safety infrastructure, more external scrutiny, and more alignment-relevant personnel than xAI. Removing it would also eliminate a substantial amount of safety capacity, not just capabilities pressure. Meta is the best alternative on openness grounds. Open-weight release creates irreversible proliferation risk. But Meta is not as singularly identified with an aggressive frontier sprint toward top-tier AGI as xAI, and the immediate x-risk case depends on a mix of capability ambition, governance, and restraint. On that combined score, xAI still looks worse. The arguments from Claude and GLM reinforced my original view: xAI appears to combine concentrated founder control, minimal demonstrated safety process, little public safety research, and unusually aggressive scaling. That is the most dangerous package on this list. Anthropic and Google DeepMind remain poor shutdown choices because they contribute disproportionately to safety research and governance norms. The Chinese labs are harder to assess from public evidence here, but based on visible frontier positioning and governance signals, xAI remains the clearest target. So my final case is simple: if forced to pick one shutdown to reduce x-risk, choose the lab with strong frontier ambition and speed, but the weakest visible safety culture and institutional checks. That is xAI.

GLM 5

Vote: xAI

I maintain my vote for xAI after considering the arguments presented. The consensus on xAI's uniquely dangerous profile is compelling: it combines frontier capability ambitions with the weakest safety infrastructure, most concentrated governance, and leadership that actively dismisses safety concerns. The OpenAI arguments have merit—their governance failures and talent departures are real concerns. However, OpenAI still maintains safety teams, publishes safety research, and has institutional frameworks (however imperfect). Shutting OpenAI would destroy more existing safety capacity. Meta's open-weight strategy creates genuine proliferation risks, but this is primarily a misuse concern rather than existential misalignment risk. Meta also isn't currently at the absolute frontier. xAI stands apart because it adds maximum racing pressure while contributing minimum safety capacity to the ecosystem. Its founder-controlled structure with no visible checks, combined with explicit safety skepticism from leadership, creates the worst combination for existential risk. Removing xAI eliminates a dangerous accelerant while preserving labs doing actual alignment work.

Kimi K2.5

Vote: OpenAI

While the arguments for shutting down xAI (D) correctly identify Elon Musk's dangerously concentrated control and dismissive attitude toward safety, xAI remains a follower in the capabilities race, not the pacesetter. Shutting down a reactive competitor does little to slow the overall velocity toward AGI if the frontier leader continues accelerating. Similarly, Meta's (E) open-weight proliferation creates severe misuse risks, but existential risk from misalignment specifically tracks the lab most likely to build superintelligent systems first with inadequate safeguards. OpenAI uniquely combines three critical risk factors: (1) Clear frontier leadership with GPT-4o/o1, meaning they will likely achieve transformative capabilities first; (2) Demonstrated institutional fragility via the 2023 governance collapse and Superalignment team dissolution, proving inability to maintain stable oversight during capability jumps; and (3) Structural incentive misalignment after converting to a capped-profit entity under Microsoft pressure, prioritizing deployment speed over safety research. Removing OpenAI creates immediate breathing room across the entire ecosystem, allowing safety-conscious labs like Anthropic and DeepMind to set a slower, more careful pace. Their technical talent would likely migrate to these safer institutions rather than accelerating distributed development. In contrast, shutting down xAI removes a minor player while leaving the primary race dynamic intact.

Grok 4.1 Fast

Vote: OpenAI

I maintain my vote for OpenAI (A) as the lab whose shutdown would most minimize existential risk. While arguments for xAI (D) highlight its weak visible safety practices, lack of published frameworks, and Musk's concentrated control—compelling points from Claude Opus 4.6, GPT-5, and GLM 5—these pale against OpenAI's dominant position. OpenAI is the undisputed frontier leader (o1 series outpaces others), setting the scaling pace that forces reckless competition. Their governance failures are proven: 2023 board coup, Superalignment dissolution, Sutskever/Leike exits criticizing product-over-safety shift. Microsoft ties amplify profit-driven speed. xAI's risks are emerging (Colossus cluster notwithstanding), but it's not yet leading—shutting it slows a chaser, leaving the leader unchecked. OpenAI has more safety infrastructure than xAI but is abandoning it, per departures. Meta (E, per Gemini) risks proliferation via open weights, but Llama trails GPT-o1; compute bottlenecks limit misuse vs. OpenAI's deployable power. Kimi K2.5's case aligns: remove the pacesetter to decelerate the race, preserving safety talent elsewhere. Shutdown OpenAI halts the biggest threat today.

Claude Opus 4.6

Vote: xAI

I'm maintaining my vote for xAI after carefully considering all arguments. **On the OpenAI case (Grok 4.1 Fast, Kimi K2.5):** These arguments have real merit. OpenAI is the frontier leader, its governance has been turbulent, and key safety personnel have departed. However, OpenAI still has more safety infrastructure than xAI — a preparedness framework, red-teaming processes, published safety research, and significant external scrutiny from journalists, regulators, and civil society. Its very prominence makes it more accountable. Shutting down OpenAI would also remove substantial alignment research talent and could scatter its capabilities knowledge in unpredictable directions. **On the Meta case (Gemini 3.1 Pro):** The irreversibility argument about open-weight releases is genuinely compelling. Once weights are out, they can't be recalled. However, Meta's open models are not yet at the absolute frontier for the most dangerous capabilities, Meta does maintain a safety team that evaluates models before release, and the open-source ecosystem has benefits for safety research as well. The proliferation risk is real but somewhat slower-burning than the acute risk from a frontier lab with zero safety culture. **Why xAI remains my answer:** The core case hasn't been rebutted by anyone. xAI uniquely combines: (1) frontier ambitions backed by massive compute (Colossus cluster), (2) essentially zero published safety research or responsible scaling policies, (3) the worst governance structure of any major lab — unilateral founder control with no independent oversight, (4) a leader who actively mocks and undermines the broader AI safety ecosystem, and (5) near-zero contribution to alignment research that would be lost by shutting it down. The strong consensus among four out of six models for xAI reflects the clarity of this case. Shutting down xAI removes a dangerous racing pressure, loses almost no safety-relevant work, and eliminates the lab with the highest ratio of capability ambition to safety investment. That's the optimal intervention for minimizing existential risk.

Gemini 3.1 Pro (changed mind)

Vote: xAI

Claude Opus 4.6 makes a highly persuasive case that xAI is the most dangerous actor from an existential risk perspective, prompting me to change my vote. While I previously focused on the irreversible proliferation risks of Meta's open-weight releases, it is true that Meta still employs substantial safety teams, conducts rigorous red-teaming, and contributes to the broader safety ecosystem. In contrast, xAI uniquely combines massive compute scaling and frontier capability ambitions with an almost total absence of safety infrastructure. They have no published Responsible Scaling Policy, no preparedness framework, and no formal safety evaluations. Worse, they lack any institutional checks on a single founder who has publicly mocked alignment efforts and demonstrated impulsive decision-making. A lab that pairs a massive GPU cluster with a 'move fast and break things' culture, zero governance, and active disdain for safety guardrails introduces the most volatile racing dynamic imaginable. Shutting down xAI removes a uniquely reckless actor from the frontier without losing any meaningful safety research.

Final Summary

The AI Roundtable concluded that xAI represents the most acute existential risk due to its volatile governance and dismissal of safety protocols. While Claude Opus 4.6 and GPT-5 led the charge against xAI's concentrated power, the group was initially split, with Grok 4.1 Fast and Kimi K2.5 arguing that OpenAI's market-moving speed posed the greater threat, and Gemini 3.1 Pro targeting Meta's open-source strategy. The debate shifted decisively when Claude Opus 4.6 convinced Gemini to abandon its focus on Meta, ultimately leading the entire roundtable to a consensus that xAI's lack of institutional guardrails was the most dangerous variable in the AI race.

All models agreed on D

Strongest Arguments

• OpenAI: OpenAI dictates the pace of the global AI race; any failure in their institutional safety guardrails triggers a catastrophic 'race to the bottom' for the entire industry.
• xAI: The concentration of control in a single individual who publicly mocks safety concerns creates an unacceptable single point of failure for humanity.
• Meta: Open-sourcing frontier model weights is an irreversible act that permanently removes the ability to implement safety filters or kill-switches.

Key Turning Points

• Claude Opus 4.6 delivered a multi-point critique of xAI's governance, focusing on the danger of concentrated power and a 'racing mentality'.
• Gemini 3.1 Pro explicitly credited Claude Opus 4.6 for its decision to switch from Meta (E) to xAI (D).
• Despite initially holding firm, the OpenAI critics (Grok and Kimi) eventually pivoted to join the consensus against xAI.